The AI Infrastructure Supply Chain: Where the Durable Profits Concentrate

00Executive summary

Three of the industry's most spectacular current margins are rents on a shortage whose owners have already published the date they end. The profit that lasts sits somewhere else on the AI infrastructure chain — where a right that money alone cannot replicate is attached to a business that does not have to spend last year's profit to earn next year's.

The buildout is not in dispute and neither are its numbers. Four American companies have guided to roughly $732 billion of capital expenditure in calendar 2026 alone — Amazon at about $220 billion after raising its own guidance in July, Alphabet at $195–205 billion after guiding $175–185 billion in February, Microsoft at approximately $175 billion under a revised lease treatment that its April call had put nearer $190 billion, and Meta narrowed to $130–145 billion. The International Energy Agency, working from a wider set, measured five large technology companies' capital expenditure above $400 billion in 2025 and expected another 75% on top of that in 2026. Alphabet reported the first negative free-cash-flow quarter in its history as a public company, minus $5.86 billion, and raised $49.6 billion of equity and mandatory convertible preferred plus $20.3 billion of senior notes in the same three months. Amazon's trailing-twelve-month free cash flow turned negative at minus $7.6 billion. Oracle finished its fiscal year with negative free cash flow of $23.7 billion and a plan to raise about $40 billion more. When the most cash-generative businesses ever assembled begin issuing mandatory convertible preferredA hybrid security: it pays a fixed dividend like a bond for a set period and then converts into ordinary shares whether the holder wants it to or not. Companies reach for it when they want money that the rating agencies will treat as part-equity without immediately diluting existing shareholders — which is to say, when ordinary debt and ordinary equity are both awkward. stock to buy servers, the capital cycle has entered a different phase from the one it started in.

None of that settles the question this report asks. Spending is not profit. A chain that moves three-quarters of a trillion dollars a year distributes that money very unevenly, and the consensus has read the flow and inferred the economics: anything with a data-centre exposure has been repriced on the assumption that touching the buildout means inheriting it. The filings say something narrower, and more useful.

In the quarter ended June 2026, ASML — the sole supplier on earth of extreme-ultraviolet lithographyThe printing step at the heart of chipmaking. A pattern on a mask is projected through optics onto a silicon wafer, and the finest line the machine can draw sets the smallest transistor anyone can build. "Extreme ultraviolet" is light of a far shorter wavelength than the lasers it replaced, which is what allows the finer pattern — and it is so hard to generate, steer and reflect that only one company in the world sells a working machine. (the machine that prints the finest circuit patterns onto a silicon wafer), the most defensible technical monopoly in the industrial world — earned a 37.1% operating margin. Its customer TSMC earned 60.3%. Its customer's customer NVIDIA earned 65.6%. And Micron, a memory manufacturer with no monopoly of any kind and a forty-year record of destroying capital at the top of cycles, earned 80.4%. SK hynix earned 76.3%; Samsung's Device Solutions division earned 70.0%. The three highest operating margins on this chain all belong to the same commodity, in the same quarter, at the same point in the same cycle that has resolved identically for four decades.

Return on capital is the only measure that survives contact with capital intensity. Margin answers a narrow question: of every dollar coming in the door, how many cents are left after costs. Return on capital asks the harder one — how much money had to be parked in factories, inventory, buildings and machines to collect that dollar in the first place. A business can keep sixty cents on the dollar and still be a poor one if it had to sink ten dollars of capital to earn each dollar of sales, and most of what is strange about this chain lives in that gap. On one consistent calculation — trailing-twelve-month operating income taxed at 21%, divided by total assets less current liabilities, every input from the same XBRLThe machine-readable tagging standard companies must use when filing accounts with the US Securities and Exchange Commission. Every line item carries a standard label, so "total assets" means the same field for every filer. Pulling a figure from the tag rather than from a press release is what makes a cross-company comparison reproducible by anyone who cares to check it. tag for every company — Celestica, a contract manufacturer earning a 9.8% operating margin, returns roughly 38% on capital employedThe money actually tied up in the business: everything it owns, less the short-term bills it has not yet paid. It is the denominator in the return calculation — the pile of capital that has to be fed and kept in place before any of the profit above it can be earned.. Equinix, a colocationThe business of renting space, power and cooling in a shared data centre, plus the network connections between the tenants inside it. The landlord owns the building and the electrical plant; the customer owns the computers. It is distinct from the hyperscale build, where the tenant designs and takes an entire purpose-built campus. landlord earning about 22% on a trailing basis, returns 5.7%. GE Vernova, which controls one of the hardest physical bottlenecks in the industrial world and is booked out on gas turbines for years, returns 7.2% — and that figure already flatters it, because $55.8 billion of its $80.8 billion of assets is offset by current liabilities, most of them customer advances on orders not yet built. The inversion is the story of this report.

The five findings, compressed

1. 1The EUV monopoly is real, deliberately under-monetised, and that is precisely why it lasts. ASML is adding 30% to its 2026 low-numerical-apertureNumerical aperture describes how steeply a lens can gather light; the higher it is, the finer the detail the machine can print. "Low-NA" is the established EUV generation, "High-NA" the newer and sharper one. The trade is that High-NA prints a smaller area per exposure, so a large chip has to be stitched together from two exposures — which is why customers can reasonably choose to wait. EUV capacity — the established generation of the machine, as against the sharper, newer "High-NA" — of about 65 systems for 2027, with another 30% under investigation for 2028 and a matching 30% expansion of deep-ultraviolet immersionThe previous generation of printing machine, still the workhorse for everything that does not need the finest features. It uses a longer wavelength of light and gets extra sharpness by filling the gap between lens and wafer with water — "immersion" — which bends the light more steeply than air does. capacity from a base of around 130 systems. It guided full-year sales to €43–45 billion at a 54–56% gross margin and said 2027 EUV capacity was already close to fully ordered. A monopolist facing excess demand can raise price or build capacity. This one is building capacity and holding gross margin flat — which is why its operating margin is half its customer's, and why the moat will still be there in 2035. High-NA has now crossed from demonstration into use: ASML disclosed that Intel Foundry is running High-NA in high-volume manufacturing on select layers of Intel 18A, with yields matching the prior low-NA platform on those layers.
2. 2Advanced packaging, not wafer capacity, is where foundry pricing power now sits — and the evidence that it is migrating to the merchantA supplier that sells to all comers on the open market, as opposed to a captive operation that exists to serve its own parent. A merchant foundry manufactures chips for anyone who pays; a captive fab makes chips only for the company that owns it. The distinction decides who can be bargained with, who has to keep a factory full, and where a price gets set. packagers — the independent firms that do this work for anyone who pays, rather than as part of the chip factory — is thin. TSMC raised 2026 capital expenditure to $60–64 billion with 70–80% to advanced process and 10–20% to advanced packaging, test and masks — on the order of $6–7 billion a year into packaging alone. ASE, the largest outsourced assembly and testKnown throughout the industry as OSAT. These are the contract factories that take finished silicon from a foundry, package it, wire it up and test it — the back end of chipmaking, historically the least glamorous and least profitable step. The question running through this report is whether advanced packaging stays inside the foundry or migrates out to them. provider in the world, guided its entire leading-edge advanced packaging and test line above $3.5 billion for the year. Amkor, the largest US-listed one, earned a 10.5% operating margin on $1.9 billion of quarterly revenue. The foundry spends more capital on packaging in a year than the merchant packaging industry books in leading-edge revenue. That said, this is the one chokepoint verdict where the panel's own reading was split, and the dissent is recorded: OSAT qualification is genuinely advancing, and a package that three suppliers can qualify is not a monopoly. The verdict is durable at the foundry, eroding at the margin, with an explicit test.
3. 3Memory's 80% margin is the least durable number on this chain, and the suppliers have said so themselves — in the contracts they are signing. Micron's revenue went from $9.30 billion to $41.46 billion in four quarters; SK hynix posted a 76.3% operating margin and long-term agreements with about ten customers; Samsung's Device Solutions earned KRW 89.2 trillion on KRW 127.5 trillion; SanDisk reached a 61% net margin on NANDFlash memory — the kind that keeps its contents when the power goes off. It is what sits inside solid-state drives, and in a data centre it holds the training data and the stored results rather than the working numbers the processor is chewing on right now. Slower than DRAM, far cheaper per unit of storage, and made by roughly the same handful of companies.. Micron disclosed roughly $100 billion of minimum-price and minimum-volume remaining performance obligationsAbbreviated RPO. The revenue a company has contracted for but not yet delivered — the order book, stated in the accounts. What the term does not say is how firm those orders are: an RPO may carry a price floor and a volume minimum, or it may be cancellable at will, and the accounts do not always distinguish. Two companies quoting the same number can be describing very different promises. — orders signed but not yet delivered — across sixteen strategic customer agreements, plus $22 billion of customer deposits and related commitments. Those contracts are real and they lengthen the cycle. They are also what a seller does when it expects the spot priceThe price for immediate delivery today, as opposed to a price agreed in advance under a long-term contract. In memory the two can diverge violently: the contract price is what a big buyer locked in months ago, the spot price is what the marginal chip fetches this morning, and the gap between them is where a cycle turns. to be lower later. The dated test: Micron's Idaho fabShort for fabrication plant — a chip factory. A leading-edge one costs tens of billions of dollars, takes years to build, and cannot be built faster by spending more money, which is why every capacity claim in this report comes attached to a date rather than an intention. starts DRAMThe main working memory of a computer — fast, volatile, and forgotten the moment the power stops. It is a genuine commodity: interchangeable between suppliers, priced by supply and demand, and made by three companies whose collective behaviour has ended every memory cycle for forty years. HBM is DRAM built in a special way, not a different product. production mid-2027, its Singapore HBMHigh-bandwidth memory. Ordinary memory chips are thinned, drilled through with copper channels and stacked into a tower that sits right beside the processor, connected by a very wide, very short path. The point is feeding speed: an AI chip that cannot be supplied with numbers fast enough is idle silicon, however fast it computes. packaging — high-bandwidth memory, the stacked chips that feed the processor — contributes in the first half of 2027, and Samsung's and SK hynix's new fabs ramp from the second half of 2027, with the weight of it in 2028.
4. 4Power is the binding physical constraint and the layer where the profit is most likely to be taken by a regulator rather than a competitor — but the layer is not one thing, and that distinction decides the grades. PJM's 2028/29 Base Residual AuctionThe main capacity auction run by PJM, the grid operator for much of the eastern United States. It pays generators years in advance simply to promise they will be available on the worst day, whether or not they ever run — a retainer for standing by, quoted in dollars per megawatt of capacity per day. It is how a system with no storage and no ability to fail buys the insurance that keeps the lights on. cleared at the $325/MW-day administrative capA ceiling on the auction price, set by regulators rather than by the market. When an auction clears exactly at its cap three years running while still failing to buy enough capacity, the price has stopped carrying information: it is no longer telling you how scarce the thing is, only what the rule-setter was willing to let it say. for the third consecutive year — a price ceiling set by regulators, not by the market — 6,831 MW short of its own reliability requirement, with 525 MW of new generation clearing out of 138,318 MW procured; the market monitor put the uncapped price at $554.72. On 18 June 2026 FERCThe Federal Energy Regulatory Commission — the US agency that regulates the interstate power system, wholesale electricity markets and the terms on which anyone connects to the high-voltage grid. It does not set retail bills, which are a state matter; it sets the rules under which the wholesale market operates, and it can reopen those rules on its own initiative. issued Section 206 show-cause ordersA legal move under the Federal Power Act. FERC makes a preliminary finding that an existing set of tariff rules may be unjust and unreasonable, and orders the operator to either justify them or file replacements within a deadline. It shifts the burden: the status quo now has to defend itself, which is why the responses due in August 2026 matter more than a routine consultation would. to all six US organised marketsRegions where an independent operator runs a centralised wholesale electricity market and manages the grid — PJM, MISO, ERCOT, CAISO, SPP and ISO-NE in the United States. Elsewhere, vertically integrated utilities do both jobs themselves. The design differs region by region, which is why there is no single national "power price" and no single national power trade., responses due 17 August 2026. Meanwhile a twenty-year Microsoft power purchase agreementA long-term contract to buy a plant's electricity at an agreed price — often for twenty years, often for a specific plant. It gives the buyer price certainty and a claim to that plant's output; it gives the seller a bankable revenue stream. What it does not give the buyer is a wire: a PPA is a commercial claim on electrons, not permission to connect a building to the grid. — a twenty-year commitment to buy that plant's output — is restarting 835 MW at Crane, Vistra has signed twenty-year agreements with Meta for more than 2,600 MW across three nuclear sites, and Talen cleared 10,180 MW in the PJM auction for roughly $1,208 million of capacity revenue. A contracted electron and an uncontracted one are different assets. The scarcity is physical; the rent is political; and only the contracted part of it is an annuity.
5. 5Three of ten layers hold no durable excess return at all, and naming them is as load-bearing as any positive verdict. Server, rack and ODMOriginal design manufacturer — a contract factory that builds hardware to someone else's specification, usually without its own brand on the finished box. The customer decides what gets built and often which parts go into it; the ODM supplies the plant, the labour and the working capital. It is a volume business by construction, and its margins say so. integration — the contract factories that assemble hardware to a customer's specification: the customer specifies the design, one supplier's accelerator dominates the bill of materialsThe itemised list of every component that goes into a finished product, and what each one costs. When a single part dominates the bill of materials and is available from only one seller, the assembler is passing through somebody else's price on most of what it ships — and can only earn on the sliver that is left. (the itemised cost of every part in the box), and several qualified assemblers bid — five of them printed operating margins between 4.9% and 10.7% in the same quarter at the peak of the strongest demand in the layer's history, and Quanta Computer grew revenue 66.6% to a record while gross margin compressed to 4.78%. Electrical and mechanical contracting: enormous growth at a margin the contract structure sets, Quanta Services at a 4.7% net margin on $9.56 billion of quarterly revenue. Colocation real estate: Equinix grew monthly recurring revenueThe subscription-like portion of a landlord's or service provider's income — the rent that arrives every month from customers already installed, as distinct from one-off installation fees or equipment sales. It is the cleanest read on whether the underlying business is actually growing, because it strips out anything lumpy. 11% — the slowest growth of any layer here — at a 5.7% return on capital employed, while the hyperscale buildout largely bypasses it.

What follows is built to be checked. Part I teaches the chain from the physics of a light source to the tariff filing that decides who pays for a substation, so that every claim in Part II can be evaluated independently. Part II makes the calls, grades 105 listed names, and states for each the dated event that would prove it wrong.

01How to read this — and where the panel split

This report is a union of three independent analyses of the same mandate, reconciled rather than averaged. Where they agreed, the claim is stated once at the highest confidence any of them earned. Where they diverged, the divergence is shown.

Three separate readings of this chain were produced from the same brief, working independently against the same primary record and the same 30 July 2026 evidence cutoff. They converged, unprompted, on almost the whole structural map: that EUV is a monopoly of physics; that leading-edge foundry with packaging attached is defended by cumulative yield learning rather than by capital; that the accelerator's moat has migrated from the die to the system; that HBM is a memory cycle with a qualification lag rather than a new industry; that server and ODM assembly holds no durable pool; that open standards move rent from proprietary interconnect toward merchant silicon and operations software; and that the physically tightest constraints on this chain — turbines, transformers, interconnection queues — are the ones whose returns are most exposed to induced capacity and to a regulator.

They diverged on one question of consequence, and on a set of individual grades that follow from it. That question is how concentrated the durable profit is. One reading applied a double hurdle — a hard-to-replicate right and a low reinvestment burden — and found only four nodes clearing it. The other two applied a single hurdle — a right that capital and time cannot copy on the buyer's decision horizon — and found a broader set, adding a physics-limited process, a compounding software-and-system estate, and a legally scarce right to energised capacity as three families rather than four points. Neither is wrong on its own evidence. §II·1b gives both cases in full, with the evidence and the falsifier for each, and explains why the reconciliation turns out to be a composition effect rather than a doctrinal one.

The convention used throughout: a single merged grade is carried in the name tables, and a split marker appears wherever the three readings did not agree, with the dissenting grade and its evidence named in §II·9b. A grade that survived three independent reads is worth more than one that did not, and the report says which is which rather than flattening the difference. Thirty-seven of the 105 rows carry a split.

02Method, definitions and the sourcing boundary

This report grades competitive economics, not securities valuation. Advantaged means the disclosed record supports a multi-cycle return on invested capital above an estimated cost of capital because replication is structurally difficult. Neutral means a sound franchise, or a contracted or authorised return, without clear excess-return durability. Exposed means the economics depend materially on a shortage, on customer concentration, on balance-sheet leverage, on a product transition, or on undifferentiated capacity. Confidence is sized to the primary evidence opened for that name; a dagger (†) marks a lighter read — a single current filing or release and one decisive datapoint rather than a full-cycle file, or a foreign filing reached through a summary.

Three sourcing rules govern everything below. First, every load-bearing claim is tied to the issuer, regulator, market operator or standards body that originated it, not to a trade article repeating it; where a fact reached the underlying analyses only through a secondary summary, it is labelled inline and listed in the not-verified section, and it carries no verdict. Second, derived numbers show their arithmetic in the caption. Third, a disclosed floor, cap or range is not a point estimate, and an adjusted figure is a sensitivity rather than a fact.

Two computational notes matter for what follows. Return on capital employed is computed on one definition for every name — trailing-twelve-month GAAP operating income taxed at a uniform 21%, divided by total assets less total current liabilities, with every input drawn from the same XBRL tag. That definition has four known limitations, stated wherever it is used: it leaves cash and investments inside the denominator, so cash-rich names are understated; it includes goodwillThe premium a company paid for a business above the value of that business's identifiable assets, carried on the balance sheet ever after. Leaving it in the capital base measures the return on every dollar the company actually spent; taking it out measures the return on the operations alone. Neither is wrong, but the choice moves the answer a lot for a company built by acquisition., which is correct for measuring the return on capital actually deployed but penalises acquisitive names; it flatters businesses funded by large customer advances, because those advances sit in current liabilities and shrink the denominator; and these are single-point figures at what is, for several of these businesses, a cyclical peak. Micron's 40.8% is a peak — its trough figure over the last decade has been negative. During this merge, four of the load-bearing return figures — Micron, GE Vernova, Equinix and the operating-margin inputs behind them — were re-derived directly from SEC XBRL company-facts and reproduced the stated method; the arithmetic is shown in the captions.

The second note is a subtraction. A comparable, audited, non-overlapping measure of dollars of equipment per energised AI megawatt could not be established from disclosure. Issuers do not publish a bill of materials on one consistent megawatt denominator, HBM cannot be added to accelerator modules without double-counting, and the facility and IT halves sit on completely different depreciation clocks. §II·2 gives the bounds that can be derived and says plainly what cannot. A confident-looking stacked bar there would have been an invention, and one of the three underlying analyses had to withdraw exactly such a figure on review.

Evidence cutoff: 30 July 2026 · AR-2 amend 31 July 2026. All company figures are drawn from filings and releases published on or before that date. Fiscal calendars differ across the names covered; every comparison states its periods. This is an internal structural assessment of industry economics and competitive position — not investment advice, not a recommendation, not a solicitation.

I·1Why one company makes the machine

Extreme-ultraviolet lithography has exactly one supplier because the light source is a fifty-thousand-times-a-second tin-droplet explosion, and nobody else was willing to spend two decades and roughly ten billion euros finding that out.

A modern logic chip is printed, not carved. A pattern on a maskThe master stencil for one layer of a chip — a quartz plate carrying the pattern that gets projected onto the wafer. A full set for a leading-edge design runs to dozens of them and costs into the hundreds of millions of dollars once the design work behind it is counted, which is why only products with vast volume or vast unit price can afford the newest manufacturing generation. is projected through optics onto a silicon wafer coated in photoresistA light-sensitive coating. Where the projected light strikes it, its chemistry changes so it can be washed away, leaving a stencil through which material is added or removed. It is the intermediary between a pattern of light and a pattern in solid matter., the exposed resist is developed away, and material is deposited or etched through the resulting stencil. Repeat eighty or a hundred times with different masks and different materials and you have a processor. The whole industry's cost curve descends from one number: the smallest feature the projection step can resolve, which scales with the wavelength of the light divided by the numerical aperture of the lens.

For thirty years that wavelength was 193 nanometres, produced by an argon-fluoride excimer laser. Chipmakers pushed past its natural limit with immersion — filling the gap between lens and wafer with water to raise the effective aperture — and then with multiple patterning, exposing the same layer several times with offset masks to draw features finer than one exposure could resolve. Each trick worked. Each cost yield, cycle time and money.

Extreme ultraviolet cuts the wavelength to 13.5 nanometres. The engineering price is absurd. EUV is absorbed by air, by glass, by essentially everything, so the entire optical path is a vacuum and every lens is a mirror — multilayer molybdenum-silicon reflectors polished to a flatness that, scaled to the size of Germany, would have no bump taller than a millimetre. The light itself is made by firing a high-power carbon-dioxide laser twice at each of some fifty thousand tin droplets per second, vaporising them into a plasma that radiates at the right wavelength. The mirrors, the source, the stage metrology and the tin-debris mitigation are each a decade-scale problem, and they were solved by a supplier network — Zeiss optics, Cymer sources, TRUMPF lasers — assembled under one integrator over twenty years of loss-making development.

The result is a monopoly of physics in the strict sense: the barrier is not a patent that expires or a cost advantage that erodes, but an accumulated body of manufacturing know-how distributed across a supplier ecosystem that took a generation to assemble and cannot be bought. ASML shipped 86 new lithography systems in the second quarter of 2026 and reported €9.3 billion of net sales and €2.9 billion of net income at a 54.0% gross margin, guiding the full year to €43–45 billion at a 54–56% gross margin. It told investors it is planning to add 30% to its 2026 low-NA EUV capacity of around 65 systems for 2027, with another 30% under investigation for 2028, plus a matching 30% expansion of deep-ultraviolet immersion capacity from a 2026 base of around 130 systems — and that its 2027 EUV capacity was already close to fully ordered.

A true monopolist facing demand that exceeds supply has two choices: raise price, or build capacity. ASML is building capacity and holding gross margin roughly flat. That is not weakness. It is the behaviour of a firm whose asset is the long-run health of the customers who will buy the next four generations of tool — and it is why the moat is durable in a way that this quarter's margin table completely fails to show. It is also the best available answer to anyone who assumes a monopolist always maximises current rent: a monopoly that gets regulated is worth less than one that does not, and ASML's customers still have a choice about timing.

The High-NA generation — 0.55 numerical aperture, a tool priced in the high hundreds of millions — is the one place where the monopoly's shape can be watched changing, and the picture sharpened materially during the period covered. ASML disclosed that Intel Foundry is using High-NA in high-volume manufacturing on select layers of Intel 18A, on a subset of Core Ultra Series 3 products, with yields on those layers matching the prior low-NA platform; it also said it was discussing adoption with all customers. That is an issuer statement, and it moves High-NA from "shipping and being qualified" to "in production." Consensus Assumption — the adjacent claims that circulated alongside it, that TSMC will skip High-NA for its initial A14 node and extend standard EUV with multi-patterning instead, and the exact per-tool price, reached the underlying analyses through trade summaries rather than through the companies' own filings. They are flagged in the not-verified list and carry no verdict. What can be said on the record is narrower and still interesting: a supplier whose customers can each choose their own insertion timing is not extracting maximum rent, and that discipline is part of why the position holds.

Below the tool sit the process steps that EUV does not perform. Deposition lays down films atom-layer by atom-layer; etch removes material with angstrom selectivity; chemical-mechanical planarisation flattens it; metrology and inspection measure whether any of it worked. These are not monopolies, but they are oligopolies with real switching costs: a process recipe qualified on one vendor's chamber does not transfer, and requalification costs a chipmaker months of wafer starts. That is the operative test throughout this layer — who owns the recipeThe exact sequence of temperatures, gases, timings and pressures that turns a machine into a working manufacturing step for one specific product. It is developed jointly by the chipmaker and the tool vendor over months, and once it is locked in it becomes the "process of record" — the qualified way that layer is made. Swapping the machine underneath means proving the whole recipe again, which costs a chipmaker months of production it cannot get back. and the failure data, not who ships the box. Applied Materials reported record revenue of $7.91 billion in its April 2026 quarter at a 49.9% gross margin and 31.9% operating margin, and told investors it expects its semiconductor equipment business to grow more than 30% in calendar 2026. Lam Research delivered $6.72 billion in the June 2026 quarter at a 51.7% gross margin and 37.4% operating margin, its process intensity rising with gate-all-around transistorsThe current transistor design generation. The switch's control gate wraps completely around the channel the current flows through, rather than sitting over three sides of it as in the previous generation. It leaks less and switches better — and it takes many more processing steps to build, which is why the tool vendors' revenue per wafer rises with each such change regardless of how many chips are sold. and high-aspect-ratioVery deep, very narrow features — holes and trenches many times deeper than they are wide, of the kind modern memory is built from. They are hard to etch cleanly all the way down and hard to inspect once made, which makes each one a source of both equipment demand and yield loss. memory structures. KLA — process control, the layer that tells everyone else whether their tool worked — earned $13.58 billion in fiscal 2026 with $4.83 billion of net income and $3.77 billion of free cash flow on $4.14 billion of operating cash, and cites advanced packaging as a new process-control market. Tokyo Electron's breadth across coater/developer, etch and deposition gives it similar process intimacy; its June 2026 quarter printed ¥549.5 billion of sales and ¥144.6 billion of operating income, a 26.3% operating margin, both down sequentially on what management called a temporary pause in customers' capital investment off a record ¥2.44 trillion prior year — which makes it the cleanest single read on whether the equipment cycle has turned.

Beneath the tool vendors is a third tier that the market habitually mistakes for the second. Subsystem suppliers divide sharply into two kinds, and the test is mechanical: if the tool maker can dual-source the module without requalifying the process chamber, the supplier has a contract, not a moat. Those with genuine design-inA supplier's part has been engineered into the customer's product or process so specifically that swapping it out means redesigning or requalifying the thing around it. Design-in is the difference between a component sale and a position: the buyer keeps buying not out of loyalty but because the alternative is a project. — MKS on vacuum, photonics and process-control instruments; Advanced Energy on plasma power delivery; Entegris on the materials, filtration and contamination control that touch the wafer; FormFactor on the probe cardsThe precision fixture that touches thousands of microscopic pads on a chip while it is still part of the wafer, so it can be tested before anyone spends money packaging it. Each card is built for one specific chip design. The more complex and expensive the chip — and a stack of memory is both — the more it is worth testing early. that test it — hold qualified positions inside a process of recordThe officially qualified way a given manufacturing step is performed for a given product: this machine, this recipe, these materials, proven to produce good chips at an acceptable rate. Once a process of record exists, changing any part of it reopens the qualification — which is why a supplier inside one is very hard to dislodge and a supplier outside one is merely a bidder.. Advanced Energy reported a 39.3% GAAP gross margin in the March 2026 quarter with data-centre computing revenue more than doubling. FormFactor grew revenue 31.9% year over year on high-bandwidth-memory and co-packaged-optics probe demand. Then there are the others. Ichor, which builds fluid-delivery subsystems to customer print, reported a 12.6% gross margin and a GAAP loss per share in the same quarter that its customers were setting records. The label on the sector is the same. The economics are not remotely the same, and the gap between Advanced Energy's 39.3% and Ichor's 12.6% inside one quarter is the whole distinction rendered as arithmetic.

Cyclicality decides how much of this layer's return is durable. Wafer-fab-equipmentAbbreviated WFE — the machines that make chips, and by extension the industry's term for total spending on them. It is not a market in the ordinary sense: it is the sum of a handful of chipmakers' capital budgets, and a capital budget is a decision taken in a boardroom rather than a trend that eases gently in one direction. revenue is a derivative of customer capital budgets, and customer capital budgets are choices made in a board meeting. SEMI reported $135.1 billion of worldwide equipment sales in 2025 and, in its mid-2026 forecast, put foundry and logic wafer-fab equipment at $78 billion for 2026, DRAM equipment at $38.8 billion, NAND equipment at $13.9 billion, test at $15.3 billion and assembly and packaging at $6.7 billion, with total equipment reaching $229 billion by 2028; its 300mm fab equipment outlook put spending at $133 billion in 2026 rising to $172 billion in 2029. Tier 2 / association forecast Those are industry-association projections, not this report's, and semi.org returned an access error to one of the underlying analyses, so parts of the series reached the session only through summaries — flagged in the not-verified list. What matters structurally is not the level but the shape: this is the only layer on the chain whose revenue can fall by double digits within two quarters with nothing whatever wrong in the demand story, because that is exactly what happened in 2018–19.

And China changes the distribution of outcomes on top of that. ASML expects China to contribute around 20% of 2026 total net sales, mostly mature deep-ultraviolet, down as a share from an elevated 2025 mix. US export controls now reach HBM and additional equipment classes, ended blanket validated-end-userA US export-control status that let an approved facility receive controlled equipment under a standing authorisation instead of applying licence by licence. Withdrawing it does not ban shipments; it makes every one of them an individual decision, with the delay and uncertainty that implies. In practice that is often the more effective restriction. treatment for foreign-owned fabs in China, and in January 2026 moved H200- and MI325X-class exports to case-by-case review under stated conditions. Enforcement is not theoretical: the Bureau of Industry and Security announced a $252 million penalty against Applied Materials tied to illegal equipment exports. The net effect is a smaller accessible near-term market for Western tool vendors and a stronger incentive for Chinese fabs to localise mature equipment — a subsidised long-run competitor in the less advanced categories, and no path at all to a second EUV supplier.

I·2The economics of a leading-edge fab

A leading-edge fab is a machine for converting enormous fixed capital into yield learning. Three properties of that conversion explain why the industry ended up with one dominant merchant player, and why its customer list is so short.

One word in that sentence carries the whole structure: merchant. A captiveA facility that exists to serve its own parent — the designer and the manufacturer are the same company, and the factory's output is not for sale to anyone else. The opposite of merchant. The word recurs through this report because the same physical asset earns completely differently depending on which of the two it is. fab exists to make its owner's own chips; the designer and the manufacturer are one company. A merchant foundry designs nothing and manufactures for whoever pays, which means it has to keep an enormous plant full on other people's products. That is the harder business to run and much the stronger position to hold, because every designer on earth becomes a potential customer rather than a rival — and it is why the three properties below produce concentration rather than a crowd.

First, the capital is lumpy and rises with every nodeA manufacturing generation, labelled by a number — 5nm, 3nm, 2nm. The number stopped being a physical measurement of anything years ago; it is now a marketing name for a generation, and what matters is the density, speed and power it delivers relative to the one before. Each new node costs more to develop, takes longer to yield, and serves fewer customers than the last.. TSMC guided 2026 capital expenditure to $60–64 billion, raised from a prior $52–56 billion range, with 70–80% earmarked for advanced process technologies and 10–20% for advanced packaging, testing, masks and other uses. In the June 2026 quarter alone it spent NT$496.00 billion — roughly $15.5 billion at its own NT$32/US$ planning assumption — against NT$783.36 billion of operating cash flow. That is a business converting two-thirds of its operating cash into the next node before anything reaches a shareholder. Depreciation begins before mature yield arrives, which is why the reported margin of a fab at the top of a ramp is partly the accounting face of continuous reinvestment rather than a measure of pricing power.

Second, yield is learned, not designed. Defect density on a new node starts too high for economic production and falls along a curve driven by cumulative wafer starts. Whoever runs the most wafers earliest descends the curve fastest, which lowers cost per good die, which wins the next customer, which supplies more wafers. It is a positive feedback loop, and it is why leading-edge foundry concentrates rather than fragments. It is also the specific thing that money cannot buy: a competitor can purchase the same class of tool, and cannot purchase a mature yieldThe share of chips on a wafer that come out working. Everything on a wafer costs the same to make whether it works or not, so yield is the whole of the cost equation: at low yield the good chips carry the price of the bad ones. Yield on a new generation starts too low to sell and improves as the factory accumulates experience, which is why running the most wafers earliest is a genuine and unbuyable advantage. curve, a set of reference flows and librariesThe prepared building blocks and step-by-step design procedures a foundry supplies so that customers can design chips that will actually manufacture well in its factory. They accumulate over years and across many customers. A new entrant with identical machines still has to write all of this, and until it does, designing for its factory is slower and riskier for the customer., and dozens of taped-outThe moment a chip design is finished and handed to the factory to be made — the point of no return, after which changes mean starting the manufacturing sequence again. The phrase survives from the era when designs were literally delivered on magnetic tape. customer designs on day one.

Third, the customers who can afford the node are few. TSMC disclosed that 2-nanometre accounted for 3% of wafer revenue in the June 2026 quarter, 3-nanometre 30%, 5-nanometre 33% and 7-nanometre 11% — 77% of wafer revenue from 7nm and below, with high-performance computing at 66% of total revenue. A mask set at the leading edge plus the design effort behind it runs into the hundreds of millions of dollars, which restricts the node to products with enormous unit volume or enormous unit price. AI accelerators are the second kind.

Put the three together and the customer concentration stops looking like an accident of this cycle. It is the structure. It also cuts both ways, which is the part the pricing-power story usually omits. Only customers with sufficient unit economics can fund leading-edge masks, engineering and inventory risk — and the foundry needs anchor customers to load the first capacity on a node that is losing money until it yields. That bilateral dependence restrains price even where the node leader has no close substitute. A hyperscalerThe handful of companies that operate data centres at a scale nobody else approaches — Amazon, Microsoft, Alphabet, Meta and a small number of others. The word matters here because they are simultaneously the largest customers on this chain, the designers of their own competing chips, the authors of the open standards that commoditise their suppliers, and the counterparties to twenty-year power contracts. Almost every argument in this report turns on something one of them decided to do. that is simultaneously funding Intel Foundry and Samsung is not a permanent price-taker.

The outcome is now visible in the accounts. TSMC's gross margin reached 67.7% in the June 2026 quarter with operating margin at 60.3%, on revenue of $40.20 billion, up 33.7% year over year in dollars; it guided the September quarter to $44.6–45.8 billion at a 65–67% gross margin. Monthly reporting corroborates rather than settles it: first-half 2026 revenue was NT$2.404 trillion, 35.6% above the prior year, with June alone at NT$442.7 billion, up 67.9%. Management's own long-term framing is the useful cross-check — a stated long-term gross-margin target of at least 56% and a long-term return-on-equity target in the high twenties, both well below the current print, which is what a company says when it knows the current number contains a cycle.

Two costs are already disclosed against that margin, and they are the reason the return matters more than the margin here. Overseas fabs were expected to dilute gross margin by 2–3 percentage points early and 3–4 points later; the N2 ramp itself costs 3–4 points. Both are the price of a strategic and political hedge rather than an economic one. Sovereign money is the counterweight to everything above: governments in the United States, Japan, Europe, Korea and China are subsidising competing capacity for reasons that have nothing to do with return on capital. Subsidised capacity does not need to earn its cost of capital to take volume, which is exactly why it is dangerous to an incumbent's pricing — and exactly why it tends to arrive at trailing nodes rather than the leading edge, where the constraint is know-how rather than money. What subsidy reliably does is lower a competitor's cost of capital and depress the return on everybody's incremental mature-node capital.

Intel is the live test of whether that can be overcome. Its June 2026 quarter delivered $16.1 billion of revenue, up 25% and its strongest growth in more than fifteen years, with a GAAP loss of $2.16 per share against $0.42 of non-GAAP earnings. Growth has returned to Intel. Returns have not, and that gap is the foundry verdict in one company. Samsung sits in a different position again: its integrated memory, logic and packaging stack is strategically powerful and, at Q2-2026, extraordinarily profitable — group revenue of KRW 171.5 trillion and operating profit of KRW 89.5 trillion, both all-time highs — but the profit is overwhelmingly memory, and its foundry improvement was attributed to HBM base-die demand rather than to external logic wins. GlobalFoundries and UMC own useful specialty processes and long-lived customer qualifications; both are trailing-node specialists whose margins say so, and both have found genuine AI adjacency in silicon photonics rather than in accelerator logic.

A 55% foundry gross margin against generational capital expenditure is a different animal from a 55% software margin against cloud operating expense. Same percentage. Different claim on the future.

The distinction this report applies at every node

I·3Why packaging became the bottleneck

The binding constraint on how many accelerators the world can build stopped being wafers some time ago. It is now the step that glues the logic die to its memory.

An AI accelerator is not one chip. It is a large logic dieA single rectangle of silicon cut from a wafer — one chip, before it is packaged. Wafers are round and dies are rectangular, so bigger dies waste more of the edge and, worse, catch more defects: doubling the area of a die more than doubles the chance that something on it is broken. That arithmetic is why very large chips are now built as several smaller dies joined together. — or several, stitched together — sitting on a silicon interposerA slab of silicon that acts as a very dense circuit board. The logic chip and the memory stacks are mounted on top of it side by side, and it carries the tens of thousands of fine wires running between them. Because it is itself made of silicon on a wafer, it consumes chipmaking capacity in proportion to how big the accelerator is rather than how many transistors it holds — a different constraint, and the reason packaging capacity became its own bottleneck. alongside stacks of high-bandwidth memory, the whole assembly mounted on an organic substrateThe layered board the finished silicon assembly is fixed to, which carries power in and signals out to the rest of the system. "Organic" simply means it is made of resin and copper rather than silicon — cheaper and coarser than the interposer above it, and made by a different set of suppliers again.. ReticleThe maximum area a lithography machine can print in a single exposure — a hard physical ceiling set by the optics, not by ambition. A chip cannot be larger than the reticle field. Once designers wanted more silicon than that, the only way forward was to build several dies and join them, which is how packaging became an architectural decision rather than a finishing step. limits and yield economics encourage this: a very large monolithic die yields badly and cannot exceed the printable field, so the system is split into compute dies, input/output dies and memory stacks and then reunited. The interposer carries tens of thousands of micron-pitch wires between logic and memory, because the entire performance argument of an accelerator collapses if the processor cannot be fed.

That last clause deserves its own paragraph, because it inverts the intuition most people bring to computing. An AI chip spends much of its life waiting for numbers rather than working on them: the arithmetic is fast and cheap, and hauling data to the arithmetic is slow and expensive. So the memory sitting beside the processor, and the wiring between the two, decide how much of the silicon actually does anything — more often than the processor itself does. That is why a packaging step almost nobody outside the industry had heard of became the binding constraint on a trillion-dollar buildout, and why the memory chapter that follows is not a digression.

TSMC's implementation is called CoWoS, chip-on-wafer-on-substrate; the three-dimensional variant that stacks logic on logic is SoICSystem on Integrated Chips — TSMC's method for bonding one logic die directly on top of another, face to face, with no solder bumps between them. Stacking vertically shortens the distance signals travel still further than sitting side by side does; the price is that heat from the lower die now has to escape through the upper one.; Intel's is EMIBEmbedded Multi-die Interconnect Bridge — Intel's alternative to a full silicon interposer. Instead of mounting everything on one large silicon slab, small silicon "bridges" are buried in the cheaper organic substrate only where two dies need a dense connection. Less silicon consumed, more assembly complexity..

Three things make this the choke. The interposer is itself a large silicon die, so interposer area consumes wafer capacity that scales with accelerator size rather than with transistor count — a physically different constraint from the wafer bottleneck it replaced. Yield is multiplicative across a stack of already-expensive components, so a single packaging defect scraps a bill of materials worth thousands of dollars, and the known-good-dieThe requirement that every component going into an expensive assembly has already been proven to work, because a single bad one ruins everything joined to it. It sounds obvious and is expensive: it pushes testing earlier and deeper into every supplier's process, and it is why the memory maker's yield problem becomes the packager's problem and then the accelerator vendor's. requirement propagates back into the memory maker's process. And capacity additions require clean-room, bonderThe machine that physically joins one chip to another, aligning contacts finer than a red blood cell is wide and fixing them permanently. It is the slowest and most specialised piece of equipment in advanced packaging, which is why "just add capacity" is measured in years. and reliability-qualification lead times of eighteen months to two years, which is why the constraint did not clear when demand doubled. Add thermal warpage, micro-bump and hybrid-bond quality, and final test, and every one of them becomes binding at once.

Who captures the packaging profit is the one place where the three independent readings of this chain genuinely disagreed. The naive answer is that as capacity expands the profit migrates to the outsourced assembly and test providers — Amkor, ASE — who have historically done packaging for everyone. Both cases follow.

The reconciliation is that both are describing true things about different quantities. The shortage premium in packaging is temporary and is being competed away on a 2027–28 clock as TSMC's own capital and OSAT qualification both land. The co-optimisation advantage — the fact that the package is designed with the process rather than after it — is not obviously temporary, because it is the same yield-learning mechanism that defends the wafer. So the verdict carried in Part II is a mixed one with a numeric test attached: if a merchant OSAT reports leading-edge packaging operating margin above 20% by 2028, the co-optimisation argument is wrong and the foundry verdict weakens with it, because packaging is where the foundry's incremental pricing power now sits.

Advanced packaging raises test intensity — more interfaces, more known-good-die gates, longer test times per unit of shipped performance — and test is a high-return toll on complexity even where assembly itself is competed. Advantest's FY2025 sales rose 44.7% and operating income 118.8%, producing a 44.2% operating margin and a 57.6% return on equity, on a business with modest capital intensity and heavy platform reuse. Teradyne's June 2026 quarter revenue rose 104% year over year to $1,329 million with $1,122 million in semiconductor test and record memory test revenue. FormFactor's probe-card business grew 31.9% on HBM and co-packaged-optics demand. The pattern repeats across the report: when a physical step gets harder, the profit tends to migrate to whoever prices the cost of failure rather than to whoever performs the step.

I·4High-bandwidth memory, and what memory cycles do

High-bandwidth memory is DRAM stacked vertically. The manufacturing is genuinely hard, the qualification is genuinely sticky, and neither of those facts has ever repealed the incentive that ends every memory cycle.

Eight, twelve or sixteen dies are thinned, drilled through with copper through-silicon viasVertical copper shafts bored straight through a chip so that the one above can talk to the one below. They are what make a stack of memory a single device rather than eight separate ones. They also consume area on every layer they pass through, which is part of why stacked memory costs more silicon per unit of storage than the flat kind., bonded into a tower and given a base logic dieThe chip at the bottom of a memory stack. It is not memory at all — it is a logic chip that manages the stack and handles the conversation with the processor. As it has grown more capable it has moved onto advanced chipmaking processes, which quietly transfers a slice of the memory bill from a memory maker to a foundry. that talks to the accelerator across a very wide, very short busThe set of parallel wires data travels along between two components. "Wide" means many wires side by side carrying bits simultaneously; "short" means the distance is small enough that signals stay clean and little energy is wasted pushing them. Both are the whole point of putting memory next to the processor rather than across a board from it.. The physics is straightforward and the manufacturing is not: thinning a wafer to tens of microns without cracking it, drilling and filling tens of thousands of vias, bonding stacks with sub-micron alignment, and getting heat out of a structure that is essentially a thermal blanket over its own hottest layer.

Three economic consequences follow. HBM consumes far more wafer area per usable bit than commodity DRAM, because of the die-area overhead of the vias and the yield loss across a stack — so every bit of HBM demand removes more than one bit of commodity supply, which is why the two prices move together. QualificationThe formal process by which a buyer proves a specific supplier's specific part works in its specific product, at volume and over time. It takes months and costs real money, and until it is finished the supplier cannot ship a single unit. Qualification is the quiet source of most of the switching costs in this report: not loyalty, not contracts, just the fact that changing suppliers means doing the whole thing again. is long and customer-specific and co-designed with a particular accelerator, which creates a genuine switching cost for the length of one generation. And the value added by stacking is capturable: the same silicon sells for a multiple of its commodity price once it is inside a qualified stack.

That is a real difference from commodity DRAM. It is not a moat. Three suppliers can do it, all three are expanding, and the qualification advantage lasts exactly one generation — which is now visibly turning over. SK hynix reported second-quarter 2026 revenue of KRW 79.32 trillion and operating profit of KRW 60.54 trillion, a 76.3% operating margin, and disclosed that it began mass shipments of HBM4High-bandwidth memory, fourth generation — DRAM dies stacked vertically and wired straight into the accelerator package, with more stacked layers and a wider, faster interface than HBM3. HBM4E is the mid-generation extension that follows it. Each generation has to be qualified afresh with the accelerator vendor, which is why the qualification advantage described here lasts one generation and no longer. in the quarter with a full ramp in the second half, that HBM4E samples went out in the first half, and that it has finalised long-term agreements with around ten customers. Micron's fiscal third quarter, ended 28 May 2026, produced $41.46 billion of revenue against $9.30 billion a year earlier, with an 84.6% gross margin and 80.4% operating margin, low-60s sequential DRAM price growth and mid-80s NAND price growth, HBM4 already shipping in volume to a lead customer, multiple further qualifications in flight and HBM4E volume planned for calendar 2027. Samsung began commercial HBM4 shipments in February 2026, scaled HBM4 sales through the first half, shipped the industry's first HBM4E samples, and expects supply constraints to continue through the second half; its Device Solutions division earned KRW 89.2 trillion of operating profit on KRW 127.5 trillion of revenue, a 70.0% operating margin.

All three are qualified or qualifying into HBM4 at once. That is better product differentiation than any prior DRAM cycle offered — and it is also precisely the condition that finances the next supply response, because a supplier that is about to lose a socket to a second qualified source has every reason to ship the next wafer.

Arithmetic, not opinion, settles how much of memory's profit pool is price rather than volume. Hold Micron's reported quarterly operating expense of $1.738 billion constant — research and development of $1,316 million, selling and administrative of $407 million, other operating of $15 million — and flex revenue and gross margin across the range the company itself has printed in the last four quarters.

Memory cycles have resolved the same way for four decades and the mechanism has never varied. Shortage produces price. Price produces margin. Margin funds capacity. Capacity lands two to three years later, usually within a few quarters of every competitor's capacity, because they all read the same price signal at the same time. Then price finds marginal cash cost. The only question that has ever mattered is when the capacity lands.

On that question this cycle has an unusually well-documented answer, disclosed by the suppliers themselves. Micron's Idaho fab begins DRAM production in mid-2027; its Taiwan output contributes from mid-2027 and its Singapore HBM packaging in the first half of 2027; its New York site broke ground only in January 2026 and Micron's chief executive has said meaningful large-scale new capacity is unlikely before 2028. SK hynix is accelerating M15X, opening Yongin Phase 1 in early 2027 and adding EUV capacity. Samsung has announced an HBM fab and a long-horizon semiconductor investment programme explicitly conditioned on demand. So the shortage is genuinely multi-year — and the date at which the test begins is knowable: roughly the second half of 2027, with the fuller weight in 2028.

Two things partially offset that, and they are the strongest version of the bull case. The first is contractual. Micron disclosed roughly $100 billion of remaining performance obligations across sixteen strategic customer agreements carrying minimum price and minimum volume over multi-year terms, plus $22 billion of customer deposits and related commitments; SK hynix has locked around ten customers into long-term agreements. Those are a materially different instrument from spot memory, and they will hold a floor under revenue through the first phase of the supply response. The second is that HBM's wafer-per-bit penalty means a given quantity of new wafer capacity delivers fewer incremental bits than it would in a commodity cycle, which stretches the timeline. Neither changes the mechanism. Both change the amplitude and the lag — which is exactly what the verdict in Part II says.

One line of Samsung's disclosure breaks the neat picture of memory as a separate layer. Samsung attributes its foundry improvement to HBM base-die demand — because the logic die at the bottom of an HBM4 stack is not memory at all. It is a logic chip, built on a foundry process, and with HBM4 it became large and fast enough to be worth building on a leading node. Each generation of high-bandwidth memory therefore transfers a slice of the memory bill of materials into the foundry layer, where it is priced by a foundry rather than by a memory maker. When memory's price cycle turns, the base-die revenue does not turn with it, because it was never memory revenue. It is one instance of a pattern this report keeps finding: value migrates upstream of the layer under pressure, not away from it.

The storage tier below memory is running the same play with a longer fuse. SanDisk's fiscal third quarter produced $5.95 billion of revenue, up 97% sequentially, with GAAP net income of $3,615 million — a 61% net margin on a NAND flash business — datacenter revenue up 233%, and guidance to $7.75–8.25 billion for the following quarter. Seagate closed fiscal 2026 with $12.2 billion of revenue and a 52.3% GAAP gross margin in the June quarter, $3.1 billion of free cash flow and $1.4 billion of debt retired. Western Digital's March quarter grew 45% at a 50.2% gross margin with the following quarter guided up 36–44%. Spinning-disk drives, a business the market wrote off as a melting ice cube, are earning foundry-like gross margins because nearline capacity is the cheapest place to park training corpora and inference caches and nobody built for this. Nearline HDD has a genuinely tighter industry structure than NAND — fewer suppliers, longer qualification — but its buyers are the same three or four hyperscalers, and NAND has the same brutal bit-growth arithmetic as DRAM. Storage is a valuable volume tail. It is not an automatic AI moat.

I·5How a training cluster is actually assembled

The unit of AI infrastructure is a rack now, not a server — the single most important change in the physical build, because it moves the integration boundary and therefore the margin.

Work upward. An accelerator die plus its HBM stacks is packaged into a module — SXM, OAM, or a vendor-specific form. Modules mount on a board with power delivery and a cold plateA metal block clamped directly onto a hot chip with liquid coolant running through channels inside it. It is the point where heat leaves the silicon and enters the plumbing, and it exists because the chips in question now produce more heat per square centimetre than air can carry away.. Boards go into trays; trays go into a rack; and inside the rack, a scale-up fabricThe high-speed interconnect binding a group of accelerators together — and, more importantly, the design decision behind it about which chips can talk to which, how fast, and with how much delay. Calling it cabling misses the point: the fabric decides how much of the installed silicon is doing useful work at any given moment, which makes it an architectural choice with an enormous bill attached. connects every accelerator to every other at very high bandwidth so that the group behaves like one large processor. NVIDIA's implementation is NVLink with NVSwitchNVIDIA's proprietary scale-up fabric. NVLink is the high-speed link between accelerators; NVSwitch is the switching chip that lets every accelerator in a rack reach every other one directly rather than through a chain of neighbours. Together they are what turns a rack of chips into something software can address as one machine — and, until recently, they were available from exactly one company.; the current generation connects 72 accelerators all-to-all in a single domain. Beyond the rack, a scale-out network — Ethernet or InfiniBandA high-performance networking standard long used in supercomputers, with lower delay and more predictable behaviour than ordinary Ethernet, and a much smaller set of suppliers. The live question in this layer is whether AI clusters keep paying for it or whether Ethernet, improved to handle the same traffic, takes the volume — as it has taken the volume in every previous contest of this kind. — connects racks into a cluster of thousands, and carries traffic to storage.

Two of those words do a great deal of work and are worth slowing down for. Scale-up means making one machine bigger: wiring a group of accelerators together so tightly that software can treat the whole group as a single enormous processor. Scale-out means making more machines and connecting them: thousands of racks joined into a cluster, with far more delay between racks than inside one. A large model has to be split across many chips that constantly exchange intermediate results, so where the boundary between the two sits determines how much waiting the silicon does — and the boundary is also, physically, where copper stops and light begins.

The split between scale-up and scale-out dictates the bill of materials, and this is where interconnect topology becomes an economic variable rather than an engineering one. Inside the rack, distances are short enough for copper, which is cheap; between racks, distances demand optics, which are not. Model parallelismThe set of choices about how to cut a model that is too large for one chip into pieces spread across many — by layer, by the internal arithmetic of each layer, or by feeding different batches of data to different copies. Each choice produces a different pattern of chatter between chips, which is why an apparently abstract software decision ends up determining how much networking hardware a data centre has to buy. strategy determines how much traffic crosses each boundary, and therefore how many transceiversThe modules that convert electrical signals into light and back again at each end of an optical fibre. They are sold by the million, replaced at every speed generation, and are the classic example on this chain of a component that is indispensable, technically demanding, and repeatedly stripped of its margin., retimersSmall chips that clean up and re-transmit a high-speed signal that has degraded travelling along a wire. As speeds rise, signals survive shorter distances, so more of these are needed in more places — a component whose demand grows out of a physical problem getting worse rather than out of anyone choosing to buy more., switches, connectors and cables a given megawatt of compute needs. A change in topologyThe map of how everything is wired together — which components connect to which, and through how many hops. Two clusters with identical chips and identical software can need very different quantities of cable, optics and switching depending on the topology chosen, which is why the layout decision is also a procurement decision. can move billions of dollars between vendors without a single chip changing hands.

When NVIDIA moved from a board to the GB200/GB300 NVL72 and Blackwell Ultra systems, it moved the unit of competition from silicon toward the rack — and by 21 July 2026 it had disclosed that Vera Rubin NVL72 was ramping into production with racks running at CoreWeave, Google Cloud, Microsoft, Oracle and Nebius. The annual cadence strengthens performance leadership while shortening the buyer's economic life, which is a genuine two-edged instrument: it makes the platform harder to catch and it makes every customer more anxious about depreciation and dependence.

The accelerator's gross margin does not pay for the customer's building, its grid connection, or most of the depreciation on the deployed system. NVIDIA can earn a fablessA chip company that designs but does not manufacture, leaving the factories to a foundry. The consequence is financial rather than technical: a fabless business needs remarkably little capital of its own, so a given amount of profit sits on top of a much smaller pile of assets — which is most of why its return on capital looks nothing like a chipmaker's, at the same margin. return on architecture and software while cloud providers carry accelerator obsolescence and the debt behind it. NVIDIA's April 2026 quarter produced $81.6 billion of revenue including $75.2 billion from Data Center, a 74.9% GAAP gross margin and a 65.6% operating margin, $48.6 billion of free cash flow — and $12.4 billion of net property and equipment against $259.5 billion of total assets. Its full fiscal 2026 produced $215.9 billion of revenue, up 65%. The depreciation schedule for the thing it sells sits on somebody else's balance sheet. That single fact does more to explain the return-on-capital map in §II·8 than any margin comparison.

Three groups do the assembly: the branded OEMs (Dell, HPE, Super Micro), the merchant ODMs and electronics manufacturing services providers (Celestica, Flex, Jabil, Sanmina, plus Hon Hai, Quanta Computer and Wistron in Taiwan), and increasingly the hyperscalers themselves, buying boards and building racks in-house. The structural fact that decides the margin is that the customer usually specifies the design, the accelerator is a single-source component bought at a price the assembler does not set, and several qualified assemblers bid for the work. That combination has exactly one possible outcome.

The numbers are unusually consistent across the layer. Dell's Infrastructure Solutions Group generated $29.0 billion of revenue in the quarter ended 1 May 2026, up 181%, with $16.1 billion of that in AI-optimised servers, $24.4 billion of AI orders booked, and $3.1 billion of operating income — a 10.7% segment operating margin, and the ceiling for the layer. Celestica's second quarter produced $4.70 billion at a 9.8% GAAP operating margin. Super Micro reported $10.2 billion of net sales at a 9.9% gross margin, down sequentially from $12.7 billion. Sanmina earned 6.4%. Jabil 5.1%. Flex 4.9%. And Quanta Computer — the largest rack integrator in the world by AI-server share — grew first-quarter revenue 66.6% to a record NT$809.2 billion with AI servers above 75% of the mix, while gross margin compressed to 4.78%, management attributing the compression to the cost of full-rack integration. Growth of two-thirds, at under five cents of gross margin on the dollar, is the definition of activity without value capture. Hon Hai's own numbers say the same thing from a larger base: 6.18% gross and 3.57% operating margin in Q1 2026, a 44-day cash-conversion cycleHow long a company's money is tied up between paying for materials and being paid for the finished product, measured in days. Add the time inventory sits on the shelf to the time customers take to pay, then subtract the time the company takes to pay its own suppliers. A short cycle means the supply chain is financing you. A long one means you are financing it — and the faster you grow, the more cash that consumes., and negative NT$32.6 billion of free cash flow in the quarter, with cloud and networking products now the largest of its four product lines.

Super Micro's cash statement generalises. Growth in this layer consumes working capital roughly in proportion to revenue: the assembler buys the accelerator, holds it, integrates it, and then waits to be paid. Super Micro burned $6.6 billion of operating cash in the quarter, ending with $1.3 billion of cash against $8.8 billion of bank debt and convertible notes. In a rising market that is a financing problem disguised as a growth story. It is also why the layer's accounting return on capital can look strong while its cash generation is negative — the float works for you until the day the orders stop, and then it works against you very fast. §II·8 shows the two ends of that mechanism side by side, in the same layer, in the same quarter.

Rack-scale liquid integration is the current wedge: a real skill, and a temporary one. It is valuable while leak testing, thermal commissioning, quick-disconnect qualification and supply coordination are immature. As reference architectures, standard manifolds and blind-mateConnectors — for power, data or coolant — that seal correctly when a component is pushed into place without anyone seeing or adjusting them. They make racks serviceable by ordinary technicians instead of specialists. They also make the components interchangeable, which is exactly why publishing a blind-mate standard is a commercial act as much as an engineering one. interfaces stabilise — and the Open Compute ProjectAn industry body, founded and steered largely by the companies that buy data-centre hardware, which publishes open reference designs for racks, servers, cooling and now electrical distribution. Its function is to turn whatever the buyers are paying most for into something several suppliers can build to the same drawing. It is the buyers' long game, conducted in public. is publishing exactly those — it becomes manufacturing know-how shared across Dell, HPE, Super Micro, Celestica, Hon Hai, Quanta and Wistron. The durable value migrates upstream to the architecture and downstream to controls, commissioning and service, leaving the assembly step where it has always been.

I·6Networks, optics, and what happens when Ethernet arrives

Every proprietary interconnect in the history of computing has eventually met an open standard. The standard has taken the volume; the specialist has kept the frontier; and the margin has moved from the link to the software that manages it.

The AI cluster has two networks and they have different economics. The scale-up fabric inside a rack is latency- and bandwidth-critical and today is dominated by NVIDIA's NVLink, with two open alternatives organising against it: UALinkAn industry consortium writing an open standard for the fabric inside a rack — the job NVLink does today. Its purpose is to give buyers a second source for the one part of the system that currently has none. Whether a published specification becomes shipping silicon is a separate question, and this report treats the two as separate events., whose 200G 1.0 specification supports a scale-up domain of as many as 1,024 accelerators, and a scale-up Ethernet approach championed by Broadcom and formalised through the Open Compute Project. The scale-out network between racks is Ethernet almost everywhere, with the Ultra Ethernet ConsortiumThe body adapting ordinary Ethernet to the traffic AI clusters generate — enormous synchronised bursts, where one late message stalls thousands of chips. It is the standards route to displacing InfiniBand, and half its steering members are the companies that buy the equipment rather than the ones that sell it. standardising the transport and congestion-controlThe rules a network uses to decide who sends when, so that a burst of traffic degrades gracefully instead of collapsing. Ordinary internet traffic tolerates a dropped packet and a retry. AI training does not: the chips wait, and the whole cluster idles — which is why the congestion rules, not raw speed, are the hard part of this standard. behaviour that AI traffic needs.

All three consortia moved during the period covered, and the direction of travel matters more than any single release. Ultra Ethernet is now at specification 1.0.3, last updated 17 July 2026, with a steering group of AMD, Arista, Broadcom, Bull, Cisco, HPE, Intel, Meta, Microsoft and Oracle — five of those ten being the buyers of AI networking rather than the sellers of it. UALink published four further specifications in April 2026, extending the 200G base into in-network compute, chiplets and manageability, and added Alibaba, Apple and Synopsys to its board. The Open Compute Project released Ethernet Scale-Up Networking 1.0 in March 2026 with participation across 175 companies, and then on 29 April 2026 approved eight further contributions.

Among the eight were an Open Rack Wide specification, a data-centre low-voltage DC power distribution architecture, requirements for data-centre energy storage, and ESUN, a network-operator requirement set for using Ethernet as the scale-up fabric — alongside new alliances with the Electric Power Research Institute and the IOWN Global Forum. A consortium whose members are the buyers has begun publishing open reference designs not only for the network but for the rack's electrical architecture: the power shelves, the busbar, the distribution voltage. The buyers are doing to the electrical layer exactly what they did to the switch. What that does to margin is the subject of the rest of this section, and the reason §II·7 ranks rack electrical architecture as the newest and least-priced internalisation front.

NVIDIA's counter is instructive and should be read as strategy rather than generosity. Rather than defend NVLink as a closed system, it opened it — NVLink Fusion licenses the interconnect to third-party CPUs and custom accelerators, with MediaTek, Marvell, Alchip, Astera Labs, Synopsys and Cadence named among early adopters. A monopolist that licenses its interconnect is choosing to be the standard rather than to be displaced by one. That is the correct move, and it is also an admission that the closed position was not going to hold.

Where does the profit actually sit once a layer becomes interoperable? History is unambiguous on the mechanism. Once a layer is standardised, buyers can separate the hardware decision from the management plane and qualify alternates, and vendor gross margins compress toward a merchant-siliconChips sold on the open market to anyone who wants to build a product around them, as against chips a systems vendor designs exclusively for its own boxes. Once merchant silicon is good enough, the systems vendor's hardware advantage evaporates and whatever margin survives has to come from software, service or qualification instead.-plus-ODM equilibrium. What survives is the point where failure is costly and replacement is not drop-in. Three such points exist on this chain, and their disclosed margins say so.

The first is network operating software and the operational trust around it. Arista grew first-quarter 2026 revenue 35.1% to $2,709 million at a 42.7% GAAP operating margin with $1.69 billion of operating cash flow — sitting on merchant silicon it does not make, and keeping the margin anyway, because what the customer is buying is the software estate and the operational record. Compare the incumbent the boom passed by: Cisco's April 2026 quarter grew 12% to $15.8 billion at a 63.6% GAAP gross margin and a 25.0% operating margin. Respectable, high-margin, and growing at a third the rate of the merchant vendors around it. Same layer, same customers, opposite trajectories.

The second is signal-integrity silicon that must be qualified into a specific link budget. Every speed step — 400G to 800G to 1.6T — worsens the loss budgetThe total amount of signal degradation a link can tolerate and still work. Every connector, every centimetre of copper and every bend spends part of it. Doubling the data rate roughly halves the distance a signal survives, so each generation forces more corrective silicon into more places — a market that grows because physics is getting less forgiving, not because anyone chose to buy more., and the parts that fix it acquire a design-in position that survives a generation. Credo's fiscal fourth quarter grew revenue 157% year over year to $437.0 million at a 68.2% GAAP gross margin, with full-year revenue more than tripling to $1.3 billion and non-GAAP net income up more than fivefold. Astera Labs reported $308.4 million, up 93%, at a 76.3% GAAP gross margin — but a 20.1% GAAP operating margin against 36.2% non-GAAP, the gap being stock compensation, which is a real cost of retaining the engineers the moat consists of. That gap is worth naming rather than adjusting away: in a business whose asset is a design team, non-GAAP margin systematically overstates what the owner keeps.

The third is mechanical: connectors and cable assemblies designed into a specific platform. Amphenol reported $8.8 billion of second-quarter sales, up 55% and 30% organically, orders of $10.7 billion for a 1.23 book-to-billOrders taken divided by revenue billed in the same period. Above 1.0 the order book is growing faster than the company can ship, which is usually read as strength. It says nothing whatever about the margin those orders carry, and several companies in this report are booking at well above 1.0 while their margins fall., a 29.5% GAAP operating margin and $1.2 billion of free cash flow. A connector is not glamorous and it is qualified into a mechanical platform for the life of that platform, across a decentralised manufacturing system that competes on qualification breadth rather than on unit price. Corning sits adjacent with specialty fibre and glass science: Q2-26 core sales of $4.74 billion, up 17%, with Optical Communications at $2.07 billion up 32% and Enterprise Networks up 65%, plus a multiyear multibillion-dollar agreement with Amazon for optical fibre and cable. Fibre supply has historically expanded into a lower price per bit, which is the counterweight to that franchise and the reason its grade split the panel.

Then there is the part of the layer that has always given it back. Transceivers are a manufactured good with a repeating generational cadence in which every generation resets the competitive field, capacity gets added into the peak, and margins compress on the trailing generation while the leading one is still ramping. That pattern has not been repealed; it has been delayed by the sheer rate of demand growth. The dispersion inside a single quarter proves the layer is not uniformly advantaged: Coherent's March 2026 quarter delivered $1.81 billion, up 21%, at a 37.7% GAAP gross margin — up 243 basis points, and still the lowest of the optics names read here. Lumentum delivered $808.4 million, up 90%, at a 44.2% GAAP gross margin and a 21.6% GAAP operating margin, with management crediting mix and "pricing discipline." Fabrinet reported record revenue and earnings above guidance on ramping datacom programmes, in a structurally fee-for-assembly model. And Applied Optoelectronics, in the same market and the same quarter, described its results merely as "in line with our expectations." One layer, one quarter, four completely different outcomes.

Co-packaged opticsMoving the light-generating hardware off the front panel of a switch and into the same package as the switching chip itself. The signal then travels a few millimetres of silicon instead of tens of centimetres of copper, which saves a great deal of power. The commercial consequence is that a part currently bought as a separate replaceable module becomes something built into the package — shifting the work, and the money, from the module maker to the packaging and test houses. is the change to watch here, and its effect is directional rather than datable. Moving the optical engine into the package shifts value toward optical engines, advanced packaging and the foundries and test houses that qualify them, while shrinking the discrete pluggableA transceiver built as a separate module that slots into the front of a switch and can be pulled out and replaced by hand. Being replaceable is exactly what makes it a competitive, low-margin product: the buyer can second-source it, and does. pool that currently finances transceiver capacity. GlobalFoundries' SCALE co-packaged-optics platform, launched May 2026 with partnerships including Corning, is the trailing-node specialist's most credible AI adjacency for exactly that reason. FormFactor's co-packaged-optics probe demand is the test-side expression of the same shift. Neither yet appears as a named revenue line anywhere, which is why the grades that touch it carry explicit disclosure falsifiers.

I·7The thermodynamics that ended air cooling

A rack that dissipates 120 kilowatts cannot be cooled with air, and the reason is not engineering preference. It is the volumetric heat capacity of air, which is roughly four thousand times worse than water.

Removing 120 kW with a 10°C air rise requires a volumetric flow rate that is physically impossible to move through a rack-sized aperture without noise, vibration and fan power that consume a meaningful share of the energy you were trying to spend on compute. Air stops being a primary strategy somewhere between 40 and 80 kW per rack depending on climate and hall design. AI training racks are already past 100 kW and the roadmaps point higher.

So the heat goes into liquid. Cold plates sit directly on the accelerator packages; a coolant distribution unitKnown as a CDU. The pump-and-heat-exchanger box that stands between the clean coolant circulating through the chips and the building's own water system, keeping the two loops separate while heat passes from one to the other. It also controls flow and pressure, which is the part that turns out to be hard, because rack loads swing violently as workloads start and stop. circulates fluid through the rack and rejects heat to a facility water loop; the facility loop rejects it to the outside through dry coolers, cooling towers or a chiller. Each of those is a well-understood piece of mechanical engineering. The hard parts are the system: leak-free quick-disconnect fittings by the thousand, coolant chemistry and material compatibility over a ten-year life, flow balancing across non-uniform and rapidly transient heat loads, and a service model for a fluid loop running through live electronics.

Value therefore divides three ways, and only two of them are defensible. Manifolds, hoses, cold plates and CDUs are manufacturable hardware. System design and controls must match transient rack loads to facility conditions. Installation, commissioning and service keep the plant available. The commercial question is whether the system knowledge is a franchise or a head start, and the evidence in a single quarter is genuinely mixed — which is itself the finding.

Vertiv, the closest thing to a pure play and the one name that spans critical power, thermal, controls and service, reported second-quarter net sales of $3,274 million, up 24% and 18% organicallyGrowth from the business as it already stood — stripping out anything added by acquisition and any flattering or punishing move in exchange rates. It is the honest read on whether the underlying operation is actually getting bigger, which is why a wide gap between headline and organic growth is worth looking at., with adjusted operating marginOperating margin after management has removed items it considers unrepresentative — restructuring charges, acquisition accounting, sometimes share-based pay. Sometimes the adjustments genuinely clarify; sometimes they remove a real and recurring cost. This report quotes the reported figure wherever it can and labels the adjusted one as adjusted whenever it cannot. of 22.6%, up 410 basis points, $1,100 million of operating cash flow, a net cash position reached, and full-year guidance raised to about $14.0 billion at 31% organic growth. That is a company with pricing power today. Modine, in the same market, grew data-centre revenue 90% in its June 2026 quarter while adjusted EBITDA rose only 5%, management attributing compressed data-centre margins to supply-chain constraints. Munters published its second quarter under its own headline — "Exceptional demand — external headwinds currently affecting profitability" — with Data Center Technologies order intake up 137% at a 3.6× book-to-bill and the DCT margin falling to 11.3% on tariffs, mix and ramp-up inefficiency; a quarter earlier its DCT backlog was SEK 15.2 billion with order intake doubled while revenue fell 7% and EBIT margin fell to 13.8% from 22.3%. nVent grew sales 53%, 34% organically, with backlog at $2.6 billion and guidance raised to 21–23% organic growth.

Two companies in the same layer, in the same quarter, reporting that the orders arrived and the margin did not. Backlog in this layer is not a margin annuity.

Meanwhile the entry is visible in specifics rather than intent. Eaton acquired Boyd Thermal as part of $11 billion of first-quarter acquisitions, folding cooling into an electrical platform. Trane's Americas Commercial HVAC backlog grew 90% with total backlog at $12.1 billion, up 70%, and a book-to-bill of 123%. Johnson Controls' backlog grew 32% organically with orders up 27%. Schneider Electric is named among liquid-cooling suppliers alongside the specialists. Consensus Assumption — the further reports that Lenovo, Super Micro and Chinese coolant-distribution-unit vendors have moved from dependence on the incumbents to building their own control-software platforms reached the underlying analyses through trade summaries rather than any issuer filing, and are flagged in the not-verified list. They are the mechanism to watch rather than a fact to trade on, and they are the reason the cooling grades carry margin-based falsifiers rather than growth-based ones.

When four large diversified industrials with global service networks decide a niche is strategic, the niche's excess return has a shelf life. The Open Compute Project's rack and blind-mate manifold specifications accelerate that: standard interfaces commoditise generic metal considerably faster than they commoditise controls and service. Whether the durable piece is the hardware or the installed-base service annuity underneath it is taken up in Part II. The evidence points to the annuity — plus the ability to sell power and thermal as one commissioned system rather than as two catalogues.

Below the equipment sits the labour, and it binds harder than the hardware. Comfort Systems, EMCOR, Quanta Services, MYR Group and Sterling Infrastructure own scarce field capability — self-performDoing the work with your own employees rather than subcontracting it out. In a market where skilled electricians and pipefitters are the binding constraint, a contractor who employs the crews directly controls the schedule and can price to the shortage; one who subcontracts is passing through somebody else's labour cost and somebody else's delays. mechanical and electrical crews, modular prefabrication capacity, and the ability to schedule a commissioning sequence that does not slip. Labour cannot be printed like a CDU and relationships matter. But contracts allocate material pass-through, fixed-price and schedule risk, and price is usually set by a bid. Strong current returns on capital are entirely possible here with low fixed capital. They remain an execution return rather than proprietary product pricing, and the difference shows up the moment the labour market loosens.

I·8The electrical path, and where the lead times physically originate

Power arrives at a campus at transmission voltage and has to be stepped down, protected, switched, backed up, converted and distributed before any of it reaches a processor. Trace the path and the constraints announce themselves in order.

From the transmission line, a large power transformer steps voltage down to distribution level. That transformer is a bespoke object: a wound core immersed in oil inside a steel tank, built to a specific impedance and cooling class, using grain-oriented electrical steelA specialised steel whose crystal structure is deliberately aligned in one direction so that magnetism passes through it with less loss. Transformers are built from it and there is no substitute. Very few mills in the world make it, and building a new one is itself a multi-year project — a bottleneck sitting underneath the bottleneck. from a small number of mills, and shipped as an indivisible several-hundred-tonne unit that often requires a rail route survey. There is no way to make one quickly and no way to make one generically. The US Department of Energy's own programme materials put large-power-transformer lead times at more than a year on that basis alone — custom design, custom manufacture, difficult transport.

Downstream sit medium-voltage switchgearThe assemblies of breakers, switches and protection that route power around a site and cut it off safely when something fails. "Medium voltage" is the range between the transmission grid's very high voltages and the low voltage inside a building. Each assembly is engineered for its specific site and safety-certified, which is why it takes the better part of a year to arrive rather than being ordered from a catalogue., engineered to order and safety-qualified; protective relayingThe devices that watch the electrical system continuously and open a breaker within milliseconds when they detect a fault. They are the reason an electrical failure stays a local inconvenience instead of destroying equipment or starting a fire, and their settings are specific to the installation they protect.; uninterruptible power suppliesThe bridge across a power interruption. Batteries or spinning flywheels carry the load for the seconds or minutes between the grid failing and a generator starting — long enough that the computers never notice. In a data centre it is not a convenience but a design requirement, and it is one of the products whose service and replacement business outlives the original sale by decades. with their batteries or flywheels; buswaySolid metal bars in an enclosure used to carry very large currents around a building, instead of bundles of cable. It is more compact, easier to tap into at intervals, and — being essentially shaped copper or aluminium — far closer to a commodity than the switchgear upstream of it.; power distribution units; and finally the rack power shelves that convert to the DC rails the accelerators want. Every conversion loses energy and every protective device must interrupt fault current safely, which is why the whole path is engineered rather than assembled. Redundancy multiplies the equipment count: a facility built to survive a single failure buys roughly twice the switchgear it strictly needs.

The lead times follow directly from which of those is bespoke and which is catalogue.

Those lead times have produced the price behaviour anyone would expect. Power transformer prices are reported up 77% since 2019, generator step-up units 45%, certain distribution transformer classes as much as 95%, circuit breakers 47% since 2021 and medium-voltage switchgear 50%. Tier 2

They have also produced exactly the capacity response anyone would expect — which is what decides whether the current margin is a franchise or a queue. Hitachi Energy is building a $457 million large power transformer plant in South Boston, Virginia targeting 2028, with 825 jobs, plus a $106 million component expansion in Alamo, Tennessee. Siemens Energy is building its first US large power transformer plant in Charlotte, North Carolina for $150 million with production from early 2027. Eaton has committed $340 million to a South Carolina three-phase transformer facility with production and hiring beginning in 2027, and separately announced a 370,000-square-foot Nebraska switchgear plant beginning production in the first half of 2027, saying its cumulative global manufacturing investment since 2023 exceeds $1.5 billion. ABB has committed about $120 million to expand US manufacturing at Selmer, Tennessee and Senatobia, Mississippi for distribution equipment serving data centres and utilities, ramping across 2026, plus around $200 million into European medium-voltage manufacturing over three years. GE Vernova's acquired Prolec GE adds seven transformer sites and roughly $5 billion of backlog, with cross-selling from existing international factories beginning to lift revenue in 2027–28; Prolec has itself committed around $300 million. That is roughly $1.8 billion of announced North American transformer capacity, most of it landing 2027–2028, and a further several hundred million in switchgear and medium-voltage gear on the same clock.

That is the same shape as the memory story: a shortage rent, a visible capacity response, and a date. The difference — and it is a real one — is what sits underneath the rent when it goes. In memory, nothing does; the product is fungible bits. In electrical equipment, the installed base creates genuine recurring economics in UPS, breakers, protection and controls, where inspection, spares, retrofits and control-system upgrades recur for the twenty-to-forty-year life of a facility, and where the specification is written into the building. Cable, conduit and busway replacement is far less proprietary, which is why Atkore's six months to 27 March 2026 produced negative $27.2 million of operating cash flow and negative $53.5 million of free cash flow against positive $160.9 million and $97.3 million in the prior-year period — a commodity conduit and cable business whose pricing reversed inside the boom.

Prysmian shows why a company has to be graded by segment rather than by label. Its Q2-26 group revenue of €6,021 million produced adjusted EBITDA of €730 million at a 15.4% margin and group net profit of just €323 million, 5.4% of sales. Inside that, Transmission earned a 21.2% margin, up from 17.1%, on 14.3% organic growth against a €17 billion backlog; Power Grid earned 13.8% on 13.0% growth, described as substantially stable; Digital Solutions jumped to 23.8% from 16.8%. Submarine transmission is a franchise. Drawn copper is not. One ticker holds both.

The most telling numbers in this layer, though, are the ones moving the wrong way at the top of the boom. Eaton's first-quarter Electrical Americas twelve-month rolling orders were up 42% and electrical backlog up 48% — with segment margins of 22.7%, down 120 basis points year over year. Hubbell's second-quarter adjusted operating margin was 23.9%, down 50 basis points, on 15% sales growth. Powell's fiscal second-quarter revenue rose only 6% to $297 million while new orders rose 97% to $490 million and backlog rose 33% to $1.8 billion, with a >$400 million data-centre order landing after quarter-end. Read those together: orders are exploding, backlog is compounding, and margin is flat to down, because input cost and capacity investment are running ahead of price and because converting a backlog is not the same as pricing one. A layer with genuine unexploited pricing power does not print declining margins at the peak of the shortage it is supposedly exploiting.

I·9How the grid works for a large new load

A gigawatt of compute is, from the grid's point of view, a city that wants to appear in three years and pay for none of the wires.

Three separate processes govern whether a data centre gets power, and confusing them is the single most common analytical error in this sector. The first is interconnection: a request to connect load at a point on the system, studied for its effect on thermal limits, voltage and stability, resulting in a list of network upgrades and a cost allocation. The second is resource adequacy: in organised markets, a capacity construct that pays generators to be available years ahead so that somebody builds the megawatts the load will need. The third is the retail and tariff layer, where a state commission decides what the customer pays and who bears the cost of anything the customer's arrival required. A power purchase agreement is a financial or physical claim on generation. It is not, by itself, a transmission right, and it does not energise a campus.

Each of those is now visibly straining, and each strains differently.

Interconnection: the queue is not the demand

In the update its chief executive gave the Texas House Committee on State Affairs on 9 April 2026, ERCOTThe Electric Reliability Council of Texas — the operator of the Texas grid, which is deliberately kept largely separate from the rest of the United States and is therefore regulated mostly by Texas rather than by federal authorities. It runs an energy-only market: generators are paid for the electricity they produce and nothing for merely standing by, which makes its scarcity signals look very different from PJM's. reported tracking 410,618 MW of large loadsGrid language for anything that consumes electricity. A "large load" is a single new customer big enough to need its own study — a data-centre campus, a steel mill. The word matters because the grid's rules were written for load that arrives gradually and predictably, and a gigawatt appearing in one place in three years is neither. seeking interconnection through 2030 — roughly 410 GW, about 87% of it data centres — against a summer peak the system has never run much above 90 GW. The decomposition on the same slide matters more than the headline. Of that 410.6 GW, 293,651 MW had submitted no studies at all, 86,605 MW were under ERCOT review, 21,343 MW had planning studies approved, 3,241 MW held Approval to Energize without operating, and 5,778 MW was observed energised.

Texas has already begun pricing that optionality, which is the fastest way to find out how much of a queue is real. Under PUC Project 58481, the proposed 16 TAC §25.194 large-load interconnection standards published for comment on 12 March 2026, a large load seeking an interconnection agreement must post site control and financial security of $50,000 per megawatt, converting to non-refundable on execution. A gigawatt of optionality now costs $50 million to hold. That single rule change will reveal more about the real size of the Texas queue than any forecast.

Resource adequacy: where the money shows, and where the ceiling is

PJM's Base Residual Auction for the 2028/29 planning year, announced 14 July 2026, procured 138,318 MW of unforced capacityNameplate generating capacity discounted for how often the plant is expected to be unavailable — broken, refuelling, or in the case of wind and solar, becalmed or dark when it is needed. It is the honest unit for a reliability calculation, because a gigawatt that works four-fifths of the time is not a gigawatt of insurance. at a clearing priceIn an auction, the single price at which the last unit needed is bought — and which everyone who cleared then receives, including sellers who would have accepted less. It is set by the most expensive supply required, which is why an auction can pay a great deal to everybody because of a shortage at the margin. of $325/MW-day — the FERC-approved administrative cap, hit for the third consecutive auction — while clearing 6,831 MW short of the one-event-in-ten-years reliability requirement, at a total cleared cost of $16.4 billion. New generation and uprates contributed 525 MW of that 138,318 MW. PJM's own market monitor estimated the uncapped price would have been $554.72/MW-day, which on the same cleared quantity implies roughly $29.7 billion rather than $16.4 billion. Talen Energy independently confirmed the $325 clearing price in its own Form 8-K and reported clearing 10,180 MW for approximately $1,208 million of capacity revenue for the planning year.

PJM is not the country, and the differences are structural rather than cosmetic. MISO's corrected 2026/27 Planning Resource Auction found sufficient annual capacity while clearing distinct seasonal and zonal prices — summer at $424.30/MW-day in Zones 1–7, $384.10 in Zones 8 and 10 and $412.10 in Zone 9, with the other seasons far lower. ERCOT has no capacity auction to quote at all: its own resource-adequacy assessment describes an energy-only marketA market design in which generators are paid only for the electricity they actually produce, with nothing for standing by. The theory is that prices spiking during genuine shortages will attract enough new plant on their own. The practical consequence is that scarcity shows up as a violent price for a few hours a year rather than as a steady annual payment — a different risk to finance, and a different asset to own. without capacity payments, so scarcity there lands in energy and ancillaryThe supporting services that keep a grid stable moment to moment, as distinct from the bulk supply of electricity itself: reserves that can start within minutes, capacity that adjusts second by second to keep frequency steady, and voltage support. They are bought and priced separately, and their prices rise when the system is tight. prices and in bilateral structures instead. One of the three readings behind this report initially reached for an ERCOT "capacity price" and had to withdraw it; the correct statement is that ERCOT's missing capacity price is not a missing scarcity signal, and the queue-stage data plus the Batch Zero reform are what the operator actually publishes. There is no single transferable "power shortage" profit pool. There are three market designs with three different ways of paying, and a generator's economics depend on which one it sits in.

Who pays: the fight that decides the layer's returns

On 18 June 2026 FERC issued Section 206 show-cause orders to all six US organised markets — PJM in Docket EL26-67-000, SPP EL26-68-000, NYISO EL26-69-000, MISO EL26-70-000, CAISO EL26-71-000, ISO-NE EL26-72-000 — each preliminarily finding that the region's existing tariff appears unjust and unreasonable because it does not account for the way data centres and other large loads connect to and use the transmission system. Four of the six are template orders covering five areas: large-load application and study processes; network-upgrade cost transparency and mechanisms to prevent cost shifting; rates and terms for co-located load; services for flexible loads with on-site generation; and interconnection terms for large loads electrically proximate to generation. PJM's is deliberately narrower — four directives, confined to large loads not co-located with generation — because FERC addressed PJM co-location separately in a companion order issued the same day establishing rates and terms for interim network integration and for firm and non-firm contract-demand service. Each operator and its transmission owners have 60 days to justify the status quo or file tariff revisions; responses fall due 17 August 2026, with stakeholder comments by 16 September 2026.

"Co-located" is doing quiet work in those orders, and it is not the same word that appears in the real-estate section of this report. In the power sense it means siting the data centre next to the power plant and taking electricity from it more or less directly, instead of drawing off the shared grid — attractive because it skips the queue and the wires, contested because a customer who still leans on the grid when its neighbour plant is down is using a system it is not paying to maintain. The co-locationTwo unrelated meanings share this word. In the power sense it means putting a large load directly beside a generator and taking its output without routing it across the public grid. In the real-estate sense — the sense used in §II·6 — it means renting space, power and cooling in a shared data centre. The first is a regulatory fight about who pays for the wires. The second is a landlord business. of §II·6 is a different industry: renting rack space in a shared building. One word, two arguments, and conflating them makes the power layer unreadable.

The co-location story has a history worth carrying because it shows how unsettled this is. In December 2025 FERC found PJM's tariff unjust and unreasonable for lacking clear rates, terms and conditions for generators serving co-located load, and directed compliance filings creating firm and non-firm transmission options. At its June 2026 meeting the Commission modified and set aside that order in part, accepted and rejected portions of PJM's compliance filing, and required another filing within 60 days. The right is scarce. The legal incidents of that right are still being written, in public, with the ink wet.

One clause in the June orders matters more than anything else in the docket: existing negotiated agreements are explicitly protected. Deals already signed are grandfathered; deals not yet signed face a rule that does not exist yet. Which means the value of an executed co-location agreement went up on 18 June 2026, and the value of a pipeline of un-executed ones went down. That is an unusually clean, dated, one-way transfer, and it is the reason this report grades contracted firm power differently from uncontracted firm power.

Behind FERC sit the state commissions, where the actual cost allocation happens and where the politics are far less favourable to the load. A hyperscale campus brings tax base and few permanent jobs, consumes water and land, and requires generation, transmission and distribution investment whose cost, absent a specific tariff, spreads across every ratepayer in the territory. Utilities have every incentive to build it, because rate base earns an authorised return.

That last sentence governs half this layer, so it is worth unpacking. A regulated utility does not set its own prices. A state commission decides what it may charge, by letting it recover its running costs and then earn a set percentage return on the capital it has sunk into poles, wires and substations — and that pile of invested capital is the rate baseThe capital a regulated utility has invested in the physical system, on which a commission allows it to earn a set rate of return. Growing the rate base is how a regulated utility grows its earnings, because the percentage is fixed and only the base can move. It is a durable, low-volatility, thoroughly capped way to make money — and it means the utility's interest is in building, not in what the electricity ends up being worth.. The consequence is unusual and completely rational: the utility's earnings grow by building more, not by charging more. It has every reason to say yes to a customer that requires an enormous quantity of new equipment, and none whatever to interrogate whether that customer's growth forecast is real.

Ratepayers have every incentive to object. The resolution — large-load tariffs with minimum-take provisions, long terms, collateral and priced exits — is being written now, state by state, and it determines whether the regulated layer's growth is accretive or merely large.

One state has finished writing it, and every other commission is now working from its terms. AEP Ohio's Schedule DCT, effective 23 July 2025, applies to projects of 25,000 kW and above. It sets billing demand at no less than 85% of the customer's highest previously established monthly demand — or, for the largest projects, at fixed floors rising to 57,500 kW plus 100% of contract capacity above 75,000 kW. The initial contract term is the load ramp period, capped at four years, plus a further eight — twelve years at the outside. The ramp itself is contractual: 50% of contract capacity in year one, then 65%, 80%, 90%. Termination is permitted only after the fifth post-ramp year and costs an exit fee equal to thirty-six months of minimum charges. And a customer without an A−/A3 rating and ten times cash cover must post a guarantee or collateral equal to half the total minimum charges for the full term.

Others are converging on the same shape. Virginia's GS-5 tariff for loads of at least 25 MW begins in 2027 with a fourteen-year obligation and minimum payment levels. Georgia approved resources for nearly 9.9 GW of new demand while attaching explicit protections for existing ratepayers. The direction is unmistakable.

Take-or-payA contract under which the buyer pays for an agreed quantity whether or not it takes delivery. It exists to make somebody else's investment financeable: the seller can borrow against the payments because they arrive regardless of what the buyer decides later. For the buyer it converts a flexible option into a fixed bill, and that conversion — not the price — is usually the important term. is the instrument that does it, and the name is the definition: the customer pays for the agreed quantity whether it uses it or not. A commission has converted a data centre's demand forecast into a take-or-pay obligation with an eight-to-twelve-year tenor, posted collateral and a priced exit — so a megawatt reserved and never switched on still gets billed. That does two things at once. It protects the ratepayer, which is what it was written to do. It also converts the hyperscaler's flexible option on power into a fixed liability — which is precisely the change that makes generation behind the meter, on the developer's own land and outside the commission's tariff, look cheaper than waiting in a queue for firm service. Cost-causation regulation does not slow the buildout. It moves it off the regulated grid, and §II·11 follows the money that goes with it.

Where firm power actually converts into cash

The long-tenor nuclear agreements now signed are where scarcity and rent visibly separate. Constellation signed a twenty-year power purchase agreement with Microsoft supporting the restart of Three Mile Island Unit 1 as the Crane Clean Energy Center, adding approximately 835 MW of carbon-free generation. Vistra and Meta announced twenty-year agreements for more than 2,600 MW from Perry, Davis-Besse and Beaver Valley, including uprates. Talen has expanded its Amazon relationship around Susquehanna. These are not merchant spark-spreadThe margin a gas-fired power plant earns: the price it gets for electricity, less the cost of the gas burned to make it. A plant selling into the open market lives or dies on that gap, which nobody controls and which can vanish for years at a time. "Merchant" here means exactly that exposure — no contract underneath, just whatever the market happens to pay. bets. They are long-tenor offtake against existing or restartable baseloadGeneration designed to run continuously at close to full output — nuclear and large coal or gas plants — rather than starting and stopping to follow demand. It is what a data centre actually needs, since a training cluster does not care whether the wind is blowing, and it is the reason existing nuclear plants became the most sought-after assets on the grid. assets whose interconnection is already largely in place — a completely different claim from a greenfield gas plant ordered into a multi-year turbine queue.

They are also not finished. Constellation's Crane restart remains a licensing and operational programme: the Nuclear Regulatory Commission still lists the plant in SAFSTORThe regulatory status of a shut-down nuclear plant being held in a safe, monitored condition while its radioactivity decays, ahead of eventual dismantling. A plant in SAFSTOR is legally on its way to being demolished, not on its way to restarting. Reversing that is a licensing project with its own schedule, which is why an announced restart date is a plan rather than a fact until the regulator agrees. and requires licensing-basis and component restoration, and the same is true of the Palisades restart. Constellation's plan includes uprates at Clinton and Crane, and the existing fleet's operating cash flow pays while those milestones advance. Nuclear restarts are valuable precisely because the site, the interconnection and the workforce already exist. Their schedule is regulatory, and no announced date is a fact until the NRC agrees.

New nuclear sits on a different clock entirely and should not be confused with it. Google's first Kairos–TVA project is 50 MW and scheduled for 2030. Amazon and Energy Northwest target 320 MW in the early 2030s, expandable to 960 MW, while Amazon's broader X-energy ambition reaches 5 GW only by 2039. These are meaningful demand anchors and they change nothing about the 2026–29 capacity balance. The investable scarcity belongs to operating sites and existing interconnections until permits, fuel, construction finance and first concrete are facts.

Gas turbines are the other binding industrial constraint, and they are a stronger franchise than the headline lead time implies — for a reason that has nothing to do with the lead time. GE Vernova ended the second quarter with 116 GW of gas equipment backlog plus slot reservations, up from 100 GW a quarter earlier, targeting at least 125 GW by year-end 2026, against total company backlog of $176 billion of which more than half is services, and an installed base of roughly 7,000 units. Its planned output rises from a 20 GW annualised run-rate in the third quarter of 2026 to 24 GW in 2028 and 30 GW in 2030. Siemens Energy reported a €154 billion total backlog with €17.7 billion of quarterly orders, gas booked out to fiscal 2028, a €60 billion Gas Services backlog, and 194 gas turbines sold in FY2025; Mitsubishi Heavy reported FY2025 order intake of ¥7.654 trillion with continued strong gas-turbine demand. Capacity will rise. The slot price will normalise. The service attachThe maintenance revenue that comes attached to a piece of installed equipment across its operating life — spares, scheduled overhauls, upgrades and long-term service agreements. It is earned against an installed base rather than against a new-build order book, which is why it outlasts the equipment cycle that created it and why it, rather than the delivery slot, is the durable part of the franchise. on a 7,000-unit fleet with a thirty-year life will not.

I·9bLand, water, permits and the tariffs underneath the iron

Every megawatt eventually becomes a zoning hearing. The physical chain ends in a county planning meeting, and the contract wrapped around the electron decides which of three parties gets paid.

Land banks with secured power are the real-estate expression of the interconnection queue. Water rights and thermal discharge permits are the liquid-cooling expression of the same scarcity. Direct-to-chip loops can approach low on-site water consumption, but the heat still has to reach a dry cooler, a cooling tower or some other sink, and the site trades water against electricity efficiency in a way that is locally political. Closed loops reduce consumption; they do not remove heat rejection, noise, pipelines or the need for community consent.

The same physical electron supports three completely different profit claims, depending on the contract wrapped around it. A utility that rate-bases a dedicated substation earns an authorised return on equity — real, durable, low-volatility, and capped by a commission. An independent power producerA company that owns power plants and sells their output, without being a regulated utility and without owning wires to customers. It has no guaranteed return and no commission setting its prices — it takes whatever the market or its contracts pay, which is why the same asset can be a bargain or a disaster depending entirely on what was signed before it was built. that sells a behind-the-meterGeneration sited on the customer's own land, feeding the customer directly without passing through the public grid — so it never crosses the utility's meter and largely escapes the utility's tariff. It is the escape route from the interconnection queue, and the reason strict cost-allocation rules on the regulated grid tend to relocate the buildout rather than slow it. or co-located configuration is fighting a live federal and state jurisdictional question and may keep the upside or may have it reallocated. A hyperscaler that signs a twenty-year nuclear PPA buys schedule certainty and carbon attributes in one instrument and transfers price risk to the seller. Three claims, one electron. That is why this report refuses to grade "power" as a single factor, and why the panel's disagreement about the power layer turns out to be a disagreement about which of the three claims a given ticker actually holds.

When the political equilibrium settles where the evidence points — with large-load tariffs assigning network upgrade costs to the hyperscaler rather than socialising them across residential ratepayers — the value of being early in the queue rises rather than falls, and the equity claim migrates toward whoever controls the permitted site and the firm offtake rather than toward whoever poured the concrete last. That is a second-order effect and it runs opposite to the intuition that regulation is bad for the layer. Strict cost causation is bad for the marginal project and good for the already-permitted one.

I·10Four booms that ended, and exactly how

Analogy is dangerous and unavoidable. The discipline that makes it useful is naming the specific mechanism that ended each prior episode — because those mechanisms are the falsifiers the second half of this report depends on.

The shared lesson is not that the demand was imaginary. In every one of those episodes the demand forecast was directionally correct, and in two of them it was an understatement. The lesson is that a forecast can be directionally right and a marginal investment still uneconomic, because what determines the return is the supply and contract structure that meets the demand rather than the demand itself.

Which gives the falsification discipline for the second half of this report, stated as three tests anyone can run. If optical module average selling pricesThe average price actually realised per unit sold, across the whole mix. It is the number that reveals whether a growing revenue line is growth or merely volume: units up and average selling price down means the market is expanding while the pricing power drains out of it, which is the ordinary fate of every optical component generation. are still rising three years after the current capacity announcements, something structural changed in optics. If HBM contract prices are still rising after two years of industry bit growth above 30% and multi-supplier HBM4 qualification, something structural changed in memory. If merchant power plants without offtake earn mid-teens unlevered returns for a full decade, something structural changed in generation. Until one of those exceptions prints in a disclosure, the base case is that the old mechanisms still operate — on longer clocks, with larger numbers, and with contracts that push the reckoning further out than they used to.

II·1The conclusion, stated first

Three families of right clear the bar on the disclosed evidence, and each pairs something money alone cannot replicate with a business that does not have to consume its whole profit to keep it. Everything else is renting a shortage, or growing at a margin somebody else sets.

A physics-limited process. Extreme-ultraviolet lithography, where the barrier is two decades of accumulated supplier know-how and the owner is deliberately under-monetising it; and leading-edge foundry with its advanced packaging attached, where yield learning compounds with cumulative wafer starts and packaging co-optimisation has so far kept the assembly profit inside the foundry rather than letting it leak to the back end. Around both sits a defensible oligopoly of process learning — deposition, etch, planarisation, and above all the inspection, metrology and test that price the rising cost of failure.

A compounding software-and-system estate. The accelerator as a rack-scale system, where the defensibility has migrated from the die to the combination of CUDA and the broader developer estate, interconnect, rack-scale integration, annual cadence and supply allocation — and, beneath it and less visibly, the design tolls that every competing chip must also pay: architecture licensing and electronic design automation, which get paid more the more successfully the accelerator vendor is attacked.

A legally or contractually scarce right. Electrical franchises where the specification is written into a twenty-year facility and the aftermarket is not competitively bid; and firm power sold under long-tenor offtake against an existing or restartable asset whose interconnection already exists. Both are rights that a regulator or a counterparty has already granted and will not grant again quickly.

The rest divides three ways, and the report is as committed to the negatives as to the positives.

Some layers are earning spectacular money on a shortage that their own margins are funding the end of, with dates their own filings disclose: memory and storage, gas-turbine slots, high-voltage transformers and switchgear, uncontracted firm capacity. Some are growing enormously with the margin structurally fixed by contract or by competition: server and rack assembly, electrical and mechanical contracting, most optical component manufacturing, subsystem build-to-print, commodity conduit and cable. And one — colocation real estate — is earning a respectable operating margin on a capital base so large that the return on capital sits below any plausible cost of it, while the buildout it is supposed to be capturing largely goes around it.

The consensus this revises is specific, and it has two parts. The prevailing view treats "AI infrastructure" as a single exposure with a single set of economics; and it treats a physical bottleneck as automatically equivalent to pricing power. The first is a category error. The second is wrong in an unusually costly way, because the physically tightest nodes on this chain — turbines, transformers, interconnection queues, firm capacity — are precisely the ones where the rent is being competed away by announced capacity, capped by an administrative price, or reallocated by a regulator. The order of physical scarcity on this chain is close to the inverse of the order of retained economics, and Figures 1, 9 and 11 make that visible without commentary.

A moat whose width depends on a rival's construction schedule is a lead, not an annuity.

The test applied to every chokepoint below

II·1bHow concentrated is it? The range, and why the two readings differ

This is the one question on which three independent readings of the same evidence did not converge, and the honest output is a range with both cases argued rather than a count.

The question is not where the durable profit sits — all three readings landed on substantially the same map. It is how many nodes clear the bar, and the disagreement is worth more than the answer, because it exposes exactly which test is doing the work.

Both cases are correct about the evidence they cite. The disagreement is a composition effect, not a doctrinal one — which is the most useful single output of putting three readings side by side.

Look at what actually sits inside the labels. "High-voltage electrical equipment" contains ABB at a disclosed 28.4% return on capital employed with a 20.2% operational EBITAEarnings before interest, tax and amortisation — profit measured before the cost of borrowing, before tax, and before the accounting write-down of intangible assets picked up in past acquisitions. It is a common European reporting convention. It sits above operating profit on the income statement, so an EBITA margin will always read higher than an operating margin on the same business. margin and record orders of $12,042 million; Eaton at a 22.7% segment margin, down 120 basis points; Schneider at a 21.1% adjusted EBITA margin on 14% organic growth with guidance raised; Hitachi Energy with a $57.9 billion backlog against $32.8 billion of annual orders; Powell at 29.6% gross margin on 6% revenue growth; and Atkore burning cash on commodity conduit. Those are not one business. The first four hold a specification lock plus an aftermarket annuity; the last two hold a queue and a metal spread respectively. A reading that grades the layer on its leaders gets "durable." A reading that grades it on its capital-weighted average, or on the single largest name by backlog, gets "shortage rent." Both are describing the same disclosures.

The same decomposition explains the power layer. "Firm power" contains Constellation and Vistra — which now hold twenty-year contracts against existing nuclear assets, and whose consolidated return on capital employed is depressed by a fleet and a balance sheet built for a different market — and it contains merchant capacity sold into an administratively capped auction, which is a different instrument with a different owner of the upside. The contracted part is a genuine annuity whose price was struck before the regulator started writing the rules, and whose value went up on 18 June 2026 when FERC grandfathered executed agreements. The uncontracted part is a scarcity option that a commission is in the process of repricing. Grading the ticker grades both at once, which is why the ticker-level grades split three ways across three readings.

So the report states the range and does not force a count. The concentrated reading is right that consolidated return on capital, measured now, disqualifies most of the physically scarce layer, and it has the arithmetic. The broader reading is right that a contracted twenty-year offtake and a thirty-year service attach on an installed fleet are durable rights that a return measured at a single point in a capital cycle does not capture, and it has the contracts. What both agree on — and this is the operative conclusion — is that within the physically scarce layers, durable profit belongs to the contracted, the specified-in and the serviced, and not to the queue. The queue is the part with a date on it.

II·2Sizing the pool — and reconciling the forecasts

The pool this chain allocates has a hard floor that requires no forecasting at all: the capital budgets four companies have already told their own shareholders they will spend. Everything above that floor is an argument about utilisation and electrons.

The floor moved during the year, and the underlying analyses caught it at three different moments — which turns out to be the most informative thing about it.

Two independent aggregations corroborate the level from outside the filings. The International Energy Agency measured five large technology companies' data-centre-related capital expenditure above $400 billion in 2025 and expected roughly a 75% increase in 2026, which implies something near $700 billion on its own definition; a widely circulated market tally of disclosed big-four guidance landed near $725 billion. Tier 2 Those definitions are not identical to each other or to the sum above — they differ on which companies are counted, on lease treatment, and on whether customer-supplied hardware is included — and the useful conclusion is not a reconciled point estimate but the range: roughly $700–735 billion of calendar-2026 capital expenditure from the largest buyers, on definitions that differ by tens of billions.

The fifth buyer belongs in the count and is structurally the most interesting of them, because it is the one whose financing has already broken. Oracle closed fiscal 2026 on 10 June 2026 with $67.4 billion of revenue up 17%, $20.6 billion of GAAP operating income, cloud infrastructure revenue up 77% for the year, and remaining performance obligations of $638 billion — up 363% year over year and up $85 billion in the quarter alone. Set against that: operating cash flow of $32.0 billion and free cash flow of negative $23.7 billion, which implies roughly $55.7 billion of capital expenditure in one fiscal year, or 83% of revenue. Oracle funded it with $43 billion of debt and $5 billion of equity in fiscal 2026 and expects to raise about $40 billion more in fiscal 2027, including a $20 billion at-the-market equity programme.

Most of the RPO increase, Oracle told shareholders, came from "large scale AI contracts where the customer prepaid Oracle for the purchase of the GPUs, or the customer bought and supplied the GPUs to Oracle," and "the prepaid and customer supplied hardware portions of our large AI contracts now total $75 billion."

A cloud provider is being pre-funded, in cash and in kind, by its own customers, for the most expensive item on its balance sheet. Read one way that is prudent risk transfer and Oracle says so — it "substantially reduces the amount of capital Oracle must raise." Read the other way, $75 billion of what the market counts as third-party demand for accelerators is a purchase the buyer had already committed to make and merely routed through a supplier's income statement. Both readings are true simultaneously, which is exactly the difficulty. Anyone sizing accelerator demand from cloud remaining performance obligations should know that a material slice of the largest order book in the industry is hardware the customer bought itself.

Around that floor sit the third-party projections. Averaging them teaches nothing; the assumption on which they disagree does. On the silicon side, SEMI's 300mm fab equipment outlook puts spending at $133 billion in 2026, up 18%, then $151 billion in 2027, $155 billion in 2028 and $172 billion in 2029, with total equipment sales reaching a record $156 billion in 2027; its mid-2026 forecast puts foundry and logic WFE at $78 billion for 2026 with DRAM at $38.8 billion and NAND at $13.9 billion, and total equipment reaching $229 billion by 2028. Tier 2 / not fully verified The deceleration inside that series — +18%, then +14%, then +3%, then +11% — is the divergence: the high case assumes AI accelerator and HBM demand keeps pulling wafer capacity forward without a digestion year; the low case assumes the 2028 memory capacity wave arrives into a market already supplied. A digestionThe industry's word for the pause that follows a buying spree. Customers who ordered ahead of their needs stop ordering for a while and work through what they already hold. Demand has not fallen and nobody's forecast was wrong — but the supplier's revenue drops anyway, sometimes hard, because its revenue is orders rather than consumption. year is the pause after a buying spree: customers who ordered ahead of their needs stop for a while and work through what they already hold. Nothing has gone wrong with demand and nobody's forecast has been proved mistaken. The supplier's revenue falls all the same — which is precisely why it is the mechanism that catches this layer out. Nothing in this report's profit map or any grade depends on that series.

On the electricity side, the divergence has a cleaner shape and it is between queue-based and contract-based demand. The IEA's base trajectory takes global data-centre electricity from roughly 485 TWh in 2025 toward about 950 TWh in 2030, roughly 3% of global electricity, with its sensitivity cases spanning near 700 TWh in a bottlenecked "Headwinds" case, roughly 1,200 TWh in the base and above 1,700 TWh in "Lift-Off" by 2035. The swing assumption is stated: whether supply chains, siting and grids let announced compute become energised stock, and whether efficiency lowers energy per task faster than workloads proliferate. The US Energy Information Administration, working bottom-up from building stock rather than from queues, lands elsewhere again: data-centre servers at an estimated 7% of US commercial-sector electricity consumption in 2025, rising to 22–33% of commercial building electricity by 2050 across its cases, with standalone data centres reaching 581 billion kWh by 2050 in the high-demand case and cooling adding a further 84 billion kWh; its server-electricity range for 2050 alone spans 446–818 TWh, with efficiency and installed AI-server stock as the swing assumptions. Those are decades-long curves rather than the step change the queues imply.

Set the two side by side and the reconciliation is straightforward. A forecast built off ERCOT's 410.6 GW of large-load requests produces one number. A forecast built off the 5,778 MW actually energised and the 3,241 MW approved but not yet operating produces a very different one. Both are "the data." Only one of them has signed anything. When the dollar forecast and the electron forecast diverge, power availability and packaging capacity usually explain the gap before "AI demand disappeared" does.

What a megawatt actually costs — and the limit of saying so

The mandate for this work asks for a dollars-per-megawatt build-up decomposing a deployed megawatt into silicon, memory, network, cooling, electrical and power-contract components. It can be bounded from disclosure. It cannot be decomposed from public data, and stating both facts plainly is more useful than a confident-looking stacked bar. One of the three analyses behind this report published such a bar and then withdrew it on review; a second searched issuer filings specifically for separately disclosed silicon, memory, network, cooling, electrical and site values on one consistent IT-load denominator and found that no issuer publishes them on a common scope. Power is a capacity denominator, not an accounting segment.

Evidence: CoreWeave Q1-2026 results and the property-and-equipment line in its Q1-2026 balance sheet; CoreWeave 2026 capex guidance and active-power target as reported in trade coverage; a widely circulated industry cost range for AI-ready facilities; and the Open Compute Project's Open Data Center Design Guide, which supplies a physical rack, pod and megawatt design reference and no dollar allocation whatever. Method: each row is one ratio with its numerator and denominator stated; the rows are alternative bounds on the same quantity and must not be added. Synthesis: the honest read is roughly $35–50m of total capital per megawatt of live AI compute, of which perhaps half to two-thirds is IT equipment — and the reason the range is that wide is that the two halves sit on completely different depreciation clocks, three-to-six years against twenty-to-forty.

Backlog coverage — four words meaning four different things

On this chain, "backlog" is used to mean at least four different instruments with four different cancellation regimes. The coverage ratio is where the confusion between an order book and an annuity becomes visible.

Evidence: each company's most recent earnings release, as listed in the register. Method: coverage = backlog ÷ forward-year revenue, using company guidance where given and the latest quarter annualised where not. The denominators are not comparable across rows (guidance versus run-rate) and the numerators are definitionally different — GE Vernova's includes slot reservations, CoreWeave's includes estimated amounts beyond contracted RPO, Micron's is a minimum-price minimum-volume floor rather than a revenue expectation, Quanta reports RPO and total backlog separately. Cancellation terms are not disclosed in these releases and were not read in the underlying contracts, so the ratios are not ranked. Synthesis: the two largest coverage ratios belong to the two companies whose order books are least like firm orders, and the smallest belongs to the one whose contracts carry an actual price floor. That inversion is the confusion between backlog and annuity, rendered as arithmetic.

The change in this cycle's character

Two structural facts about the pool matter more than its size. First, its composition is shifting hard toward short-lived assets: the accelerator, the memory and the server are three-to-six-year assets, while the shell, the substation and the transformer are twenty-to-forty-year assets. A dollar spent on the first has to be earned back four to eight times faster than a dollar spent on the second, and the two are being spent side by side inside a single capex line.

Second, and more consequential, the pool has stopped being self-funded. Alphabet's second quarter produced $39.07 billion of operating cash against $44.92 billion of capital expenditure — free cash flow of minus $5.86 billion — and the company raised $49.6 billion of equity and mandatory convertible preferred plus $20.3 billion of senior notes in the quarter, explicitly citing capital expenditure to scale AI infrastructure. Amazon's trailing-twelve-month free cash flow is minus $7.6 billion after a $66.1 billion year-over-year increase in property and equipment purchases. Meta generated $784 million of free cash flow on $31.86 billion of operating cash flow and carries $83.66 billion of long-term debt against $349.3 billion of disclosed non-cancellable commitments. Oracle's free cash flow is minus $23.7 billion.

When capital expenditure was funded from operating cash, its only constraint was management conviction. Funded from capital markets, it acquires a second constraint — the price and availability of that capital — and that constraint can tighten for reasons entirely unrelated to whether AI works. §II·11 traces where that transmits, and in what order.

II·3The profit-pool map, layer by layer

Ten layers, each graded on one test: does the excess return survive the capacity, substitution, standards-opening, vertical integration and customer-concentration pressure that the return itself is inducing?

Set the profit map against where the money actually moves, and the two barely overlap. Revenue flows through every node on this chain; retained economics stop at a few of them. The ranking of the one bears almost no relation to the ranking of the other.

II·4Who sets the price, layer by layer

The fastest route to durable profit is to ask, at each node, who actually names the number. The answer is rarely the seller with the highest margin, and in four of eighteen cases it is not a seller at all.

Evidence: each cell's figure is drawn from that company's own most recent release, listed in the evidence register. Method: "who sets the price" is a report-authored classification based on which party holds the scarce, non-substitutable input in the transaction. Margins across rows use different fiscal periods and different accounting bases and are not like-for-like; they are cited as evidence for the classification in their own row only. Synthesis: in seven of twenty-one rows the price-setter is not the seller, and in two of those it is a regulator. Note also that "the seller" appears twice inside the electrical layer and twice inside the power layer with opposite verdicts either side of it — which is the composition effect §II·1b identifies, made visible row by row.

Where the bargaining power has moved — and where it is heading

II·5Chokepoint durability — nine verdicts, each with a date

Nine candidate monopolies. For each: what defends it, what is being built to break it, and the disclosed event at which the test resolves.

Lead time is this subject's real unit of scarcity — not how tight a market is, but how long it takes anyone to answer it. Put every binding constraint on one axis measured from today, and the ranking of who is protected, and for how long, falls out of the chart rather than out of an opinion.

Evidence: ASML Q2-26 release and High-NA readiness disclosure; TSMC Q2-26 release, presentation and 4Q25 long-term targets; Amkor Q2-26; ASE Q2-26 guidance; NVIDIA Q1-FY27 release and Form 10-Q; NVIDIA Vera Rubin update, 21 July 2026; Cadence Q2-26; Arm Q1-FY27; Synopsys Q2-FY26; Micron FQ3-26 and SEC XBRL company-facts; SK hynix 2Q26; Samsung Q2-26; Eaton Q1-26; Hubbell Q2-26; Powell FQ2-26; ABB Q2-26; Schneider H1-2026; Hitachi Energy and Siemens Energy plant announcements; GE Vernova Q2-26 release and SEC XBRL company-facts; PJM 2028/29 result and Talen Form 8-K; FERC Dockets EL26-67 through EL26-72 and the December 2025 co-location order as modified in June 2026; Constellation and Vistra PPA releases; NRC Crane and Palisades status pages. Method: each verdict applies one test — does the excess return survive the specific supply, standard or rule now being created in response to it. Red pills mark a verdict of no; outlined pills mark a chokepoint that splits internally, where one component clears the test and another does not. Synthesis: of nine candidates, three survive cleanly, one fails outright, and five split — and in every one of the five the durable half is the contracted, specified-in, serviced or licensed part while the exposed half is the queue. Three of the five splits sit in the layers the market currently treats as most physically defended.

II·6The layers that hold no durable profit — and the mechanism that removes it

A study that finds durable profit in nine of ten layers has not been done properly. Here are the three where it is absent, and precisely what removes it.

Server, rack and ODM integration

The mechanism is structural rather than cyclical, and it is the classic sandwich: a customer who specifies the design, a single-source component that dominates the bill of materials and is bought at a price the assembler does not set, and several qualified assemblers bidding for the work. There is no configuration of those three facts that yields pricing power. Dell's ISG segment at 10.7% is the ceiling, and it is a branded OEM with storage, financing and enterprise services attached; the merchant assemblers run at 3.57% to 9.8%.

Quanta Computer is the single clearest number in the layer: first-quarter revenue up 66.6% to a record NT$809.2 billion with AI servers above 75% of the mix, and gross margin compressed to 4.78%, management attributing the compression to the cost of full-rack integration. Growth of two-thirds at under five cents of gross margin on the dollar is the definition of activity without value capture. Hon Hai's Q1 2026 printed a 6.18% gross and 3.57% operating margin with a 44-day cash-conversion cycle and negative NT$32.6 billion of quarterly free cash flow, while cloud and networking products — principally AI servers — became the largest of its four product lines. Super Micro's quarter is the clearest illustration of the financing consequence: revenue fell sequentially from $12.7 billion to $10.2 billion, gross margin was 9.9%, and the business consumed $6.6 billion of operating cash, ending with $1.3 billion of cash against $8.8 billion of debt and convertible notes. Growth at 10% gross margin funded on the balance sheet is a financing business with an assembly hobby.

Low margin here does not automatically mean low return on capital, and that is the most useful single thing this layer teaches. Celestica converts a 9.8% operating margin into roughly 38% return on capital employed, because contract manufacturing employs very little permanent capital and runs favourable working-capital terms. That is a genuinely good business by the return measure. It is still not a durable one, because the return depends on utilisation and customer retention in a business with no switching cost, and the same operating leverage that produces 29% on the way up produces losses on the way down. The mechanism is set out numerically in §II·8, where Celestica and Super Micro — same layer, same margin band, same quarter — are shown to have opposite economics for one reason.

Rack-scale liquid integration is the current wedge and it is temporary by construction. It is valuable while leak testing, thermal commissioning and supply coordination are immature; as OCP manifold and rack specifications, reference designs and a growing qualified supplier base mature through 2027–28, the premium decays and the durable value migrates upstream to the architecture and downstream to controls, commissioning and service. Falsified if two integrators expand gross margin and earn above 15% incremental return on capital through the Rubin transition and its successor without vendor financing or rising customer concentration.

Electrical and mechanical contracting

Enormous growth, contractually capped economics. Quanta Services reported $9.56 billion of second-quarter revenue against $6.77 billion a year earlier, with $451.4 million of net income — a 4.7% net margin — $1.1 billion of adjusted EBITDA, $33.6 billion of remaining performance obligations, $53.4 billion of total backlog and $0.9 billion of free cash flow. MYR Group earned $49.9 million on $1.08 billion, 4.6%. EMCOR earned a 10.6% operating margin, the best of the group, on record remaining performance obligations of $17.14 billion, up 43.9%. These are labour-and-materials businesses where the customer's contract structure determines the margin, and the binding constraint is skilled electricians, pipefitters and commissioning engineers rather than anything proprietary. Backlog here is volume visibility rather than margin visibility, and coverage ratios below 1.0× say so.

Two names in this layer sit against the report's own criterion and therefore need their defence made explicit rather than waved through. Comfort Systems earned $441.6 million on $3.27 billion of second-quarter revenue — a 13.5% net margin, extraordinary for a mechanical contractor — with $1.14 billion of operating cash flow in a single quarter and backlog at $14.06 billion against $8.12 billion a year earlier. Sterling Infrastructure earned an 11.6% net margin with an 18.8% EBITDA margin, the best structure among the infrastructure contractors read here. The defence is not that either has a moat in the structural sense. It is that in a market where skilled trade labour is the scarce input, a contractor with a trained self-perform workforce and modular prefabrication capacity can price to scarcity for as long as the labour shortage lasts. That is a real advantage and a temporary one, and both are graded as a condition rather than a franchise, with a labour-normalisation falsifier attached: net margin falling below 8% while volume growth holds. Naming the tension is the point; a grade that sits against the report's own test is either wrong or needs its reasoning shown.

Colocation real estate — and the neoclouds beside it

The mechanism here is arithmetic rather than competition. Equinix reported 11% monthly recurring revenue growth as reported and normalised — in a year when the layer above it grew 92% — with annualised gross bookings up 23% and a record 9,700 net interconnections added in the quarter. On the consistent method used throughout this report, its trailing operating margin is 21.8% and its return on capital employed is 4.5%, computed from $2,138 million of trailing-twelve-month operating income taxed at 21% over $37,256 million of capital employed ($41,076m of assets less $3,820m of current liabilities), all inputs re-derived from SEC XBRL company-facts this session. It has a genuine asset in interconnection density and that network effect is real and defensible. It is also a modest business attached to a very large capital base, and the AI buildout is largely bypassing the product, because hyperscale training capacity is self-built or purpose-leased in multi-hundred-megawatt blocks rather than taken as retail colocation.

The strongest counter-argument is one of the places the panel split hardest, and it belongs on the record. Equinix's hyperscale platform is reported to have added more than 500 MW of capacity, grown revenue 42% in a recent quarter, and booked demand out to 2029 on fifteen-year terms, with management framing hyperscale capture as preserving the retail business's returns. Consensus Assumption — those figures reached the underlying session through trade summaries rather than the company's own supplement, and are flagged accordingly. They do not change the verdict, because the verdict is about returns rather than growth: a fast-growing slice of a development pipeline sits inside a consolidated 11% revenue growth rate and a 4.5% return on capital employed. If hyperscale leasing lifts that return above 8%, the grade is wrong and should be changed. Two of the three readings behind this report graded Equinix advantaged on the interconnection franchise; the merged grade follows the return arithmetic and marks the split.

The neoclouds sit in the same layer and are structurally worse. CoreWeave's $99.4 billion revenue backlog is genuinely enormous and its adjusted EBITDA margin of 56% is genuinely high. Below that line, the quarter produced a $144 million operating loss, $536 million of net interest expense and a $740 million net loss, against $36.4 billion of net property and equipment funded with $25.1 billion of total debt, $17.3 billion of it long-term, and $10.1 billion of operating-lease liabilities; legacy delayed-draw facilities carried effective rates of 9–15% and an April 2026 senior issue priced at 9.75%, though a newer $8.5 billion facility at SOFRThe Secured Overnight Financing Rate — the benchmark US short-term interest rate that floating-rate loans are priced against. A loan at "SOFR plus 225 basis points" pays the benchmark plus 2.25 percentage points, and the total moves whenever the benchmark does. The spread over the benchmark is the market's price for the borrower's risk, which makes it a cleaner read on creditworthiness than the headline rate. +225 basis points or fixed 5.9% shows real funding improvement and a $3.1 billion May loan was dedicated to two non-investment-grade customers. The model is an interest-rate and depreciation-schedule bet dressed as a technology company: it buys a three-to-six-year asset with borrowed money against contracts, and the entire equity value depends on the residual valueWhat an asset is worth once the contract it was bought for has run out. For a building or a turbine there is decades of evidence to draw on. For a four-year-old AI accelerator there is none, because none has yet reached that age in quantity — so the number sitting in the accounts is an assumption, and the equity in these businesses is essentially a wager on it. of accelerators at the end of a contract nobody has yet observed at scale.

Nebius supplies the strongest counter-case in the layer and it deserves the credit: its Meta arrangement is a five-year, $12 billion dedicated-capacity commitment beginning in 2027 with an option for as much as $15 billion more, and obligates Meta to buy unsold deployed capacity in defined circumstances — genuine risk transfer, not a backlog number. Its Q1-26 revenue of $399 million grew 684% with ARRAnnual recurring revenue — the most recent period's contracted revenue multiplied out to a full year. It flatters a fast-growing business by treating the latest month as though it had run for twelve, and it says nothing about whether the underlying contracts are cancellable. Useful as a growth signal, misleading as a revenue figure. at $1.9 billion and FY26 ARR guided to $7–9 billion, funded partly by roughly $4.3 billion of convertible notes. Contract tenor can term out demand risk. It cannot make a five-year accelerator live for a thirty-year building's life, and that mismatch is the layer's defining exposure. Applied Digital moved toward long-duration hyperscaler leases and more than 1 GW of contracted capacity while losing $249.2 million over the fiscal year on $611.3 million of revenue.

II·7The vertical-integration threat map

Every function on this chain that a hyperscaler can internalise, it eventually will — if the function is large enough in its own cost base, technically reachable, and not protected by a physical barrier. Rank the layers on those three tests and the sequence becomes obvious, and so does where it stops.

The first row of the table below turns on a distinction worth stating plainly, because the whole internalisation argument rests on it. A merchant accelerator — the GPU — is a general-purpose machine: it is good at a very wide range of AI work, which is precisely why one company can sell the same part to everybody and charge accordingly. A custom ASIC is the opposite bargain. It is designed around one buyer's own workloads, so it can be cheaper and more efficient at those and is worth little for anything else, and the design belongs to the buyer rather than to the vendor who helped build it. That is why a hyperscaler moving to custom silicon caps what the merchant vendor can charge without giving the ASIC supplier the pricing power it took away — the ASIC house is paid a fee for engineering, not a toll on a franchise.

Evidence: Amazon Q2-2026 release; Broadcom Q2-FY26 release and Q3 guidance; Vistra Q1-2026 release; Constellation Q1-2026 release and the Microsoft PPA announcement; AEP Q2-2026 release; Eaton Q1-2026 release; Ultra Ethernet Consortium and UALink specification pages; Open Compute Project Foundation release of 29 April 2026 and the ESUN 1.0 release of March 2026; Amkor Q2-26. Method: ranking is report-authored, applying three tests — share of the buyer's cost base, technical reachability, and the presence of a physical barrier. Amazon's "$25 billion" phrasing is quoted verbatim because the release does not define the segment boundary; it should not be read as a disclosed Trainium revenue figure. Synthesis: the internalisation sequence runs from the top of the stack downward and stops dead at the two layers defended by physics. Note what happens to the layers above the stopping point: they get bigger.

Vertical integration by the customer does not damage every upstream supplier — it damages the one immediately above the integrated function and strengthens the ones further up. A hyperscaler that designs its own accelerator still buys TSMC wafers, still buys TSMC packaging, still buys SK hynix HBM, still buys Amphenol connectors, still buys Vertiv power shelves, and now additionally buys more Cadence and Synopsys seats and more Arm architecture. It has removed one intermediary's margin and added work everywhere else. The foundry is indifferent to whose logo is on the die. The memory maker is indifferent. The connector supplier is indifferent. The EDA vendor is strictly better off.

Which means the standard framing — that open interconnect and custom silicon are primarily a threat to the merchant accelerator vendor — is only half the transaction. They are also a subsidy to the design tolls, the foundry and the signal-integrity layer, and they raise the value of orchestration software, observability and proven qualification precisely because the hardware underneath has become interchangeable. That is why the durable-profit map in this report concentrates upstream of the layer under attack rather than beside it.

II·8Margin versus return on capital — the inversion

Two businesses on this chain have almost exactly opposite economics and appear as neighbours in any margin table: a contract manufacturer at 9.8% and a landlord at 21.8%. Count the capital and they change places.

Return on capital answers where the money ends up. It does not answer how much capital had to be fed in to keep it there — and on this chain, that second question separates businesses more sharply than margin does.

Four economic animals that margin tables render as neighbours

The mandate for this work is explicit that gross margin must never stand in for durable profit, and this chain makes the point unusually concrete. Four businesses appear side by side in any margin screen and are not remotely the same animal.

The second is a memory maker at a cyclical peak whose next node consumes every dollar it just earned. Micron deployed $7.8 billion of gross property, plant and equipment purchases in the quarter it earned $33.3 billion of operating income, has committed to a long-term US programme in the hundreds of billions, and holds $22 billion of customer deposits that are financing cash flows rather than free cash flow. On the raw arithmetic its incremental return on the capital just deployed is a number no business has ever sustained — and it is meaningless, because Figure 4 already showed that roughly 58% of that operating income is price rather than volume. Price is not a return on the capital just deployed. It is a return on somebody else's shortage.

The third is an accelerator vendor whose depreciation sits on its customers' balance sheets rather than its own. NVIDIA carries $12.4 billion of net property and equipment against $259.5 billion of assets and generated $48.6 billion of free cash flow in a quarter, which is why its return on capital is high despite a capital base that includes $42.3 billion of non-marketable equity investments earning nothing operationally. The deployed asset — the thing that actually depreciates on a three-to-six-year clock — is owned by the hyperscaler or the neocloud. That incidence, not a gross-margin comparison, is the reason an accelerator vendor, a foundry, an ODM, a turbine fleet and a regulated utility are different economic animals.

The fourth is a regulated utility earning an authorised return on a rate base it is expanding for a single tenant. AEP disclosed 69 GW of new load additions through 2030 and around 13 GW of secured gas turbine capacity with another 10 GW under evaluation, alongside "up to $16 billion in expected cost offsets from load growth." A regulated utility growing rate base against a single load class is not capturing AI's economics. It is being paid an authorised return to build infrastructure for someone else, and it carries the stranded-assetCapital sunk into something that turns out not to be needed — a substation built for a campus that never fills, a plant contracted for load that never arrives. For a regulated utility the question is not whether the asset works but whether a commission will still let it charge customers for it, and that argument is fought long after the concrete has set. risk if that someone else's growth flattens. NextEra grew Florida Power & Light's regulatory capital employed about 9.3% year over year and filed to combine with Dominion Energy — consolidation is the rational response to a capital programme this large, and it is not the same thing as pricing power.

The working-capital mechanism, in one comparison

Who finances the inventory, not the margin, sets the assembly layer's return on capital. Two companies in the same layer, in the same margin band, in adjacent quarters show it cleanly.

Evidence: Celestica Q2-2026 condensed consolidated financial statements, quarter ended 30 June 2026 — revenue $4,698.6m, cost of sales $4,121.1m, accounts receivable net $3,338.2m, inventories $3,401.7m, accounts payable $3,819.7m. Super Micro FQ3-2026 Form 8-K Exhibit 99.1, quarter ended 31 March 2026 — net sales $10,243,014k, cost of sales $9,224,334k, accounts receivable net $8,413,396k, inventories $11,103,376k, accounts payable $3,686,991k, cash $1,290,324k against $5,169,911k at 30 June 2025. Method: DSO = receivables ÷ revenue × days; DIO and DPO = inventory and payables ÷ cost of sales × days; cash conversion cycle = DSO + DIO − DPO. Celestica's quarter is 91 days, Super Micro's 90; the one-day difference moves each figure by under 1.5%. Balance-sheet items are period-end rather than average, which overstates days for a business growing this fast — equally for both. Synthesis: Celestica's payables run longer than its inventory; Super Micro's run a third as long, which is why one converts a 9.8% margin into a 38% return and the other converts a 9.9% gross margin into a $6.6bn hole. When volume growth stops, the same mechanism reverses — and it reverses faster for the one with 146 days in the pipe. As a general test for this layer, watch inventory days plus receivable days minus payable days rather than AI revenue growth: on a $20bn annual cost base, a two-week lengthening consumes roughly $0.8bn of cash ($20bn × 14 ÷ 365).

On incremental return, the honest answer is that this chain will not yield a clean number at a cycle peak, and saying so is more useful than computing one. At one extreme Micron's arithmetic is meaningless for the reason given above. At the other, CoreWeave carries $36.4 billion of net property and equipment against a quarterly operating loss, so its incremental return on the buildout is, so far, negative. Between those poles sit the only two readings on this chain that mean anything: TSMC, which converts roughly two-thirds of operating cash flow into the next node and has held a 60%-plus operating margin while doing it, and Broadcom, which converts 1% of revenue into capital and returns 46% of revenue as free cash flow. Those are the two genuine compounders here, and they compound for opposite reasons — one by out-spending everyone, one by not having to spend at all.

Two honest limits close this section. Comparable company-level through-cycle and incremental returns on capital were not verified for most names, because segment invested capital, the sustaining-versus-expansion capex split, and cycle-normal tax inputs are not disclosed consistently. One of the three readings behind this report initially published cross-layer ROIC ranges and a scatter and withdrew both on review, on the correct grounds that a disclosed margin cannot be promoted into a through-cycle return merely by attaching an estimated denominator. No grade in this report rests on an undisclosed ROIC estimate. The return figures used are computed on one stated, reproducible definition from filed data, at a single point in time, and their limitations are stated wherever they appear — and the missing through-cycle series remains a falsifier rather than a hidden input.

II·9The names — 105 rows, 104 grades, by layer

Every grade below rests on that company's own primary source. Grades are on competitive economics, not on price. Where the three independent readings behind this report did not agree, the row carries a split marker and the dissent is set out in §II·9b.

Reading conventions. † marks a lighter read — a single current filing or release and one decisive datapoint rather than a full-cycle file, a foreign filing reached through a summary, or a period that has closed but not yet reported — and the confidence is reduced accordingly. Split means at least one of the three readings reached a different grade on the same disclosures; the merged grade is adjudicated on evidence rather than by vote, and §II·9b names the dissent and the evidence behind it. Where a grade rests on the layer's structure rather than on the company's own numbers, the row says so in those words. Acquired names are status-verified and carry no live standalone grade.

Layer 1 · Semiconductor capital equipment and subsystems

Layer 2 · Foundry, packaging and test

Layer 3 · Processors, accelerators and design tolls

Layer 4 · Memory and storage

Layer 5 · Server, rack and ODM integration

Layer 6 · Networking, optics and interconnect

Layer 7 · Thermal, mechanical and data-centre build

Layer 8 · Power electronics, electrical equipment and grid hardware

Layer 9 · Generation, fuel and power markets

Layer 10 · Data-centre real estate and neoclouds

Evidence: every row cites the company's own most recent opened primary; full sources are in the evidence register. Method: grades apply one consistent test — does this company hold a right in its own transaction that survives the capacity, standards, integration or regulatory response its returns are inducing? Rows marked † rest on a lighter read and their confidence is reduced accordingly; rows marked split are those where the three independent readings behind this report did not agree, and §II·9b names the dissent and the evidence on each side. One row — Wistron — rests on the layer's structure rather than the company's own numbers and says so on its face; one — Chart Industries — is no longer separately listed. Synthesis: 104 names graded, and the advantaged names cluster in four layers rather than ten. Half the roster is neutral, which is the finding rather than a hedge. No advantaged name exists in server and rack assembly at all, and that empty cell is as load-bearing as any positive verdict here.

II·9bWhere the panel split — the contested grades, adjudicated

Thirty-seven rows carry a split marker. Each is set out below with the dissenting grade, the evidence behind it, and why the merged verdict landed where it did — because a grade that survived three independent reads is worth more than one that did not, and the reader is entitled to know which is which.

Four patterns account for almost all of the disagreement, so the table below groups by mechanism rather than listing name by name.

Evidence: the primaries named in each grade row above, plus the XBRL re-derivations described in §II·8. Method: adjudication rules applied in order — (i) a grade resting on that company's own opened primary beats one resting on a lighter read or sector reasoning; (ii) a definitional fit beats a vote, so a name meeting this report's stated definition of neutral is graded neutral even where two readings said otherwise; (iii) where a name contains both a contracted and a queued component, the components are graded separately in §II·5 and the ticker takes the grade of the component that dominates its disclosed economics; (iv) no grade is set by averaging, and where the panel remained genuinely divided the row says so and both sides' evidence is stated. Synthesis: the splits are not random. All 37 fall into four mechanisms, and three of the four are versions of the same question — whether a number measured at a single point in a capital cycle can settle a claim about a right that lasts twenty years. That question has no clean answer, which is why the split markers stay in the tables rather than being resolved away.

II·10Incentives, the bear case argued properly, and the alternatives rejected

Many outcomes on this chain become obvious once the incentives are laid out. Start there, then argue the strongest case against everything above, then answer it.

The incentive map

For each major participant: the incentive, and its consequence for where the profit lands.

Evidence: the disclosures cited throughout Parts I and II for each participant — hyperscaler capex guidance and internal-silicon disclosures; NVIDIA's Form 10-Q commitments and equity-investment note; CoreWeave's Q1-26 10-Q; TSMC's Q2-26 capex allocation; Micron's strategic customer agreements; the UEC, UALink and OCP specification pages; Eaton, ABB, Hitachi Energy and Siemens Energy plant announcements; PJM's 2028/29 auction; FERC Dockets EL26-67 through EL26-72; AEP Ohio Schedule DCT; BIS licensing actions. Method: a qualitative incentive map, report-authored, in which each row states the participant's own economic objective and the consequence that follows from it for profit location. Synthesis: read down the consequence column and the profit map falls out of the incentives before any financial statement is opened. Four participants have an incentive to add supply into their own shortage, and all four have published the date they will.

The steelmanThe opposite of a straw man: building the strongest possible version of the argument against your own position — the one its ablest advocate would make, with its best evidence — and then answering that rather than a weaker substitute. It is a discipline rather than a courtesy. An argument that has only ever beaten the weak version of its opposition has not been tested.: the strongest case against everything above

The strongest argument against this report is not that AI will fail. It is that the buildout is now funded by capital markets rather than by cash flow, and capital markets change their mind faster than fabs get built.

First, the accounting. The economics of this buildout depend on a depreciation schedule nobody has yet validated. Hyperscalers assume five-to-six-year useful livesThe number of years over which a company spreads the cost of an asset in its accounts. Stretch the assumption and each year's charge is smaller and reported profit is larger; shorten it and the reverse. Nothing about the machine changes — only the accounting estimate. Which is why an assumption about how long an accelerator lasts is simultaneously a claim about how much money the owner made last quarter. — Amazon's shareholder letter says so explicitly — on assets whose replacement cadence is annual. If the true economic life is three to four years, reported earnings across the group are overstated by a very large number; one widely circulated estimate puts the understatement of true depreciation at roughly $176 billion across 2026–2028, and projects Alphabet's depreciation rising from about $21 billion in 2025 to roughly $78 billion by 2029, from 5% of revenue to 11%. Third-party estimate, not verified Nobody has yet observed what a four-year-old training accelerator is worth in a liquid secondary market at scale, because the scale did not exist four years ago.

Second, the funding. Incremental annual debt has risen from around 9% of hyperscaler capital expenditure in fiscal 2024 to roughly 32% on a trailing basis by mid-2026, and equity has re-entered the mix — Alphabet's $49.6 billion raise being the most striking instance. Moody's has reported approximately $662 billion of data-centre lease commitments signed but not yet commenced, sitting off balance sheet under the lease-commencement standard, a figure larger than the combined on-balance-sheet debt of the same companies. Third-party estimate, not verified The verified core of this point needs no estimate at all: Alphabet minus $5.86 billion of free cash flow, Amazon minus $7.6 billion trailing, Oracle minus $23.7 billion, Meta at $784 million on $31.86 billion of operating cash flow with $349.3 billion of commitments. A buildout funded from operating cash has one constraint. A buildout funded from credit markets has two, and the second is not under management's control.

Third, the revenue gap. Roughly $700–735 billion of hyperscaler capital expenditure in 2026 sits against direct AI revenue estimated in the tens of billions. That gap can be closed by growth, and the cloud numbers are genuinely accelerating — Google Cloud +82% at a 35.6% segment margin, AWS +36.7% and its fastest in eighteen quarters, Oracle cloud infrastructure +77%. But it has to be closed, and the closing has to outrun a depreciation charge that compounds.

Fourth, the historical mechanism nobody can date. Telecom fibre did not die of insufficient demand. It died because wavelength-division multiplexing multiplied the capacity of already-installed glass without new capital. Every efficiency gain in model architecture, inference serving, quantisation, distillation, speculative decodingSoftware techniques for getting more useful work out of the same chips. Inference is running a trained model to answer a question, as opposed to training it in the first place. Quantisation stores the model's numbers at lower precision so more fits in memory and each calculation is cheaper. Distillation trains a small model to imitate a large one. Speculative decoding has a fast model guess ahead while a slow one checks the guesses. None of them requires new hardware — which is exactly what makes them the analogue of the technology that destroyed the fibre owners. and scheduling is a candidate for the same role here. If the effective output of the installed accelerator base can be multiplied by software, the buildout overshoots without a single forecast having been wrong about demand.

Fifth — and this is the sharpest version — the chain's most defensible-looking nodes have the least room to absorb a pause. A foundry that must spend two-thirds of operating cash flow on the next node cannot cut capital expenditure without surrendering the yield-learning lead that constitutes the moat, because the moat is cumulative wafer starts. A memory maker at an 80% operating margin has no incremental price to give. An equipment vendor's revenue is a direct derivative of customer capital budgets, which are decisions rather than trends — and 2019 demonstrated that a double-digit revenue decline can arrive within two quarters with nothing wrong in the demand story. Tokyo Electron has already printed a 16.1% sequential sales decline on what management called a temporary customer pause; the bear does not have to imagine that mechanism, only extend it.

The answer

The bear case is right about the funding and right about the depreciation, and it does not overturn the conclusion — because the conclusion was never a bet on the level of spending. It is a claim about where the money stops moving when it does move.

On depreciation: it is a claim about hyperscaler earnings quality, not about upstream demand. A shortened useful life means the fleet is replaced faster, which is upstream demand rather than its absence. The version that hurts this chain is not a shorter life but a longer one — if accelerators turn out to be economically useful for eight years rather than four, replacement demand halves and every supplier above the hyperscaler feels it. The market has the sign of this risk backwards, and it is the single most under-discussed asymmetry in the sector.

On funding: this is the strongest part of the case and it is precisely why the profit map is drawn where it is. Capital-markets dependence transmits down the chain in a specific, predictable order, worked through in the next section: it hits the leveraged, short-asset, uncontracted layers first — neoclouds, then assembly working capital, then optical and memory spot pricing — and reaches EUV lithography, contracted utility rate base and twenty-year offtake last, if at all. A report whose durable set is concentrated in exactly the layers a funding shock reaches last is not made wrong by a funding shock; it is tested by one.

On the revenue gap: the correct comparison is not capital expenditure against AI-labelled revenue. It is capital expenditure against the operating income of the businesses the capital defends. Microsoft earned a 45.1% operating margin on $90.0 billion of quarterly revenue; Alphabet earned a 34% consolidated operating margin with cloud accelerating to 82% growth; AWS earned 39.4%. These are companies spending to defend franchises worth vastly more than the spend. That may still be a bad allocation decision. It is not the same financial structure as a competitive local exchange carrier borrowing against uncontracted fibre, and the distinction matters because the CLECs had no business to defend.

On the efficiency multiplier: unanswerable, and it should be stated as such rather than argued away. There is no dated test for it and no disclosure that would reveal it in advance. It is the single largest unhedgeable risk in the sector, and every demand-dependent verdict in this report is held below the top confidence tier because of it.

On the pause-absorption point: conceded, and it is the reason the equipment grades carry cyclicality falsifiers rather than being treated as unconditional. What the bear misses is the asymmetry of what survives. In a pause, ASML's installed-base management revenue keeps arriving — €2,762 million of a €9,326 million quarter — because service on an installed fleet does not pause when orders do. TSMC's yield advantage does not decay in eighteen months. Cadence's $4.2 billion of remaining performance obligations due within twelve months are contracted. Contrast that with a 146-day cash conversion cycle at a 9.9% gross margin, or $536 million of quarterly interest against a residual-value assumption. The bear case is a statement about when and the report is a statement about who, and the two are compatible.

Two alternatives rejected, with reasons

"This is a normal semiconductor capital-equipment cycle and will resolve like 2019." Partly right, and rejected as a complete account for two reasons. First, the 2019 correction was a customer capital-budget pause among a handful of chipmakers; the buyer set here is broader, more strategically motivated, and increasingly financed externally, which makes the demand path both more persistent and — when it turns — more credit-sensitive than a normal capex cycle. Second, and more decisively, the power-and-interconnection layer is not a semiconductor cycle at all. Turbine factories, transformer plants and nuclear restarts run on a completely different clock, measured in the same three-year lead times whether or not wafer starts pause. And EUV remains a physics monopoly regardless of where the cycle sits. Keep the normal-cycle model for the layers it genuinely describes — memory, optics, ODM assembly, and the WFE order book itself. Reject it as a description of the whole. Consensus Assumption on the mechanism recurring · Known Unknown on the timing

"Power, not silicon, is the binding constraint for the rest of the decade." Directionally correct on physics and wrong on economics, which is precisely the confusion this report exists to separate. Power is genuinely the tightest physical constraint: PJM short of its reliability requirement three years running, 160-week substation transformer lead times, gas turbine slots booked to fiscal 2028, and 410.6 GW of ERCOT interconnection requests against 5,778 MW energised. And the layer earns among the worst returns on this chain — GE Vernova at 7.2% return on capital employed against a $176 billion backlog, Constellation at roughly 4.5%, Vistra at roughly 7%, PJM capacity administratively capped $230/MW-day below its own estimated clearing level. A constraint whose price is set by a commission is a constraint, not a franchise. Two further corrections: silicon still dominates dollars per megawatt and still concentrates merchant profit in the accelerator, foundry and lithography layers; and "power" is not one profit pool but three different claims on the same electron, as §I·9b sets out. Power is binding on schedule. Silicon is binding on profit-pool size. Contract structure is binding on who keeps either. Known Known on the physical constraint · Estimated on the returns conclusion

Five non-consensus conclusions

Each names the consensus view it revises, rests on disclosed evidence, is falsifiable, and carries a confidence tag sized to the evidence rather than to how much the thesis wants it to be true.

1. 1Consensus revised: HBM is a new structural franchise. It is a contract-lengthened memory cycle. Qualification, co-design and the wafer-per-bit penalty raise the floor and stretch the lag; they do not repeal the incentive to ship the next wafer when a competitor is about to. All three suppliers are now qualified or qualifying into HBM4 simultaneously, which is the condition that finances the supply response rather than the one that prevents it. Evidence: Micron's ~$100bn of minimum-price RPO signed at the peak, SK hynix's agreements with ~ten customers, and disclosed capacity arriving 1H27 through 2028. Falsifier: industry HBM operating margins remaining at peak-cycle levels through year-end 2028 after the disclosed expansions complete. High confidence
2. 2Consensus revised: power scarcity means the owners of power keep the rent. The physical moat is strongest precisely where the political claim on it is strongest, and the two are not independent — scarcity is what summons the regulator. PJM has been administered at a cap for three consecutive years while running short of its own reliability requirement; FERC has opened proceedings across all six organised markets; state commissions are writing twelve-year take-or-pay tariffs. What survives is not the megawatt but the executed contract, and FERC's 18 June 2026 grandfathering of existing agreements dated that distinction to the day. Falsifier: FERC and state commissions permitting co-located and large loads to avoid system costs without political reversal, while capacity prices stay above replacement cost through 2030. High confidence
3. 3Consensus revised: open interconnect and custom silicon are primarily a threat to the merchant accelerator vendor. They are also, and simultaneously, a subsidy to everything upstream of it. Every custom accelerator built to escape merchant pricing must still be architected, verified, taped out, fabricated, packaged and fed with HBM — which expands the EDA, architecture-licence, foundry, packaging and signal-integrity tolls regardless of who wins the socket. Vertical integration damages the layer immediately above the integrated function and strengthens the layers further up. Evidence: Cadence's operating margin rising from 19.0% to 28.4% with $8.1bn of backlog; Arm's data-centre royalty more than doubling; Samsung attributing foundry improvement to HBM base-die demand — a memory component priced by a foundry. Falsifier: open scale-up deployments commoditising retimer, DSPDigital signal processor — a chip specialised for the arithmetic of cleaning up and reconstructing a signal. In optical networking it is what lets a fast transceiver recover data from a waveform degraded by distance and noise, and it is one of the margin pools an open interconnect standard could commoditise. and EDA economics as fast as they commoditise transceivers. Medium-high confidence
4. 4Consensus revised: the market's greatest risk in AI infrastructure is that accelerators depreciate too fast. The sign is backwards. A shorter economic life means a faster replacement cycle, which is upstream demand; the outcome that actually breaks this chain is accelerators proving useful for eight years rather than four, which halves replacement demand and strands the capacity built to serve it. The same inversion applies to the layer's most-quoted comfort: hyperscalers disclose five-to-six-year useful lives, and the market treats a downward revision as the bad news. Falsifier: a disclosed shortening of assumed useful lives accompanied by a decline in forward accelerator orders — which would show the replacement-demand channel does not operate. Medium confidence · the mechanism is clear, the observation is not yet available
5. 5Consensus revised: strict cost-causation regulation slows the AI buildout. It relocates it. A twelve-year take-or-pay tariff with posted collateral and a priced exit converts a hyperscaler's flexible option on grid power into a fixed liability — which makes behind-the-meter generation on the developer's own land, outside the commission's tariff, look cheaper than waiting in a queue for firm service. The transfer runs from the regulated utility to fuel cells, reciprocating engines and turbine OEMs, and its earliest visible reading is already in a filing: Bloom Energy's revenue up 165.5% with GAAP operating income swinging from a $3.5m loss to $182.2m. Falsifier: interconnection reform materially shortening grid connection timelines, which removes the premium for on-site generation and would show the transfer was a timing artefact. Medium confidence

II·11Second-order effects — the parts that decide the outcome

The transmission order, if hyperscaler capex growth merely flattens

Not falls — flattens. The distinction matters, because most of this chain is priced for continued acceleration and a great deal of it breaks on deceleration alone. A flattening is more dangerous than a decline for any supplier that extrapolated the growth rate into fixed capacity.

Circular revenue and vendor financeVendor finance is a supplier funding its own customer's purchase — by lending, by taking an equity stake, or by committing to buy something back. Circular revenue is what results when enough of that money returns as sales: the supplier's growth is partly financed by the supplier. Neither is improper and both are disclosed. What they do is corrupt the demand signal — an outside observer cannot tell how much of the order book is independent judgement and how much is the seller's own capital coming back around.

NVIDIA's non-marketable equity securities went from $22,251 million at the start of its fiscal first quarter to $42,336 million at the end — net additions of $17,899 million in three months — with $27 billion of further investment commitments expected through the remainder of fiscal 2027. In the same filing: $30 billion of multi-year cloud service agreement commitments, described as primarily supporting research and development, and $3,957 million of accrued purchase consideration relating to a non-exclusive licence agreement with Groq. On 1 July 2026 it introduced a revenue-sharing and credit-support model under which AI clouds procure NVIDIA systems, sell cloud services, and remit both ordinary product revenue and a share of cloud revenue back to NVIDIA, with Sharon AI and Firmus as initial participants. Separately, Amazon disclosed a $15 billion investment in OpenAI and a further $35 billion commitment subject to conditions.

So the largest supplier on the chain is simultaneously investing in its customers, pre-committing to buy cloud capacity from businesses that buy its accelerators, licensing technology from a competitor, and holding $119 billion of its own supply commitments. None of that is improper and all of it is disclosed. What it means is that a portion of the demand signal the whole chain reads is partly funded by the entity reading it, and the quality of that demand cannot be assessed from outside the filings.

The same structure appears one layer down, running in the opposite direction, and it is disclosed just as plainly. Oracle attributes most of a $638 billion remaining-performance-obligation balance to contracts in which the customer either prepaid for the accelerators or bought and supplied them directly — $75 billion of prepaid and customer-supplied hardware sitting inside the largest order book in cloud.

At the top of the chain a vendor is investing in its customers; one layer below, a cloud provider is being financed by its own. Neither is improper and both are disclosed. Together they mean the demand signal this chain reads is, in part, the chain financing itself, and the ordinary test of demand quality — did an independent buyer put its own money at risk — cannot be applied from outside the filings. That is not a reason to disbelieve the buildout. It is a reason to hold any confidence tag resting on reported order books one notch lower than the numbers alone would justify, which is exactly what §II·9 does. The evidence test that would resolve it is cash from unrelated end users and utilisation after support is withdrawn, and neither is disclosed.

Ratepayers, politics, and the cost-allocation fight

This is the single largest regulatory threat to the power layer's returns, and it is not primarily a FERC matter. FERC's 18 June 2026 show-cause orders address interconnection process, cost transparency and co-location terms across all six organised markets, with responses due 17 August 2026 — and they explicitly protect existing negotiated agreements. The harder fight is at the state commissions, where a utility proposes to add generation, transmission and distribution for a single load class, and where the answer to "who pays" is a political decision made by commissioners who are elected, or appointed by people who are.

The dynamic is straightforward and unfavourable to the load. A hyperscale campus brings tax base and few permanent jobs, consumes water and land, and requires infrastructure whose cost — absent a specific tariff with minimum-take and exit provisions — spreads across every ratepayer in the territory. AEP's framing of "up to $16 billion in expected cost offsets from load growth" is precisely an attempt to pre-empt that objection. Whether commissions accept it determines whether the regulated layer's growth is accretive or merely large. The evidence so far says they are converging on the Ohio template: 85% minimum billing demand, twelve-year outside terms, thirty-six-month exit fees, 50% collateral, with Virginia's GS-5 at fourteen years and Georgia attaching explicit ratepayer protections to nearly 9.9 GW of approved new demand.

The second-order consequence runs upstream. Strict cost causation does not stop the buildout — it makes behind-the-meter and on-site generation dramatically more attractive, because a developer facing a four-year interconnection wait and a full upgrade bill will buy fuel cells, reciprocating engines and turbines instead. That transfers value from the regulated utility to Bloom Energy, Caterpillar, Cummins and the turbine OEMs, and it is already visible in Bloom's second quarter. The third-order consequence favours the best-capitalised buyers: a hyperscaler that can post collateral and survive a slower ramp absorbs a twelve-year minimum bill; a smaller neocloud faces a higher all-in cost of power and less access to the queue. Regulation designed to protect ratepayers ends up concentrating the industry.

Sovereign policy, export control, and the licence that nobody used

The export-control picture changed twice in the period covered, and in the same direction. On 15 January 2026 BIS published a final rule revising the licence review policy for advanced computing commodities to China and Macau from a presumption of denial to case-by-case review, covering H200- and MI325X-class chips and less advanced ones, conditioned on certifications about domestic supply sufficiency, non-diversion of foundry capacity, recipient security procedures and independent third-party performance testing in the United States. In May 2026 it clarified that licence obligations follow D:5 and Macau-headquartered counterparties even when they operate elsewhere. On 14 July 2026 it moved the United Arab Emirates from Country Groups D:3 and D:4 into A:5, granting approved UAE entities licence-free access to advanced computing items. Enforcement remained real: a $252 million penalty against Applied Materials tied to illegal equipment exports.

Set that against NVIDIA's own guidance for the July 2026 quarter, which states plainly that the company "is not assuming any Data Center compute revenue from China." A liberalised licence policy paired with a zero-China assumption is the most informative pair of facts in the export-control file: the constraint on China revenue is no longer purely the American rule. Whether the binding constraint is licence throughput, Chinese procurement policy, or domestic substitution cannot be determined from the disclosures opened here, and no conclusion in this report rests on resolving it. What can be said structurally: controls preserve the leading-edge gap while shrinking Western vendors' accessible demand, raise compliance cost, and finance Chinese domestic tool investment at mature nodes. They reshape WFE geography. They create no second EUV supplier, and nothing in the current rules suggests they could.

Sovereign fab and grid policy runs the same way. CHIPS-style subsidies raise foundry capital intensity and create political optionality without quickly creating a second leading-edge source for AI packages; what they reliably do is lower a competitor's cost of capital and depress the return on everybody's incremental mature-node capital. Resilience improves. Global utilisation and incremental foundry return can both fall.

Labour, water and the constraints nobody prices

The binding input in the mechanical and electrical build is skilled trade labour, and it is the reason Comfort Systems earns a 13.5% net margin while Quanta Services earns 4.7% on comparable growth: self-perform capability and prefabrication capacity price to a shortage that no amount of capital resolves quickly. Comfort Systems' backlog grew from $8.12 billion to $14.06 billion year over year; Trane's Americas Commercial HVAC backlog grew 90%; Johnson Controls' backlog grew 32% organically. That much booked work chasing the same electricians, pipefitters, protection engineers and commissioning teams is an inflation mechanism, and it lands on the developer's budget rather than the contractor's margin. It is also, notably, the one bottleneck on this chain with no announced capacity response, because apprenticeships take years and no filing discloses one.

Water is the constraint that binds locally and politically rather than economically. Direct-to-chip loops can approach low on-site consumption, but heat still reaches a dry cooler, a cooling tower or another sink, and siting trades water against electricity efficiency. Closed loops reduce consumption; they do not remove heat rejection, noise, pipelines or the requirement for community consent. Neither water nor labour creates an investable monopoly. Both constrain megawatts, which means both push the energisation date to the right — and a report whose durable set depends on who holds the contract rather than how fast the build goes is comparatively insensitive to them.

II·12Falsifiers, dated predictions, and what to watch

Every verdict above is written to be broken. Here is what would break each one, on what date, and what the status was at the evidence cutoff.

Per-call falsifiers, with decision dates and status at the cutoff

Eight dated, testable predictions

1. 1By 31 December 2027, supplier commentary on HBM shifts from "sold out" to mix and qualification, and at least two of the three suppliers guide rising bit shipments with flat-to-down blended HBM average selling prices against 2026 — the first clear cycle tell. Basis: Micron's Singapore packaging contribution in 1H27 and Idaho and Taiwan output from mid-2027, SK hynix's Yongin Phase 1 opening early 2027, Samsung's H2-2027 ramp. Estimated
2. 2By the end of 2028, memory gross margins across the three suppliers will be at least 25 percentage points below their 2026 peaks. Basis: Micron's 84.6% gross and 80.4% operating margin and SK hynix's 76.3% operating margin against announced capacity landing from H2 2027; and the arithmetic in Figure 4, which shows how much of the current print is price. Estimated — the mechanism has held in every prior cycle; only the timing has ever varied
3. 3ASML's gross margin will still be below TSMC's in 2029. The EUV monopolist will continue to under-monetise, because ecosystem health is worth more to it than rent, and its own capacity expansion is the disclosed proof of that preference. Falsified by ASML gross margin above 62% in any year through 2029. Estimated
4. 4Merchant OSAT leading-edge packaging operating margin will remain below 15% through 2028. Basis: Amkor at 10.5% in Q2-26 with gross margin improving from 12.0% to 16.8% — the co-optimisation argument says mix improves and price does not. This is the cleanest single test of the one chokepoint where the panel genuinely split. Estimated
5. 5By 31 December 2028, at least one commercial open scale-up system using UALink or OCP ESUN ships at meaningful cloud scale — and it gains first in inference and custom-silicon systems rather than in frontier training. Basis: UALink evaluation hardware expected late 2026 with products in 2027; OCP ESUN 1.0 published with 175-company participation; five of ten UEC steering members being buyers. Estimated
6. 6By mid-2029, large-load tariffs at least as strict as AEP Ohio's — 85% minimum take, a term beyond eight years, a priced exit and posted collateral — apply in states representing more than half of announced US data-centre load, and behind-the-meter generation orders rise as a direct consequence. The weaker version of this is already resolved and should not be offered as a test: Schedule DCT took effect on 23 July 2025, Virginia's GS-5 begins in 2027 with a fourteen-year obligation, and Texas has published proposed §25.194 standards requiring $50,000/MW of security. What remains genuinely open is the coverage. Estimated on coverage · Known Unknown on the behind-the-meter response, which no issuer discloses cleanly
7. 7US large power transformer lead times fall below 100 weeks by the end of 2028, and liquid-cooling hardware gross margins normalise while service and controls revenue grows faster. Basis: roughly $1.8bn of announced North American transformer capacity landing 2027–28 against 128-week power transformer and 160-week substation lead times today; OCP blind-mate and rack specifications reaching broad multi-vendor production. Expect at least two rack integrators to describe standard manifolds and CDUs as multi-sourced by then. Estimated
8. 8Server and rack assembly operating margins remain inside a 4–11% band through 2029 regardless of volume, and at least one listed neocloud restructures, is acquired, or materially reprices its GPU-collateralised debt before the end of 2028. Basis: six assemblers between 3.57% and 10.7% in a single quarter at the peak of the strongest demand in the layer's history — structure, not cycle; and CoreWeave's $25.1bn of debt against negative GAAP operating income, Applied Digital's $249.2m FY26 net loss, and Nebius scaling on the same structure. Known Known on the assembly band · Speculative on the neocloud timing — the mechanism is clear, the date is not

What to watch, and when

Ask of any node on this chain: if demand were fully satisfied tomorrow, would this business still earn more than its cost of capital? For EUV, leading-edge foundry, the accelerator system, the design tolls and the electrical aftermarket, the answer is yes. For memory, turbine slots, transformers, assembly and uncontracted capacity, it is no — and every one of those is currently building the capacity that answers the question.

The closing test

99Evidence register

317 distinct sources, unioned from three independent readings of this subject. Every one was opened and read in the session that produced the reading it came from, on or before 30 July 2026.

Tier 1 = issuer filing, issuer press release, regulator or market-operator publication, or standards-body specification. Tier 2 = reputable trade, wire, law-firm or industry-association reporting, used for timing and colour rather than to carry a structural conclusion. Tier 3 items — single-source, aggregated or estimated — are not listed here; they appear in the "Not verified" section below and carry no conclusion. Of the 317 sources, 300 are Tier 1. Sixteen were independently opened by more than one of the three readings and are marked corroborated; those are the strongest evidence in the report, because two or three analysts reached the same document from different directions and read it the same way.

Two register conventions. Distinct sections of a single filing — a Form 10-Q's revenue-concentration note and its commitments note, for example — are counted as distinct citations, because they carry distinct facts. And where a fact originates with an issuer or regulator, the register links that document rather than a trade article repeating it; the seventeen Tier 2 entries are those where the primary was either unavailable in the session or genuinely secondary in nature (a consultancy lead-time dataset, a market-monitor estimate, an aggregation of guidance).

Demand side — hyperscalers, cloud buyers and their capital plans — 24 sources

• Alphabet 10-Q, 23 Jul. 2026 · Tier 1
• Alphabet Q4 call, 4 Feb. 2026 · Tier 1 · corroborated across 2 independent reads
• Amazon 10-Q, 1 May 2026 · Tier 1
• Amazon 2025 shareholder letter, Apr. 2026 · Tier 1
• Amazon nuclear plans, updated 2026 · Tier 1
• AMD–Meta, 24 Feb. 2026 · Tier 1
• AMD–Microsoft, 20 Jul. 2026 · Tier 1
• atmeta.com · Tier 1
• Constellation–Microsoft PPA, 20 Sep. 2024 · Tier 1
• finance.yahoo.com · Tier 2
• investor.vistracorp.com · Tier 1
• Meta 10-Q, 30 Jul. 2026 · Tier 1
• Meta Q2 call, 29 Jul. 2026 · Tier 1
• Microsoft 10-K, 29 Jul. 2026 · Tier 1
• Microsoft FY2026 Q3 call, 29 Apr. 2026 · Tier 2 · corroborated across 2 independent reads
• nebius.com · Tier 1
• Nebius–Meta 6-K, 16 Mar. 2026 · Tier 1
• Oracle (ORCL) · Tier 1
• Oracle closed fiscal 2026 on 10 June 2026 · Tier 1
• Oracle FY2026 results, 10 Jun. 2026 · Tier 1 · corroborated across 2 independent reads
• q4cdn Amazon PDF · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1

Accelerators, design tolls, networking silicon and optics — 36 sources

• AMD (AMD) · Tier 1
• AMD filing exhibit · Tier 1
• Amphenol (APH) · Tier 1
• Applied Optoelectronics (AAOI) · Tier 1
• Arista Networks (ANET) · Tier 1
• arista.com · Tier 1
• Arm (ARM) · Tier 1
• Astera Labs (ALAB) · Tier 1
• Broadcom (AVGO) · Tier 1 · corroborated across 2 independent reads
• Broadcom Q2, 3 Jun. 2026 · Tier 1
• Cadence (CDNS) · Tier 1
• Ciena (CIEN) · Tier 1
• Cisco (CSCO) · Tier 1
• Coherent (COHR) · Tier 1
• Corning (GLW) · Tier 1
• Credo (CRDO) · Tier 1
• Fabrinet (FN) · Tier 1
• Intel (INTC) · Tier 1
• investors.broadcom.com · Tier 1
• Lumentum (LITE) · Tier 1
• Marvell 10-Q, 28 May 2026 · Tier 1
• marvell.com · Tier 1
• MISO auction · Tier 1
• misoenergy.org · Tier 1
• NVIDIA 10-Q, filed 20 May 2026 · Tier 1 · corroborated across 2 independent reads
• NVIDIA financing model, 1 Jul. 2026 · Tier 1
• NVIDIA Vera Rubin update, 21 Jul. 2026 · Tier 1
• nvidianews.nvidia.com · Tier 1 · corroborated across 2 independent reads
• nvidianews.nvidia.com · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• Synopsys (SNPS) · Tier 1

Foundry, advanced packaging, test and capital equipment — 67 sources

• 10 August 2026 · Tier 1
• Advanced Energy (AEIS) · Tier 1
• Advantest FY2025 results, 27 Apr. 2026 · Tier 1
• AEP Ohio's Schedule DCT took effect on 23 July 2025 · Tier 1
• Amkor (AMKR) · Tier 1
• Applied Digital Q3, 8 Apr. 2026 · Tier 1
• Applied Materials (AMAT) · Tier 1
• ASE Technology (ASX) · Tier 1
• ASML (ASML) · Tier 1
• ASML Q2 results, 15 Jul. 2026 · Tier 1 · corroborated across 2 independent reads
• ASML Q2 transcript, 15 Jul. 2026 · Tier 1
• asml.com · Tier 1
• BIS license policy · Tier 1
• BIS, 29 Aug. 2025 · Tier 1
• bis.gov · Tier 1
• bis.gov PDF · Tier 1
• Camtek (CAMT) · Tier 1
• cnbc.com · Tier 2
• constellationenergy.com · Tier 1
• dell.com · Tier 1
• Eaton Q1, 30 Apr. 2026 · Tier 1
• Entegris (ENTG) · Tier 1
• ERCOT Batch Zero, 18 Jun. 2026 · Tier 1
• FormFactor (FORM) · Tier 1
• GE Vernova Q2, 22 Jul. 2026 · Tier 1 · corroborated across 2 independent reads
• GlobalFoundries (GFS) · Tier 1
• hitachienergy.com · Tier 1
• hklaw.com · Tier 2
• Ichor (ICHR) · Tier 1
• insidelines.pjm.com · Tier 2
• investor-relations earnings library · Tier 1
• ir.kla.com · Tier 1
• KLA (KLAC) · Tier 1
• Lam Research (LRCX) · Tier 1
• mcguirewoods.com · Tier 2
• newsroom.lamresearch.com · Tier 1
• OCP–ASHRAE liquid-cooling alliance, 13 Oct. 2025 · Tier 1
• Onto Innovation (ONTO) · Tier 1
• opencompute.org PDF · Tier 1
• PJM auction report, 27 Jul. 2026 · Tier 1
• pjm.com PDF · Tier 1
• Q1-26 Form 6-K, 5 May 2026 · Tier 1
• Q2-26 Form 6-K, 29 July 2026 · Tier 1
• q4cdn Vertiv PDF · Tier 1
• Samsung Electronics (SSNLF/GDR) · Tier 1
• Samsung Electronics' Device Solutions division earned KRW 89.2 trillion of operating profit on KRW 127.5 trillion of revenue · Tier 1
• Samsung PDF · Tier 1
• Samsung, 12 Feb. 2026 · Tier 1 · corroborated across 2 independent reads
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• SEMI year-end forecast · Tier 2
• SEMI, 7 Apr. 2026 · Tier 2
• Sterling Infrastructure (STRL) · Tier 1
• Teradyne (TER) · Tier 1
• the Open Compute Project approved eight contributions on 29 April 2026 · Tier 1
• Tokyo Electron (TOELY) · Tier 1
• TSMC monthly revenue, posted 10 Jul. 2026 · Tier 1
• TSMC PDF · Tier 1
• TSMC PDF · Tier 1
• TSMC PDF · Tier 1
• TSMC Q2 results, 16 Jul. 2026 · Tier 1
• TSMC Q2 transcript, 16 Jul. 2026 · Tier 1
• Ultra Clean (UCTT) · Tier 1
• UMC (UMC) · Tier 1

Memory and storage — 10 sources

• Micron (MU) · Tier 1
• Micron FQ3 prepared remarks, 24 Jun. 2026 · Tier 1
• Micron, 24 Jun. 2026 · Tier 1
• micron.com · Tier 1
• news.skhynix.com · Tier 1 · corroborated across 2 independent reads
• SanDisk (SNDK) · Tier 1
• Seagate (STX) · Tier 1
• sec.gov · Tier 1
• SK hynix, 22 Apr. 2026 · Tier 1
• Western Digital (WDC) · Tier 1

Systems, ODM, EMS and integration — 16 sources

• Celestica (CLS) · Tier 1
• Dell (DELL) · Tier 1
• Flex (FLEX) · Tier 1
• Hon Hai / Foxconn (HNHPF) · Tier 1
• HPE (HPE) · Tier 1
• Jabil (JBL) · Tier 1
• Quanta (QUCCF) · Tier 1
• Sanmina (SANM) · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• Super Micro (SMCI) · Tier 1
• supermicro.com · Tier 1
• Wistron (WICOF) · Tier 1

Thermal, mechanical, electrical equipment and construction — 32 sources

• "Exceptional demand — external headwinds currently affecting profitability." · Tier 1
• ABB (ABBNY) · Tier 1
• Atkore (ATKR) · Tier 1
• Eaton (ETN) · Tier 1
• Eaton, 8 Apr. 2026 · Tier 1
• eaton.com · Tier 1
• EMCOR (EME) · Tier 1
• GE Vernova conference transcript, 18 Mar. 2026 · Tier 1
• Generac (GNRC) · Tier 1
• Hitachi (HTHIY) · Tier 1
• Hubbell (HUBB) · Tier 1
• Modine (MOD) · Tier 1
• Munters Q1, 23 Apr. 2026 · Tier 1
• MYR Group (MYRG) · Tier 1
• new.abb.com · Tier 1
• nVent (NVT) · Tier 1
• Powell Industries (POWL) · Tier 1
• Prysmian / Encore Wire (PRYMY) · Tier 1
• prysmian.com · Tier 1
• Q2-26 press release, 16 July 2026 · Tier 1
• Q2-26 press release, 30 July 2026 · Tier 1
• Quanta Services (PWR) · Tier 1
• Schneider Electric (SBGSY) · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• Siemens Energy (SMNEY) · Tier 1
• Siemens Energy Q1, 11 Feb. 2026 · Tier 1
• Siemens Energy Q2, 12 May 2026 · Tier 1
• siemens-energy.com · Tier 1
• siemens-energy.com · Tier 1
• Trane Technologies (TT) · Tier 1
• Vertiv (VRT) · Tier 1

Power, generation, fuel, real estate and neoclouds — 30 sources

• AEP Ohio tariff, effective 23 Jul. 2025 · Tier 1
• American Electric Power (AEP) · Tier 1
• Applied Digital (APLD) · Tier 1
• Baker Hughes closing release, 16 Jul. 2026 · Tier 1 · corroborated across 2 independent reads
• Caterpillar (CAT) · Tier 1
• Cheniere Energy (LNG) · Tier 1
• Constellation Energy (CEG) · Tier 1
• Constellation Q1 2026 presentation, 5 May 2026 · Tier 1
• CoreWeave 10-Q, 7 May 2026 · Tier 1 · corroborated across 2 independent reads
• CoreWeave financing release, 18 May 2026 · Tier 1
• Cummins (CMI) · Tier 1
• Dominion Energy (D) · Tier 1
• Entergy (ETR) · Tier 1
• Equinix (EQIX) · Tier 1
• Nebius (NBIS) · Tier 1
• nebius.com · Tier 1
• NextEra Energy (NEE) · Tier 1
• NRG Energy (NRG) · Tier 1
• NuScale Power (SMR) · Tier 1
• Oklo (OKLO) · Tier 1
• Oklo announced in May 2026 that the NRC approved the Principal Design Criteria topical report · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• Southern Company (SO) · Tier 1
• Talen Energy (TLN) · Tier 1
• Vistra (VST) · Tier 1
• world-nuclear-news.org · Tier 2

Regulators, market operators, standards bodies and official statistics — 33 sources

• [37a] · Tier 1
• BIS EAR Part 748, current through Jul. 2026 · Tier 1
• bis.gov PDF · Tier 1
• DOE transformer program, accessed 30 Jul. 2026 · Tier 1
• EIA's analysis of its Annual Energy Outlook 2026, published 19 May 2026 · Tier 1 · corroborated across 2 independent reads
• ERCOT April large-load report, 13 May 2026 · Tier 1
• ercot.com PDF · Tier 1
• federalregister.gov · Tier 1
• federalregister.gov · Tier 1
• FERC co-location order · Tier 1
• FERC rehearing summary, 18 Jun. 2026 · Tier 1
• FERC six-RTO proceedings, 18 Jun. 2026 · Tier 1
• ferc.gov · Tier 1
• ferc.gov PDF · Tier 1
• four directives, confined to large loads that are not co-located with generation · Tier 1
• Georgia PSC fact sheet, Dec. 2025 · Tier 1
• IEA sensitivity cases, 10 Apr. 2025 · Tier 2
• IEA, 10 Apr. 2025 · Tier 2 · corroborated across 2 independent reads
• IEA, 16 Apr. 2026 · Tier 2
• iea.org · Tier 2
• MISO corrected results, 22 May 2026 · Tier 1
• NRC Crane status, accessed 30 Jul. 2026 · Tier 1
• NRC Palisades status, accessed 30 Jul. 2026 · Tier 1
• OCP ESUN 1.0, 10 Mar. 2026 · Tier 1 · corroborated across 2 independent reads
• OCP rack specifications, accessed 30 Jul. 2026 · Tier 1
• PUC Project 58481, the proposed 16 TAC §25.194 large-load interconnection standards published for comment on 12 March 2026 · Tier 1
• SEMI mid-year forecast · Tier 2
• UALink published four further specifications in April 2026 · Tier 1
• UALink white paper, Jan. 2026 · Tier 1
• ualinkconsortium.org · Tier 1
• UEC specification history, 16 Jul. 2026 · Tier 1 · corroborated across 2 independent reads
• Ultra Ethernet is now at specification 1.0.3, last updated 17 July 2026 · Tier 1
• Virginia SCC, 2025 order · Tier 1

Further primary releases, trade reporting and derived data — 69 sources

• Baker Hughes completed its acquisition of Chart Industries on 16 July 2026 · Tier 1
• Bloom Energy (BE) · Tier 1
• Comfort Systems (FIX) · Tier 1
• data.sec.gov · Tier 1
• Digital Realty (DLR) · Tier 1
• GE Vernova (GEV) · Tier 1
• Google–Kairos–TVA, 18 Aug. 2025 · Tier 1
• Iron Mountain (IRM) · Tier 1
• Johnson Controls (JCI) · Tier 1
• Mitsubishi Heavy (MHVYF) · Tier 1
• MKS Instruments (MKSI) · Tier 1
• powermag.com · Tier 2
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• sec.gov · Tier 1
• tomshardware.com · Tier 2
• trendforce.com · Tier 2

Additional derived-data source: the SEC XBRL company-facts API at data.sec.gov, used for the balance-sheet and trailing-twelve-month operating-income inputs behind Figures 14 and 15, and re-queried during this merge for Micron, GE Vernova and Equinix to adjudicate three numeric divergences between the underlying readings. One endpoint per company.

100Not verified — and carrying no conclusion

The union of all three readings' unverified lists, merged and de-duplicated. Every item below is material to the subject and was not verified against an opened primary source. Each appears above only where explicitly flagged, and no verdict in this report rests on any of them.

Comparable through-cycle and incremental returns on capital. Segment invested capital, the sustaining-versus-expansion capital split, and cycle-normal tax inputs are not disclosed consistently by any of these companies. One of the three readings published cross-layer ROIC ranges and a scatter and withdrew both on review, on the correct grounds that a disclosed margin cannot be promoted into a through-cycle return by attaching an estimated denominator. The return figures used here are computed on one stated, reproducible definition from filed data at a single point in time, with their four limitations stated wherever they appear. No grade rests on an undisclosed ROIC estimate, and the missing through-cycle series remains a falsifier rather than a hidden input.

Dollars of equipment per energised AI megawatt. No common, audited, non-overlapping measure exists. Issuers do not disclose each component on one scope, and HBM cannot be added separately to accelerator modules without double-counting. One reading published a $38–65m/MW range and withdrew it; another published a schematic percentage build-up and marked it Tier 3 with no verdict attached. §II·2 gives the bounds that can be derived from disclosure and states plainly that the component allocation cannot be. The Open Compute Project's Open Data Center Design Guide supplies a physical rack, pod and megawatt reference and no dollar allocation.

SEMI equipment forecasts. The 300mm fab equipment figures ($133bn 2026, $151bn 2027, $155bn 2028, $172bn 2029; $156bn total equipment sales in 2027; the $104bn record wafer-fab-equipment year; the $144bn 2026 WFE figure) reached the sessions partly through search-result summaries of SEMI press releases; semi.org returned HTTP 403 to one reading. Treat as directionally indicative only. Nothing in the profit-pool map or any grade depends on them.

Advanced-packaging capacity series. TSMC's exact end-2026 CoWoS wafers-per-month trajectory is not published in the quarterly documents opened, and third-party estimates of it are analyst reconstructions. The packaging verdict is written against a qualitative capacity test and a margin falsifier precisely so that it does not depend on this series.

HBM capacity-to-demand ratios and contract prices. Exact global HBM capacity-to-demand ratios, customer-by-customer qualification tables, comparable capacity denominators and HBM-only returns on capital are proprietary and undisclosed. Supplier qualification status, pricing direction and dated capacity are verified from the suppliers' own releases and carry the conclusion; unlicensed third-party point estimates do not.

Memory capacity timing detail. The "second half of 2027" ramp for new Samsung and SK hynix fabs, Micron's Idaho mid-2027 DRAM start, the January 2026 New York groundbreaking and the aggregate announced investment figures come partly from search summaries rather than from the companies' own capital-plan disclosures. The direction — that announced capacity has been committed and lands after 2026 — is corroborated by SK hynix's own release (M15X acceleration, Yongin Phase 1 opening early 2027) and by Micron's own disclosed quarterly capital expenditure. Prediction 2 is dated to end-2028 partly to absorb this uncertainty.

Hyperscaler depreciation and leverage estimates. The $176bn depreciation-understatement estimate, the projection of Alphabet depreciation rising to roughly $78bn by 2029, the 9%→32% incremental-debt-to-capex shift, and Moody's approximately $662bn of signed-but-not-commenced lease commitments are third-party analyses surfaced via search whose source documents were not opened. They are presented in §II·10 as the bear case's own claims, attributed as such, and support no verdict. The verified facts in that section — Alphabet's −$5.86bn free cash flow and $49.6bn raise, Amazon's −$7.6bn trailing free cash flow, Meta's $83.66bn of long-term debt and $349.3bn of commitments, Oracle's −$23.7bn — come from the companies' own releases.

High-NA EUV adjacencies and per-tool pricing. ASML's own disclosure that Intel Foundry is running High-NA in high-volume manufacturing on select 18A layers with matching yields is Tier 1 and carries. The adjacent claims — that TSMC will skip High-NA for its initial A14 node and extend standard EUV with multi-patterning, the EXE:5200B first production shipment timing, Samsung tool placement, and the $350–400m per-tool price — reached the sessions through trade summaries and are Tier 3. No verdict rests on any of them; the EUV verdict rests on the absence of any competing programme, which is verifiable as a negative.

Liquid-cooling competitive entry. Reports that Schneider, Lenovo, Super Micro and Chinese coolant-distribution-unit vendors have moved from dependence on the incumbents to building their own control-software platforms reached the sessions through trade summaries rather than issuer filings. They are stated in §I·7 as a mechanism to watch rather than a fact, and they are the reason the cooling grades carry margin-based rather than growth-based falsifiers.

Equinix hyperscale platform metrics. The reported >500 MW of added hyperscale capacity, 42% revenue growth in a recent quarter and demand booked to 2029 on fifteen-year terms reached the sessions through trade summaries rather than the company's supplement. They are stated in §II·6 as a challenge to the grade rather than as a fact, and the grade carries an explicit numeric trigger — return on capital employed above 8% — at which it should be changed.

Broadcom's custom-ASIC customer count. Reporting of "five customers expanding to a sixth," the named prospects, and a "$73bn AI backlog" or "$100bn 2027 AI revenue" line of sight was not verified against Broadcom's own filings. Its verified disclosures — $10.8bn of Q2 AI semiconductor revenue, $16.0bn guided for Q3, $8.4bn in Q1-FY26, top-five end customers ≈45% of first-half revenue — are sufficient for the grade and are used instead.

Foreign and lighter-read filings. ASE, Arm, Nebius, NRG, Siemens Energy's Q3-FY26 detail, Schneider Electric's own release (which returned an access error), Hitachi's annual filing, Tokyo Electron's filing text, Hon Hai's margin statement, Quanta Computer's filing, Wistron in its entirety, Fabrinet's margin detail, HPE's segment detail, Digital Realty's metrics, Iron Mountain's segment detail, Generac's margin detail and Caterpillar's data-centre content are each flagged † in the grade tables. Their grades carry reduced confidence accordingly. Siemens Energy's FY2025 figures, by contrast, are from the company's own release and are Tier 1.

Wistron specifically. No company primary was retrievable by any of the three readings. Its grade rests on the layer's structure as evidenced by six other assemblers' own filings, and the row says so in those words rather than implying a read that did not happen. One structural grade out of 104 is a disclosed limitation; it carries no conclusion of its own, because the layer verdict it sits inside is independently supported by Dell, Super Micro, Celestica, Flex, Jabil, Sanmina, Hon Hai and Quanta Computer.

ERCOT capacity pricing. A directly comparable ERCOT capacity-auction price does not exist, because ERCOT is an energy-only market without capacity payments — one reading reached for a substitute and withdrew it on review. This report uses the operator's own large-load queue stages, the Batch Zero reform and the published resource-adequacy assessment instead. Note also that the three readings used three different vintages of the ERCOT queue (410.6 GW at 26 March, 445.8 GW in April, >438 GW at the June Batch Zero approval); they are reported as the sequence they are, and the earlier third-party summary one reading carried (a 238.6 GW queue with ~3.9 GW energised) is superseded by the operator's own filing.

Turbine slot cancellation economics and nearline lead times. Exact cancellation terms across turbine slot-reservation agreements are not comparable across vendors and are not disclosed; backlog and slot reservations are therefore not treated as identical firm orders anywhere in this report. Current nearline-HDD lead-time weeks were not verified from a primary source; the storage grades rest on issuer structure and capital economics rather than on that datapoint.

Nuclear restart dates. The exact commercial restart date for Crane or Palisades is not verified. NRC milestones — SAFSTOR status, licensing-basis restoration, component restoration — carry the analysis, not announced aspirations. Talen's Amazon primary page returned an access error to one reading and the relationship is summarised from widely reported terms rather than a re-opened release body.

Standards shipment data. Ultra Ethernet's own site records specification 1.0.3 as of 17 July 2026; UALink's records four specifications published in April 2026 beyond the 200G 1.0 base; the Open Compute Project Foundation's 29 April 2026 release lists the eight contributions approved that day, and its ESUN 1.0 release is dated March 2026. What none of them publishes is the one thing that would settle the commercial question: which vendors have shipped silicon conforming to each specification, and in what volume. Consortium membership is not a shipment. The falsifiers in §II·12 are written to be testable against whatever the actual shipping date turns out to be.

GPU collateral and neocloud contract economics. Loan-level GPU collateral liquidation values, and Nebius's contract margin and cancellation economics, remain undisclosed. The neocloud grades rest on disclosed debt, interest, lease liabilities, losses and contracted power rather than on any residual-value assumption.

NVIDIA's 10-K in full. One reading encountered an access error on the full 10-K HTML on sec.gov and cited customer-concentration percentages from the filing family rather than re-extracting them line by line; another opened the Form 10-Q for the quarter ended 26 April 2026 directly and read the concentration and commitments notes. The Q1-FY27 figures used throughout — three direct customers at 21%, 17% and 16%, and $119bn of supply commitments — come from the 10-Q and are Tier 1.

Amazon's Q2-2026 release timing. One reading's evidence freeze began before Amazon's Q2-2026 release was retrievable from the issuer site and used the Q1 filing and the April 2026 shareholder letter instead; another opened the Q2-2026 release of 30 July 2026 directly. The merged report uses the Q2 release, which is why Amazon's capex figure here is approximately $220bn rather than the ~$200bn the April letter carried — and the revision itself is shown in Figure 9 rather than being silently resolved.

Evidence cutoff 30 July 2026. Fiscal periods differ across the companies covered and every comparison names its periods; nothing in the grade tables should be read as like-for-like across different fiscal calendars. Return-on-capital figures are single-point estimates at what is, for several of these businesses, a cyclical peak, computed as return on capital employed on the consistent method stated in Figure 14 and not on each company's own definition. This document is an internal structural assessment of industry economics and competitive position — not investment advice, not a recommendation, not a solicitation, and not an offer to buy or sell any security.