How AI Data Centers Get Built, From Chip to Power Line, and Why Memory and Grid Power Now Set the Pace

What this primer argues, and the numbers that size it

An AI data center can only be built as fast as its slowest input. In 2026 two inputs are slow at once. One is memory chips: their makers spent the last slump cutting capacity, and they need two to three years to add more. The other is grid power, where a new connectionA grid connection is the physical link between a site and the high-voltage grid, usually a substation and the lines to it, together with the studies and approvals the utility and the grid operator must finish before power flows. Section 14 walks through the steps., turbine or transformerA machine that changes the voltage of electricity. Power travels long distances at very high voltage and is stepped down in stages before it reaches a building. The large units used at power plants and substations are built to order. takes three to six years.

1. 1The spending is real and still rising. Amazon, Alphabet, Microsoft, Meta and Oracle spent about $586 billion on buildings and equipment in their latest four reported quarters, and three of them guide to $530–555 billion for calendar 2026 alone. High confidence
2. 2Memory is short now. Micron's quarterly revenue went from $11.3 billion to $54.2 billion in a year and its gross margin from 44.7% to 86.8%. SK hynix and Samsung both describe demand running ahead of what they can make. High confidence
3. 3Power decides where and when buildings fill. PJM, the grid operator for 67 million people, cleared its 2028/29 capacity auction (where power plants are paid to be available at peak demand) at its price cap and 6,831 MWMegawatt: a million watts. Watts measure power, the rate at which electricity is drawn at a given moment; the total used over time is counted in watt-hours. A gigawatt (GW) is a thousand megawatts. short of its reliability requirement, while only 525 MW of new generation and uprates (upgrades that raise an existing plant's output) cleared. High confidence
4. 4The slow steps are slow for physical reasons. GE Vernova's gas-turbine backlog and reservations (paid deposits that hold a production slot) equal about 5.8 years of its 2026 output; power transformers averaged 128 weeks to deliver in 2025 and generator step-up transformers (the units that connect a power plant to the grid) 144 weeks; a typical US power plant built in 2024 took 55 months from grid request to operation. High confidence
5. 5Most announced capacity is an option, not a commitment. AEP Ohio, an electric utility, received requests for more than 30,000 MW against about 600 MW of data-center loadIn the power industry, load is the electricity customers draw, measured in megawatts. A data center connecting to the grid is a new load; a power plant is new supply. in 2024. New tariffs (the terms and rates a regulator approves for a class of utility customer) that make large users pay for 85% of reserved capacity are designed to shrink that gap. Moderate confidence
6. 6Risk is moving into leases and supplier guarantees. Microsoft, Oracle and Amazon report about $754 billion of leases signed but not yet started. Nvidia has guaranteed up to $105 billion of payments for one OpenAI campus, against the $8.1 billion customer-financing bookCredit an equipment maker extends to its own customers so they can buy its products: loans, deferred payments or guarantees of their debts. If the customers fail, the equipment maker takes the loss. Section 21 tells Lucent's story. that Lucent, then a leading telecom equipment maker, carried into the telecom bust. Moderate confidence
7. 7The money and the delay sit in different places. On a campus filled with Nvidia's newest systems, roughly half of each capital dollar becomes accelerator-vendor gross profit (an accelerator is the processor chip built to run AI models; Nvidia is the largest seller); across all buyers, Nvidia's data-center gross profit was about a third of the five largest buyers' capexShort for capital expenditure: money spent on buildings and equipment that will be used for years, as opposed to day-to-day running costs.. The grid, the slowest step, is mostly paid for through regulated electricity rates (prices a public regulator sets so the utility recovers its approved costs from customers' bills) over decades, outside the capital budgets that are growing fastest. Moderate confidence

Most people can say that AI runs on Nvidia chips in very large buildings. That sentence leaves out a chain of more than a dozen industries, each with its own factories, suppliers and waiting times. The chain starts with a lithography machineThe machine that prints a chip's circuit pattern onto silicon by shining light through a patterned mask. The most advanced kind, which uses extreme-ultraviolet light, is made only by ASML. Section 5 explains how it works. that one company builds, runs through a chip factory that takes years to reach good yieldsYield is the share of chips on a silicon disc that work when they come off the line. A new manufacturing process starts with low yields, which climb as engineers find and fix the causes of defects. and a memory stack twelve or sixteen chips high, and ends at a transformer, a turbine and a grid connection that each take years to arrive. This primer explains how each of those steps works, why it is hard, who does it, how long it takes to add capacity, and what happens to everything else when one of them stalls.

The buildout can only move as fast as its slowest step, and the slowest step changes. In 2023 and 2024 it was advanced packaging, the step that joins the processor to its memory. By 2025 most of the industry pointed to power. The evidence in this report supports a more specific version: in 2026 memory and power are both binding, for different reasons and on different clocks. A step is binding when it sets the limit for the whole chain: adding capacity anywhere else would not get more AI computing switched on until that step expands. Memory limits how many accelerators can ship this quarter. Power limits where a finished building can be switched on and how soon.

The size of the thing

Capital spending (capex: money spent on buildings and equipment that will be used for years) is the input that sets demand for every layer below it. The five companies in Figure 1 spent about $586 billion on property and equipment over their latest four reported quarters. Microsoft's spending on property and equipment for the fiscal year to June 2026 was $115.9 billion, up 80% from $64.6 billion a year earlier, according to its fiscal fourth-quarter earnings release¹. Amazon's gross purchases over the twelve months to June 2026 were $173.0 billion, 61% more than in the prior twelve months, and its second-quarter release² says the increase "primarily reflects investments in artificial intelligence." Oracle spent $28.5 billion in the single quarter to August 2026, more than half of its $55.7 billion total for the whole previous fiscal year, according to its first-quarter fiscal 2027 release³.

Forward guidance points higher. At its May 2026 developer conference Alphabet said it expected 2026 capex of "approximately $180 to $190 billion," about six times the $31 billion it spent in 2022, according to Sundar Pichai's keynote post⁴. Meta narrowed its 2026 range to $130–145 billion in its second-quarter release⁵. Amazon raised its 2026 figure to about $220 billion on its July call, citing higher memory prices, according to reporting of the call⁶; that figure comes from the call rather than a filing and is the least firm of the three. Two of these companies are now also raising outside money for it: Alphabet issued $49.6 billion of new equity in June 2026 "to scale AI infrastructure," and Oracle raised $19.9 billion through an at-the-market share programA way of selling new shares a little at a time into the stock market at the going price, instead of in one large offering. in its latest quarter, according to their releases.

The money shows up quickly in the chain's revenue lines. Nvidia reported data-center revenue of $89.0 billion for the quarter to July 2026, up 117% from a year earlier, in its CFO commentary⁷. TSMC, which makes most of the world's leading-edgeMade on the newest manufacturing processes, which pack the most transistors into each chip. Only a handful of companies can make chips this way. chips, reported in its 2025 annual report on Form 20-F⁸ that high-performance computing, the category that includes AI accelerators and server processors, supplied 58% of its 2025 revenue, up from 43% in 2023. It cannot be read as the AI share precisely, because the category also includes ordinary server and PC processors, but it shows where the leading-edge factories now point.

Electricity is where the chain meets everyone else. Lawrence Berkeley National Laboratory estimated, in a report released by the Department of Energy in December 2024⁹, that US data centers used 176 terawatt-hours in 2023, about 4.4% of US electricity, and would use between 6.7% and 12% by 2028. A terawatt-hour is a billion kilowatt-hours, so the range spans a lot of power plants. The Energy Information Administration's September 2026 outlook¹⁰ forecasts record US electricity sales in 2026 and 2027 "driven by data center development and increased manufacturing activity." Worldwide, the International Energy Agency¹¹ put data centers at about 1.5% of electricity use in 2024 and projected consumption would more than double by 2030.

A newer US forecast is higher. The Electric Power Research Institute (EPRI), a research body for the power industry, projected in February 2026 that data centers would use 9% to 17% of US electricity by 2030, up from 4% to 5% now, in its Powering Intelligence 2026¹² scenarios. Those estimates are 60% higher than its scenarios of two years earlier. The range is nearly a factor of two wide, which shows how uncertain the buildout's pace is.

The rest of Part I explains the machine they describe, one layer at a time, starting with what AI work actually asks of hardware. Part II puts the layers together and asks which one binds. Part III looks at what comes next.

The chain on one page

Table 1 lists the fourteen layers in the order this primer teaches them, from the demand at the top to the land and rules at the bottom. The fourth column is the one to watch. It gives a rough time to add new capacity at that layer, taken from the sources cited in each layer's own section. The spread runs from about a year for a new chip generation to five years or more for a grid connection or a place in the queue for a gas turbine.

Read down the table and three kinds of step appear. The first kind is design and software. Nvidia, the chip teams at the hyperscalers (the largest cloud-computing companies), and the companies that sell design tools add capacity by hiring engineers and buying computers. They release a new product roughly every year and are rarely the reason a building sits empty.

The second kind is factory work: lithography machines, wafer fabsChip factories ("fab" is short for fabrication plant). A wafer is the thin disc of silicon on which many chips are made together before being cut apart., packaging lines, memory fabs. Here capacity comes from cleanrooms and tools that take one to three years to build and fill. These industries have a history of building too much in booms and too little afterwards, and the memory industry's 2023 slump is the reason memory is short in 2026. Section 9 tells that story.

The third kind is heavy infrastructure: transformers, switchgearCabinets of heavy circuit breakers that route power around a site and cut off any section that develops a fault., turbines, transmission lines and the grid connection itself. These take the longest. Part of the delay is physical, since a large transformer is built to order from specialised electrical steel by skilled workers who are in short supply, and a gas turbine needs a slot in a factory that is booked for years. Part of it is institutional. A grid connection needs studies by the utility and the grid operatorAn independent organisation, such as PJM or ERCOT, that runs the high-voltage grid and the wholesale power market across a region. Utilities own the local wires and serve customers; the grid operator coordinates the whole system and studies each new connection., a tariff approved by a state commissionThe state agency, often called a public utility commission, that sets the prices and terms of service for the utilities in its state., and rules set by the Federal Energy Regulatory Commission (FERC), the federal agency that regulates interstate electricity transmission. Each of those parties works on its own calendar.

Three distinctions run through the rest of this report. Designing a chip and manufacturing one are different businesses. Nvidia designs; TSMC manufactures; they have different costs, different risks and different margins. A chip and the system it ships in are different products. Nvidia now sells most of its data-center output as whole racksA rack is a wardrobe-sized steel cabinet into which servers, network switches and power equipment slide like drawers. Data-center power and cooling are planned rack by rack., and its annual report¹³ says its data-center platform is "typically delivered to customers as rack-scale systems." And capacity that has been announced is different from capacity that has been built, connected to the grid and filled with working equipment. A gigawatt in a press release and a gigawatt drawing power can be four or five years apart, and many announced gigawatts never arrive at all.

A gigawatt (GW) is a billion watts, or a thousand megawatts (MW). Data-center capacity is usually quoted as the IT load: the power available to the computers themselves, before cooling and other overhead. When a press release says "a 1-GW campus," it usually means that much IT load, and the building will draw somewhat more from the grid.

What AI work demands

A modern AI model is a very large table of numbers, called parameters or weights, plus a fixed recipe for multiplying an input by them. Training is the process of setting those numbers. The model reads a piece of text, guesses the next word, measures how wrong the guess was, and nudges every weight slightly in the direction that would have made it less wrong. It does this trillions of times. Each nudge is a long run of multiplications and additions, and the industry counts that work in floating-point operations, or FLOP: one FLOP is one multiplication or addition on a number with a decimal point.

The industry keeps buying more of this work because of a pattern observed in experiments rather than derived from theory. In 2020 a team led by Jared Kaplan reported that a language model's error falls as a smooth power law as you increase the model's size, the amount of training data and the compute spent, with some trends "spanning more than seven orders of magnitude" (Kaplan et al., 2020¹⁴). A power law means each fixed improvement costs a multiple of the compute that bought the last one. Two years later a team led by Jordan Hoffmann found that models of the time were undertrained, and that "for every doubling of model size the number of training tokens should also be doubled" (Hoffmann et al., 2022¹⁵). Their 70-billion-parameter model, Chinchilla, beat a 280-billion-parameter model trained with the same compute. Since compute is roughly size times data, doubling both quadruples it.

Epoch AI, a research group that tracks these quantities, estimates that the compute used to train frontierThe largest and most capable AI models at any given time, built by a handful of AI labs such as OpenAI and Anthropic. "Frontier labs" are the companies that build them. language models has grown about fivefold a year since 2020, roughly ten thousandfold in total, and that the power drawn by the largest training runs has been doubling each year, to "tens to hundreds of megawatts" (Epoch AI trends page, updated February 2026¹⁶). Two efficiency trends push the other way. Epoch estimates AI chip performance per dollar has improved 49% a year since 2023, and that better algorithms let the same result be reached with about three times less compute each year. So far the appetite has grown faster than the efficiency. Whether that continues is one of the open questions in Part III.

Why one chip is never enough

A frontier model does not fit on one chip, and training it on one chip would take centuries. So the work is split. The model's layers, or slices of each layer, are spread across many accelerators, and the training data is split across many copies of the model. After every small batch of data the copies must share what they have learned before taking the next step. That sharing is why AI clustersA cluster is a large group of accelerators, up to tens of thousands, wired together so they can work on one job as a single machine. need networks that carry far more data than ordinary cloud servers ever did, and why the fastest links are reserved for small groups of chips that must talk constantly. Section 11 explains how those networks are built. In practice, this means training demand arrives in very large, synchronised blocks: tens of thousands of accelerators that must be installed, powered and connected together before any of them is useful.

Inference: the usage side

Inference is the use of a trained model to answer requests. It has two phases with different physics. In the first, called prefill, the model reads the whole prompt at once, which keeps the arithmetic units busy. In the second, called decode, it produces the answer one token at a time, and for every token it must read its weights out of memory again. A token is a word or part of a word, the unit models read and write. A team led by Pratyush Patel described the split precisely: "a compute-intensive prompt computation, and a memory-intensive token generation" (Patel et al., Splitwise, 2023–24¹⁷). By splitting the two phases across different machines they reported 2.35 times the throughput for the same cost and power.

Decode is where memory bandwidth, the rate at which data can be read from memory, becomes the limit. Nvidia's Rubin accelerator carries up to 288 gigabytes of high-bandwidth memoryMemory built for AI and other high-performance chips, usually shortened to HBM. Twelve or sixteen memory chips are stacked and placed right beside the processor, so data can move between them very fast. Section 9 explains how it is made. read at up to 22 terabytes a second, according to Nvidia's January 2026 technical description¹⁸. Reading all 288 gigabytes once at that rate takes about 13 milliseconds. A model that filled the memory could therefore produce at most about 75 steps per second for a single stream of text, however fast the arithmetic units were. Batching many users' requests together shares each read across more tokens, which is how operators make decode economical, but the bandwidth ceiling is real. This is why the memory industry, rather than the logicLogic chips compute: processors and accelerators. Memory chips store data. The two are made in different factories, often by different companies. industry, has become the most stretched part of the chain in 2026.

Long conversations add a second memory load. A model keeps a running record of the conversation so far, called the key-value cache, and it grows with every token of context. When it outgrows the accelerator's own memory it spills to ordinary server memory and then to solid-state drives. Micron's September 2026 results release¹⁹ reports that its fastest data-center drives "are now shipping to market-leading customers for KV-cache applications," a use that did not exist as a product category three years ago.

How fast usage is growing

The clearest public measure of usage is Google's count of tokens processed. In May 2024 it was 9.7 trillion tokens a month. A year later it was roughly 480 trillion. In May 2026 it was "over 3.2 quadrillion per month," seven times the year before, according to Alphabet's I/O 2026 keynote post⁴. Figure 2 plots those three points on a logarithmic scale, where each gridline is ten times the last.

Some of that growth is more people using AI. Some is each use consuming more tokens. Reasoning models write out long internal chains of steps before answering, and agentic systems, in which a model calls tools and other models to finish a multi-step task, can spend thousands of tokens on what looks to the user like one request. Both raise compute per task. Both also lean on memory, because every one of those tokens is a decode step.

From tokens to watts

Every layer below converts this demand into a physical quantity. Arithmetic becomes logic wafers at TSMC. Bandwidth becomes stacks of high-bandwidth memory. The need for thousands of chips to act as one becomes switches, cables and lasers. And all of it becomes electricity. Nvidia's product page for its Vera Rubin rack describes a "100 MW AI Factory" housing 40,000 Rubin accelerators, according to Nvidia's Vera Rubin NVL72 page²⁰. That is 2.5 kilowatts of facility power per accelerator once cooling, networking, storage and losses are included, about three and a half times the 700 watts of the chip that started the boom, the H100.

The hyperscalers, the AI labs and the neoclouds, which rent out AI computing capacity, set demand for the whole chain. They are the only buyers who see end-user demand directly, and they place their orders years ahead, because the slowest layers below them need years of notice. Microsoft's 2026 annual report²¹ describes this plainly: the investments are "in advance of fully developed revenue streams," and its data centers "depend on the availability of permitted and buildable land, predictable energy, networking supplies, and servers, including graphics processing units and other components." That sentence is close to a list of the layers in this report.

Chip design, design software and licensed architectures

Nvidia's Rubin accelerator contains 336 billion transistors, according to Nvidia's technical description¹⁸. Nobody places 336 billion transistors by hand. A chip like this is designed in layers of abstraction, each handled by software, and the software comes almost entirely from two American companies.

How a modern accelerator is designed

The work starts with architects deciding what the chip is for: how many arithmetic units, what number formats they handle, how much memory sits beside them and how fast it must be read, and how the chip talks to its neighbours. Engineers then describe the logic in a hardware description language, a kind of programming language for circuits. That description is checked by simulation, and for chips this large by emulation as well, in which the design is loaded onto racks of special hardware that behave like the future chip so software can run on it months before silicon exists. Cadence's quarterly report²² lists "operating leases of hardware" among its recurring revenue for this reason: some of what it sells is emulation machines.

Next the logic is converted into a netlistA complete list of a chip's circuit building blocks and the wires between them. The netlist comes between a written description of the logic and the physical layout. of standard building blocks, and physical design software places those blocks and routes the wires between them on the specific manufacturing process the chip will use. Each foundryA company that manufactures chips designed by others. TSMC is the largest. Nvidia, for example, designs its chips and pays TSMC to make them. process has its own rules about how close wires may be and how transistors behave, published as a design kit. The software must model those rules exactly, which is why TSMC's annual report⁸ announces "EDA and intellectual property readiness" for each new nodeA generation of chip manufacturing process, named in nanometers (3-nm, 2-nm). The number is a label for the generation. Each new node fits more transistors into the same area. Section 6 explains it further. before customers can use it. Finally the design is checked for timing, power and manufacturability and sent to the foundry as a set of mask patterns, a step still called tape-out.

Electronic design automation (EDA) is the name for this software. Cadence and Synopsys supply most of it. Synopsys also sells pre-designed circuit blocks, which its quarterly report²³ describes as "pre-designed circuits that engineers use as components of larger chip designs," such as the interfaces that connect a processor to its memory. Synopsys has absorbed Ansys, a maker of physics simulation software, and Cadence has been buying simulation companies too. The reason is that at this scale the chip, its package, its heat and its power delivery have to be designed together.

Figure 3 shows how small this layer is in money. Synopsys and Cadence together reported about $4.1 billion of revenue in their latest quarters, against $96.2 billion at Nvidia. The layer matters because every leading-edge design must pass through tools certified for the foundry's process, and switching tools in the middle of a multi-year design is not a practical option. So Cadence and Synopsys earn a small amount from every leading-edge chip design.

Licensed architectures: what Arm sells

An instruction set architecture is the vocabulary of basic commands a processor understands. Software is written for a vocabulary, so a new processor that speaks an existing one can run existing software. Arm owns one of the two vocabularies used by most of the world's processors and licenses it. A company can license ready-made processor designs from Arm, or take an architecture licence and design its own processor that speaks Arm's vocabulary.

Arm is paid twice. Customers pay a licence fee up front, and then a royalty on each chip they ship. Arm's annual report for the year to March 2026²⁴ says royalties are "based on average selling price of the customer's Arm-based chip or a fixed fee per chip." Total revenue was $4.92 billion, of which royalties were $2.61 billion, up 21% on higher royalty rates per chip for its newest architecture, Armv9. The same report shows Arm moving down the chain: it lists "production silicon products, such as the Arm AGI CPU," chipletsSmall chips designed to be combined in one package so that together they work as a single larger chip. and complete subsystems among its developments. That puts Arm partly in competition with its own licensees, a tension the report flags as a risk.

Merchant chips and custom chips

A merchant accelerator is designed by a chip company and sold to anyone: Nvidia's and AMD's products are merchant chips. A custom accelerator is designed for one buyer's workloads, usually a hyperscaler or an AI lab, which owns the specification and often the software. The buyer rarely does the whole job alone. Broadcom and Marvell supply the hard parts that are not specific to the buyer, such as the high-speed input-output circuits, the memory interfaces and the physical design, and they manage the wafer, packaging and memory supply at TSMC and the memory makers.

That business has become large. Broadcom reported AI semiconductor revenue of $16.7 billion for the quarter to August 2026, up 221% from a year earlier, and guided to $21.7 billion for the next quarter, in its third-quarter release²⁵. Its quarterly report²⁶ describes "unprecedented demand for our custom AI accelerators or XPUs and AI networking solutions from our customers, including the leading frontier AI labs." Marvell's data-center business was 79% of its $2.74 billion of revenue in the quarter to August 2026, according to its 10-Q²⁷. The same filing shows how the customer relationship works: after the quarter, Marvell issued a customer a warrant to buy up to 59 million of its shares, vesting as the customer buys products, and it lists the risk that customers fail to pay for non-recurring engineering, the up-front design work.

The hyperscalers' own chips are now significant products in their own right. Google introduced two eighth-generation Tensor Processing Units (TPUs) in May 2026, one for training and one for inference, in its I/O keynote⁴. Alphabet's second-quarter release²⁸ now says Google Cloud "generates product revenues primarily from the sale of TPU systems," which means Google sells its chips to others as well as using them. Amazon said its chips business passed a $25 billion annual revenue run rate, with Anthropic and OpenAI "making multi-year, multi-gigawatt commitments" to its Trainium accelerators, in its second-quarter release².

A custom chip leaves out what one buyer does not need and avoids paying a merchant vendor's margin, which Section 17 shows is large. It gives up flexibility, because a chip optimised for today's model may suit next year's model less well, and it gives up the software and developer tools that come with a merchant chip. The buyer also takes on the supply chain: it must secure its own wafers, packaging and memory in the same queues as Nvidia.

How fast this layer can grow

Design capacity is not what limits the buildout. Design teams grow by hiring engineers and buying compute, and the leading vendors already ship a new generation roughly every year. Nvidia shipped mostly Blackwell Ultra systems in the quarter to July 2026, and its 10-Q²⁹ says the next architecture, Vera Rubin, "began production shipments" in the following quarter. A single design still takes two to three years from first architecture to volume; the annual cadence comes from overlapping teams.

Design matters to the rest of the chain in a different way: it sets the bill of materials for every layer below. Each new generation has used more memory per chip, more power per chip and a larger package. Rubin carries up to 288 gigabytes of the newest high-bandwidth memory, against 80 gigabytes on the H100, according to Nvidia's H100 specifications³⁰ and its Rubin description. A design decision in California becomes, two years later, a capacity requirement at memory fabs in Korea, packaging lines in Taiwan and substationsFenced yards of transformers and switches where power from high-voltage transmission lines is stepped down for local delivery, or where a power plant's output is stepped up onto the grid. in Ohio.

Chipmaking equipment and materials

ASML sold 44 low-NAThe standard version of ASML's extreme-ultraviolet machine. The newer High-NA version, explained later in this section, prints finer detail and costs more. extreme-ultraviolet lithography machines in 2025 for €10.4 billion, about €237 million each, and four of its newer High-NA machines for €1.16 billion, about €289 million each, according to its 2025 annual report on Form 20-F³¹. No other company sells either machine. Every leading-edge logic chip in an AI accelerator passes through them many times, and so does much of the newest DRAMDynamic random-access memory, the main working memory in computers and servers. High-bandwidth memory is built from stacks of DRAM chips. in its memory.

Lithography from first principles

A chip is built up in dozens of layers on a disc of silicon called a wafer. Each layer starts with a pattern. The wafer is coated with a light-sensitive film called photoresist, light is shone through a mask carrying the pattern, and the exposed areas of the film change chemically so they can be washed away. The pattern left behind tells the next machines where to etch material away or deposit new material. Repeat this for every layer and you have billions of transistors and the wiring that joins them.

The finest detail light can print depends on its wavelength: shorter waves draw finer lines. For two decades the industry's workhorse used deep-ultraviolet (DUV) light from argon fluoride lasers at 193 nanometers. Extreme ultraviolet (EUV) light has a wavelength of 13.5 nanometers, "more than 14 times shorter," according to ASML's explanation of its light sources³². Making it is the hard part. ASML fires molten tin droplets about 25 microns across from a generator at 70 meters a second, flattens each with a weak laser pulse, then vaporises it with a powerful one, creating a plasma that emits EUV light. It does this 50,000 times a second. The light is collected and steered by optics so precise that one supplier makes them: Carl Zeiss SMT, which ASML's annual report calls its "sole supplier of lenses, mirrors, illuminators, collectors and other critical optical components."

Each of these steps has one supplier because the machine is so hard to make. ASML's 20-F explains that for EUV "the complexity and uniqueness of parts often make multi-sourcing economically impractical," and that "the number of lithography systems we are able to produce is limited by the production capacity of one of our key suppliers, Carl Zeiss SMT." A rival would need its own light source, its own optics maker, and years of learning on customers' production lines. The four High-NA machines show how long that learning takes even for the incumbent. High-NA EUV (NA stands for numerical aperture, a measure of how wide a cone of light the optics capture) prints finer still. ASML shipped its first production-ready model in April 2025 and expects the platform "to start supporting high-volume manufacturing in 2027." In July 2026 Intel said it had qualified High-NA exposure on "select Intel 18A product layers," according to ASML's second-quarter investor presentation³³.

Deposition, etch, inspection: the rest of the fab

Lithography gets the attention, but most of a fab's tools do other things. Deposition machines lay down films of metal or insulator a few atoms thick. Etch machines carve the patterned material away, now often into deep, narrow trenches and around transistors whose channels are wrapped on all sides, a design called gate-all-around that the newest manufacturing nodes adopt. Cleaning tools remove residue between steps. Coater-developer tracks apply and develop the photoresist before and after each exposure. Inspection and metrology tools measure what was made and find defects.

Each of these has its own leaders. Lam Research, whose deposition and etch business had $23.2 billion of revenue in the year to June 2026, describes the coming "technology inflections" as 3D device scaling, multiple patterningPrinting one layer of a chip in two or more exposures, so the lines come out finer than a single pass of light could draw. The extra exposures add steps, time and cost. and advanced packaging in its annual report³⁴. Tokyo Electron reports a share above 90% in coater-developers, above 50% in dielectric etchEtching of the insulating layers that separate a chip's wires. A dielectric is an electrical insulator., and a "high share in bonders for HBM," the machines that stack memory chips, in its April 2026 results presentation³⁵. KLA, the inspection specialist, explains in its annual report³⁶ why its tools sell even when capacity is not growing: "accelerating the yield ramp and reaching high-volume production ahead of competitors are critical to manufacturers' revenue and profitability." Applied Materials reported that DRAM makers took 26% of its chip-equipment sales in the quarter to July 2026, up from 22% a year before, as memory makers invested in new generations, according to its 10-Q³⁷.

Under the tool makers sits another layer of suppliers: optics, lasers, vacuum pumps, precision stages, and materials such as silicon wafers, specialty gases and photoresists. TSMC's 20-F notes that its raw wafers come from "a limited number of suppliers located in Taiwan, Japan, Germany, and Singapore." These suppliers rarely limit the buildout on their own, but they are where a single factory fire or export rule can stop a tool from shipping.

Why equipment orders swing so hard

Chipmakers buy tools to add capacity, not to replace output. That makes tool demand depend on the change in chip demand rather than its level. When chip demand grows 20% a year, fabs must add about 20% capacity a year. A slowdown to 5% growth cuts the need for new tools to a quarter, even though chip sales are still rising. This amplification is why equipment sales fell 46% in 2009, from $29.5 billion to $15.9 billion, according to SEMI figures reported by Circuits Assembly³⁸.

The same mechanism now works in the other direction. In February 2026 ASML guided to 2026 sales of €34–39 billion; by July it had raised that to €43–45 billion, saying in its second-quarter release³⁹ that "our customers, in turn, continue to accelerate their capacity expansion plans." SEMI's mid-year forecast, as reported by PCD&F⁴⁰, puts 2026 wafer-fab equipment at $143.9 billion and total equipment sales at a record $165 billion. Tokyo Electron's April outlook was $150–170 billion a year of wafer-fab equipment in 2026 and 2027.

How fast this layer can grow

ASML's July release gives unusually concrete numbers. It plans to raise its capacity for low-NA EUV machines by 30% from "around 65" in 2026 for 2027, and is "investigating to increase capacity with another 30% for 2028," with the same 30% steps for DUV immersion machines from around 130. So the step that only one company can perform grows by about a third a year, and only if Zeiss keeps pace. That growth rate is fast for a machine this complex and slow next to demand for leading-edge wafers if AI spending keeps compounding at the rates in Section 1, so chipmakers place tool orders years ahead.

Export controls are the other limit on this layer, and they cut demand as well as supply. China was 34% of Lam's revenue in the year to June 2026 and 28% of Applied Materials' in its latest quarter. US rules issued in December 2024 added controls on chipmaking equipment and new "foreign direct productUS rules that apply American export controls to products made abroad if they were made with US technology, software or equipment." rules aimed at China's ability to make advanced-node chips, according to the Federal Register notice⁴¹. Section 16 returns to those rules.

Leading-edge foundry

TSMC's original plan for Arizona was a $65 billion investment in three fabs. In March 2025 it raised the total to $165 billion, adding three more fabs, two advanced-packaging plants and a research center, according to TSMC's announcement⁴². The first Arizona fab had been "in volume production since late 2024." The original $65 billion covered three fabs, so a leading-edge fab costs on the order of $20 billion, and years pass between the decision and the first wafers.

What a fab is, and what a node is

A fab is a building wrapped around a cleanroom, an area with almost no airborne particles, filled with the tools described in Section 5. A wafer moves through hundreds of processing steps there, cycling repeatedly through lithography, deposition, etch, cleaning and inspection. The building's own systems, supplying ultrapure water, specialty gases, vibration control and stable power, are a large part of the cost before a single tool is installed.

A process node is a generation of manufacturing recipe: a particular set of transistor designs, materials, wiring rules and tool settings. Nodes are named in nanometers, such as 3-nm or 2-nm, but the number is a generation label rather than a measurement of any single feature. What matters to a customer is what the node delivers: more transistors in the same area, less power for the same work, or more speed. A chip designed for one node cannot simply be made on another; it must be redesigned with the new design kit, which is why customers commit to a foundry's node years ahead.

Why it costs what it does

Three costs add up. The tools are expensive, and a leading-edge fab needs many EUV machines at €200 million or more each, plus hundreds of other tools. The building and its utilities are expensive. And the process itself has to be developed and debugged, which costs years of engineering before the factory earns anything. TSMC expects capital spending of $52–56 billion in 2026, "mainly for 2-nanometer and 3-nanometer nodes," according to its 2025 annual report⁸; in the second quarter of 2026 alone it spent NT$496 billion, about $15.7 billion at that quarter's exchange rate, according to its earnings presentation⁴³.

Yield learning

Yield is the share of chips on a wafer that work. A dust particle or a misaligned layer can kill a chip, and the bigger the chip, the more likely it is to contain at least one fatal defect. AI accelerators are among the largest chips made, so they are unusually sensitive to defect rates. A new node starts with low yields. Engineers trace each failure to a cause, fix the recipe, and the yield climbs over months. Every point of yield is chips that cost nothing extra to make, which is why KLA's annual report says reaching high-volume production "ahead of competitors" is "critical to manufacturers' revenue and profitability."

Intel's filings show what the early part of that curve costs. Intel began shipping its first products on its 18A node in high volume at the start of 2026. In the second quarter its product profit was $340 million lower "driven by an increased mix of higher-cost wafers manufactured on our Intel 18A process node," according to its 10-Q⁴⁴. The same filing notes $226 million of reserves released as early-ramp inventory sold through. That is yield learning seen in a profit and loss statement: the wafers cost more until the process matures.

Why so few suppliers, and so few customers

A foundry needs very large volumes to spread the cost of developing a node, and it needs customers who will commit to that node before it is proven. Customers, in turn, choose the foundry whose last node arrived on time with good yields, so the foundry that succeeded last time tends to win the next node too. TSMC's advanced nodes, 7-nm and below, produced 77% of its wafer revenue in the second quarter of 2026, according to its second-quarter release⁴⁵, and high-performance computing rose from 43% of its revenue in 2023 to 58% in 2025.

Intel's position shows the other side of the same loop. Its 10-Q says "substantially all of our Intel Foundry business supports internal manufacturing," with external foundry revenue of $293 million in the quarter, mostly from Altera, the programmable-chip business it no longer consolidates. Intel committed in the second quarter to finish developing its next node, 14A, but said the pace of its expansion "will ultimately be dictated by the amount of committed demand for Intel 14A that we are able to obtain." A foundry without large external customers cannot justify the capacity, and customers will not commit without seeing the capacity and yields. Samsung reported "2nm HPC engagements" and plans to ramp a second-generation 2-nm process for mobile chips in the second half of 2026, in its second-quarter results⁴⁶, but did not name AI accelerator customers.

Where it stands, and how fast it can grow

TSMC's business is growing faster than it planned. It expects 2026 revenue to rise "slightly above 40%" in US dollars, and its third-quarter guidance of $44.6–45.8 billion cites "the steep ramp-up of our 2-nanometer technology." Its August 2026 revenue was NT$514.8 billion, up 53% from a year earlier, according to its monthly report⁴⁷. Its next node, A16, was planned for mass production in 2026. At its April 2026 technology symposium TSMC moved that to 2027 and introduced A13 for 2029, according to a report of the symposium⁴⁸.

Taiwanese press reports summarised by TrendForce in September 2026⁴⁹ put TSMC's 2-nm capacity at about 90,000 wafers a month, rising to 110,000 by mid-2027, and its 3-nm capacity at more than 180,000, rising to 210,000. TSMC does not publish these figures, so they should be read as industry estimates. Taken together with the filed capex they describe a business adding capacity as fast as tools arrive and buildings can be fitted out, with each new fab two to four years from decision to output. Leading-edge wafer capacity is tight in 2026, but the filings describe it as a step that is growing roughly with demand rather than one that is holding the buildout back. Packaging, the next step, has a different history.

Advanced packaging and test

In September 2023 TSMC could package about 12,000 wafers a month with its CoWoS process and was trying to reach 25,000 to 30,000, according to a TechNode report⁵⁰ on the shortage. A month later TSMC said it planned to double that capacity by the end of 2025, according to the Taipei Times⁵¹. For most of 2023 and 2024 this step, not chip manufacturing, decided how many AI accelerators the world could ship.

Why the processor and its memory must sit side by side

Section 3 explained that AI work is limited by how fast data moves from memory to the processor. The answer has been to put the memory very close and connect it with a very large number of wires. A single stack of the newest high-bandwidth memory, HBM4, connects through a 2,048-bit interface: 2,048 separate data lines, each carrying up to 8 gigabits a second, according to the JEDEC HBM4 standard⁵². A leading accelerator carries several of these stacks. A conventional circuit board cannot route tens of thousands of such fine wires over even a few centimeters.

The solution is to mount the processor and the memory stacks on a slab of silicon, called an interposer, whose wiring is made with chipmaking tools and is therefore fine enough. TSMC's version is CoWoS, for chip-on-wafer-on-substrate. The name describes the steps: chips are bonded onto a wafer of interposers (chip-on-wafer), and each finished interposer assembly is then mounted on an ordinary package substrate (on-substrate) that connects it to the circuit board. Newer variants replace the full silicon slab with small silicon bridges embedded where the chips meet, which allows larger packages. Nvidia's Blackwell uses that variant, CoWoS-L, according to Epoch AI's cost analysis⁵³.

Why packaging became the limit

AI demand arrived faster than packaging lines could be expanded. The tools are different from those in a fab: bonders, precision placement machines, thinning and dicing equipment. Their makers had to scale too. When TSMC raised its equipment orders by about 30% in 2023, the new machines were not expected until the first half of 2024 and mass production not until the second half, according to the TechNode report. That one-year lag, while demand was doubling, caused the 2023–24 shortage.

Expansion since then has been large. Taiwanese reporting summarised by TrendForce in August 2026⁵⁴ puts TSMC's CoWoS capacity at about 70,000 wafers a month in 2025, rising to 130,000–140,000 by the end of 2026, with a "20% supply-demand gap" still expected in 2026. The same report says Nvidia alone reserved about 800,000–850,000 CoWoS wafers for 2026, more than half the total. A September report⁴⁹ puts the target at 260,000 a month by the end of 2028. TSMC does not publish these numbers. Read together with the 2023 figure, they describe roughly a tenfold expansion in three years that has narrowed the gap without closing it.

Epoch AI estimates that packaging accounts for about $1,100 of the roughly $6,250 it costs to make one B200, and that packaging yield loss accounts for another $1,000. When a package fails at the end of the line, the good processor dieA single chip cut from a finished wafer, before it is packaged. The plural is dies. and the good memory stacks bonded to it are lost with it. That is why the industry tests each die before assembly, and why testing has grown into a business of its own.

Who does it, and test

TSMC does most leading-edge packaging itself, a change from the industry's older division of labour in which outsourced assembly and test companies (OSATs) packaged chips that foundries made. ASE, a leading OSAT, acknowledges in its 2025 annual report⁵⁵ that "some foundry players have actively invested and expanded their advanced packaging capacity and gained some customer bases." Some of that work is now moving back to the OSATs: TSMC is reported to be outsourcing the chip-on-wafer step to ASE, SPIL and Amkor. Amkor expects to spend $2.5–3.0 billion in 2026, mainly on a new advanced-packaging plant in Arizona whose construction began in the second half of 2025, according to its 10-Q⁵⁶. TSMC's own Arizona plan includes two packaging plants, which would mean leading-edge chips no longer need to return to Taiwan to be packaged.

Test is the step that decides whether a chip ships. Each processor is tested on the wafer, again after packaging, and again in the system. As chips grow more complex the tests take longer, which raises demand for testers. Advantest, a leading maker of testers for complex chips, reported quarterly sales up 39% in the three months to June 2026, saying "demand for testers for AI-related high-performance semiconductors grew significantly," in its results review⁵⁷. SEMI's forecast puts test equipment sales up 31% in 2026, to about $15 billion. Tokyo Electron reports a leading position in probers, the machines that hold wafers during test, for leading-edge logic.

How fast this layer can grow

Packaging capacity can be added faster than fab capacity: roughly a year from tool order to production, against two to four years for a new fab. The shorter lead time is why packaging stopped being the single binding step once the industry had a year's warning. Packaging is still tight in 2026, and its next challenges are physical. Bigger packages with more memory stacks are harder to keep flat during heating and harder to cool, and the industry is developing versions built on large rectangular panels instead of round wafers, such as the CoPoS process ASE lists, to fit them. Section 22 returns to those limits.

Accelerators and the software around them

A central processing unit (CPU), the chip at the heart of an ordinary server, is built to do one thing after another very quickly, with lots of decisions along the way. A graphics processing unit (GPU) is built to do the same simple operation on a great many numbers at once. Multiplying two large tables of numbers, which is most of what a neural network does, is exactly that kind of work. A GPU devotes most of its area to arithmetic units and very little to the control logic a CPU needs, so for this one task it does far more work per second and per watt. Modern AI accelerators add special units for matrix multiplication and support ever smaller number formats: Nvidia quotes Rubin at 50 petaFLOPS for inference using a 4-bit floating-point format it calls NVFP4, in its technical description¹⁸. A petaFLOP is a thousand trillion operations. Using fewer bits per number lets more numbers flow through the same wires and arithmetic units.

What CUDA actually is

Nvidia introduced CUDA in 2006 to let programmers use its graphics chips for general computing, and its 2026 annual report¹³ describes CUDA as "the foundational NVIDIA CUDA development platform that runs on all NVIDIA GPUs, as well as hundreds of domain-specific software libraries." CUDA is several things at once: a programming model, a compiler, and a very large set of tuned libraries for the operations AI and scientific software use, plus the tools to find out why code runs slowly. AI frameworks and models are written on top of those libraries. When a lab moves to different hardware, it must make sure every one of those operations exists, is correct and is fast on the new chip.

So the software a buyer already uses shapes which hardware it buys. A buyer with a small team picks the platform its software already runs well on. A buyer with a large team, such as Google with its TPUs or Amazon with Trainium, can afford to build its own software stack and will do so if the hardware saves enough. AMD is closing the gap from the other direction: its second-quarter 2026 release⁵⁸ announced ROCm.aiROCm is AMD's counterpart to CUDA: the software platform programmers use to run code on AMD's accelerators., "an AI-native developer experience," and a "multiyear collaboration to optimize AMD Instinct GPUs and ROCm software development" with Anthropic. Large buyers are now paying engineers to narrow the software gap.

The accelerators shipping now

Nvidia remains the largest supplier by a wide margin. Its revenue for the year to January 2026 was $215.9 billion, and in the quarter to July 2026 it reported $96.2 billion, with data-center revenue of $89.0 billion. Blackwell systems were "the majority of our system shipments," and Vera Rubin "began production shipments" in the quarter that followed, according to its 10-Q²⁹. The same filing says Nvidia is "currently experiencing certain supply constraints."

Power per chip has doubled since the H100. The H100 is rated at up to 700 watts on Nvidia's product page³⁰. Nvidia's power-management documentation sets a maximum of 1,200 watts for each GB200 accelerator, and a per-node ceiling for the newer GB300 that works out to 1,400 watts each, according to its device specifications⁵⁹ and September 2026 power guide⁶⁰. Server makers' planning guides put Rubin at roughly 1.8 to 2.3 kilowatts per accelerator; Nvidia has not published an official figure, so that range comes from a server maker's guide⁶¹ and should be treated as an estimate.

AMD's data-center revenue more than doubled to $6.7 billion in the second quarter of 2026. It launched its Instinct MI400 family and its Helios rack, which it says is being deployed by "Anthropic, Cirrascale, HUMAIN, Meta, Microsoft, OpenAI, Oracle," among others, and it announced a partnership with Anthropic "to deploy up to 2 gigawatts of MI450 Series GPUs." The custom chips described in Section 4, from Google, Amazon and Broadcom's customers, are the other alternative. Figure 7 sets out the trade-off between the two kinds of supply.

Why the product is now a rack

The clearest example is the GB200 NVL72, a single rack holding 72 Blackwell accelerators and 36 Grace CPUs, joined by a switched network called NVLink with 130 terabytes a second of total bandwidth, 13.4 terabytes of high-bandwidth memory, and liquid cooling, according to Nvidia's product page⁶². Its successor, the Vera Rubin NVL72, raises the memory to 20.7 terabytes and the NVLink bandwidth to 216 terabytes a second, according to the Vera Rubin page²⁰. Within the rack, each accelerator can exchange data with any other at about 1.8 terabytes a second (130 divided by 72), roughly a quarter of the 8 terabytes a second at which it reads its own memory (576 divided by 72). Nvidia designs the rack so that software can treat the 72 chips much like one very large accelerator.

Nvidia sells it that way. Its annual report says its data-center platform is "typically delivered to customers as rack-scale systems, subsystems, or modules," and describes "extreme co-design where the infrastructure's chips, networking, systems, software, and algorithms are holistically architected." AMD's Helios is the same idea. This changes the chain in three ways. First, the rack's design is fixed by the chip vendor, and the companies that assemble it, covered in Section 10, build to that design. Second, the rack's power and cooling requirements arrive at the data center as a package: a building that cannot deliver about 120–135 kilowatts of power and liquid cooling to one rack cannot host the product at all. Third, more of the value of the system is captured at the design stage, because the buyer is paying for a working machine rather than parts.

The rack also changes what "supply" means. A shortage of any one component, whether a memory stack, a cable, a power shelf or a cold plate, holds back the whole rack, and with it 72 accelerators. Nvidia's response has been to commit very large sums to its suppliers years ahead. Its supply and capacity commitments rose from $119 billion to $279 billion in a single quarter to July 2026, and its 10-Q says they are "primarily memory and manufacturing facilities." That figure is the best single public measure of how tight the layers below the accelerator have become.

Memory and storage

In the three months to early September 2026 Micron Technology, the US memory maker, reported revenue of $54.2 billion and a gross margin of 86.8%. A year earlier the same quarter had brought in $11.3 billion at a 44.7% margin, according to its fourth-quarter release¹⁹. Three years earlier, in fiscal 2023, Micron had sold its products below cost. Those swings are the most important fact about the memory layer, and they explain why memory, more than any other chip, is short in 2026.

What high-bandwidth memory is, physically

DRAM, dynamic random-access memory, stores each bit as a tiny electric charge in a capacitor next to a transistor. The charge leaks, so the chip must refresh it constantly, and reading it is fast but not free. Ordinary DRAM chips are packaged singly and placed on memory modules a few centimeters from the processor. High-bandwidth memory (HBM) uses DRAM dies of the same basic kind but packages them very differently. Twelve or sixteen dies are thinned, stacked on top of each other on a base die, and connected vertically by through-silicon vias: holes drilled through each die and filled with metal, so that signals pass straight up and down the stack. The whole stack then sits beside the processor on the interposer described in Section 7.

The current JEDEC standard, HBM4, gives each stack a 2,048-bit interface running at up to 8 gigabits a second per line, for up to 2 terabytes a second, and allows stacks of 4, 8, 12 or 16 dies holding up to 64 gigabytes, according to the JEDEC announcement⁵² as republished. JEDEC is the industry body that sets memory standards, so that a processor designer can buy compatible memory from several makers.

Stacking is hard for physical reasons. To fit twelve dies into the same height as eight, SK hynix had to make each die 40% thinner, according to its September 2024 announcement⁶³ of the first 12-layer HBM3E. Thin silicon warps when heated, and the bonding process heats it. SK hynix developed its own process, which it calls Advanced MR-MUF, to fill the gaps between dies with a liquid that hardens and to control the warping. Every die in the stack must work, so a single bad die ruins the stack, and the stack's yield is roughly the product of the yields of each die and each bonding step. Heat is the other problem: the stack sits next to the hottest chip in the data center, and SK hynix markets each generation's heat dissipation as a feature.

The result is memory that costs much more per bit than ordinary DRAM. Because each stack carries an extra base die and loses all its dies if any one fails, it also takes more wafer capacity per usable bit. In an accelerator it is also the largest single cost. Epoch AI estimates that the memory in one B200 costs about $2,900, more than three times the $900 processor die, as Figure 8 shows.

How memory cycles work

DRAM is a commodity in the economist's sense: chips from different makers that meet the same standard are close substitutes, and price is set by supply and demand across the market. A memory fab costs many billions of dollars and takes two to three years to build, and once built it is cheapest to run full. So makers add capacity when prices are high, the new capacity arrives together a few years later, prices collapse, and makers cut spending until shortage returns. Figure 9 shows the most recent cycle at Micron. In fiscal 2023 its revenue fell 49% to $15.5 billion and its gross margin was −9%, according to its 2023 annual report⁶⁴, which blamed "declines in average selling prices for DRAM." In fiscal 2026 its revenue was $133.2 billion and its gross margin 81%.

Three things turned the slump into the present shortage. Makers cut capital spending in 2023, so less new capacity was under construction when demand returned. Shifting wafers to HBM, which uses more wafer per bit, reduced the supply of ordinary DRAM. And AI inference raised demand for ordinary server memory and solid-state drives as well as for HBM, because of the long-context caches described in Section 3. Micron's 2023 annual report already noted that HBM "had very strong demand this year, driven by demand for generative AI."

Where it stands now

All three makers describe a market where demand exceeds supply. SK hynix reported second-quarter 2026 revenue of 79.3 trillion won, up 257% from a year earlier, with a 76% operating margin, and said it is operating "in a market environment where customer demand exceeds supply capabilities," in its results release⁶⁵. It began mass shipments of HBM4 in that quarter and shipped samples of the next generation, HBM4E. Samsung reported a record operating profit of 89.5 trillion won, almost all of it from its chip division, and said its outlook is for the market to remain "undersupplied" with "supply constraints... expected to continue," according to its results release⁴⁶, which also says it scaled up sales of HBM4 in the quarter and shipped HBM4E samples to major customers. Micron reported in June that its HBM4 was "in high-volume shipments for our lead customer's platform" and that HBM4E would reach volume production in calendar 2027, in its third-quarter release⁶⁶.

The buyers' filings tell the same story. Nvidia's supply commitments, "primarily memory and manufacturing facilities," rose to $279 billion. Nvidia also said its PC chip sales were "tempered by elevated memory and systems prices," a sign that the AI build is pulling memory away from other markets. Amazon raised its 2026 capital spending plan by $20 billion, citing higher memory prices, according to reporting of its July call.

The contracts are changing as a result. Micron's May 2026 quarterly report⁶⁷ describes new strategic customer agreements "structured as take-or-pay agreements, with binding commitments for specific volumes over the multi-year contract terms," with pricing "either fixed, or... subject to minimum and maximum pricing." Take-or-pay means the buyer pays for the agreed volume whether or not it takes delivery. SK hynix says it has signed long-term agreements with about ten customers. These agreements are an attempt to break the cycle: the buyer gets assured supply, and the maker gets assured revenue to justify building fabs. Whether they hold in a downturn is untested.

How fast this layer can grow

New memory capacity takes two to three years. Micron broke ground in Singapore in January 2025 on an HBM packaging plant that will "meaningfully expand our total advanced packaging capacity beginning in the first half of calendar 2027." It broke ground on a new NANDThe flash-memory chips used in solid-state drives and phones. NAND keeps data when the power is off but is slower than DRAM, so NAND holds stored data and DRAM feeds the processor. fab in Singapore in January 2026 that will add cleanroom space "in the second half of calendar 2028." To move faster it bought an existing fab in Taiwan from Powerchip for $1.8 billion in March 2026, which it expects to ship meaningful volume from "mid-calendar 2027." SK hynix says its Yongin site will open its first cleanroom in early 2027. The faster routes, such as buying an existing fab, are limited by how many suitable fabs exist. Equipment makers see the same timing: Applied Materials' DRAM sales were 26% of its chip-equipment revenue in its latest quarter, up from 22% a year earlier.

Storage

AI also needs places to keep data that is not being worked on this instant: training datasets, model checkpoints saved during training, and the caches and logs of inference. Flash memory (NAND) in solid-state drives serves the fast tier. Hard disk drives still hold most of the bulk. Seagate's 2026 annual report⁶⁸ says hard drives support "the vast majority of exabytes shipped into large data center deployments," and its revenue rose 34% to $12.2 billion in the year to July 2026 on higher shipments of high-capacity drives and price increases. Its newest drives use heat-assisted magnetic recording, in which a tiny laser heats each spot on the disk as it is written so that bits can be packed more densely. Storage is growing with the buildout, but its makers do not describe a shortage on the scale of DRAM, and it is not what holds a data center back.

Western Digital, the other large maker of hard drives, shows the same pattern in its annual report for the year to July 2026⁶⁹. Revenue rose 36%, from 25% more exabytes sold (an exabyte is a billion gigabytes) at prices 8% higher per exabyte. Cloud customers were 89% of its revenue. An 8% price rise is modest next to what the memory makers charged in the same year, and that gap is the difference between a growing market and a short one.

Servers, racks and the companies that assemble them

Foxconn (Hon Hai), the Taiwanese contract manufacturer, reported a gross margin of 6.15% and an operating margin of 3.67% for the first half of 2026, while saying its shipments of AI server racks would "more than double" for the year, according to its second-quarter results as reported by I-Connect007⁷⁰. Nvidia's gross margin over the same months was about 75%. That gap sums up this layer: assembling AI systems is a very large business, and the assemblers keep very little of the money that passes through them.

From chip to hall

An AI cluster is built in nested levels, and each level multiplies the power it must deliver and the heat it must remove. Figure 10 follows Nvidia's GB200 NVL72 from the accelerator package to the substation. The package is fitted to a board with a cold plate, a metal block with liquid channels pressed against the chip. Boards go into a compute tray, a drawer one rack unitThe standard unit of height for equipment in a server rack, written 1U. A 1U tray is the thinnest standard drawer. high. Nvidia's DGX GB200 user guide⁷¹ describes the rack as "18x 1RU compute trays, each with 2 Grace CPUs and 4 Blackwell GPUs," plus switch trays and "eight power shelvesUnits in the rack that convert the building's alternating-current supply into the direct current the trays run on. Section 13 follows the whole path." each delivering 33 kilowatts, for a rack that consumes "approximately 120kW." Racks are arranged in rows in a hall, and the hall is fed from the campus electrical yard and the utility substation.

The weight and density are new. A single DGX H100 server, eight accelerators in an eight-unit chassis, weighs up to 130 kilograms and draws up to 10.2 kilowatts, according to Nvidia's DGX H100 user guide⁷². Nvidia's H100 data-center design guide⁷³ says those systems are "typically deployed with a rack density of four DGX H100 systems per rack," which is about 41 kilowatts and over half a tonne of servers alone. The GB200 rack packs 72 accelerators and about three times the power into the same footprint. Floors and cooling designed for the older density often do not carry over.

Power per rack across generations

Figure 11 plots power per rack for four generations, and separates the accelerators from everything else in the rack. The whole rack went from about 41 kilowatts for an H100 rack to about 135 kilowatts for a GB300 NVL72, based on Nvidia's own planning envelopes. For Rubin, Nvidia has not published a rack figure; one server maker's guide puts it at 190–230 kilowatts, and Nvidia's 100-megawatt reference design for 40,000 Rubin accelerators implies about 180 kilowatts of facility power per 72 accelerators. The gap between the two lines is the CPUs, the NVLink switches, the network cards, the power conversion and the pumps. Nvidia's May 2025 power architecture post⁷⁴ says the industry is moving to racks of "1 MW" starting in 2027.

Who designs and who assembles

Three kinds of company build these systems. Original design manufacturers (ODMs), such as Foxconn, Quanta and Wistron, build to a customer's or Nvidia's specification in very large volumes, often for hyperscalers that design their own racks. Original equipment manufacturers (OEMs), such as Dell and Super Micro, sell branded systems with their own engineering, support and financing, mostly to enterprises, neoclouds and governments. And the chip vendors themselves now design the reference rack, which the others build.

The OEM numbers show how fast this layer is growing. Dell sold $16.4 billion of AI-optimized servers in the quarter to July 2026, double the year before, and reported "a significant increase in backlog," according to its 10-Q⁷⁵. Super Micro's sales rose to $39.1 billion in the year to June 2026 from $22.0 billion, according to its annual report⁷⁶, and it now sells what it calls Data Center Building Block Solutions, from single accelerators "to complete racks, site infrastructure, management software, and professional services."

How money and inventory move

Most of the cost of an AI server is parts the assembler buys from Nvidia and the memory makers and passes through at little markup. The assembler's job is to integrate, test and ship. Its profit comes from that work and from managing the parts in between, and its main risk is holding expensive inventory when demand or a product generation changes. Super Micro's gross margin was 10.8% in fiscal 2026, down from 13.8% two years earlier, on the figures in its annual report. Its filing argues that its building-block designs mean "our inventory can generally be used with multiple products, lowering working capital requirements and reducing the risk of inventory write-downs."

Nvidia's own reporting shows how indirect this chain is. Its annual report¹³ says "many of our indirect customers often do not purchase directly from us but through multiple OEMs, ODMs, system integrators, distributors, and other channel partners," and that some indirect customers each represent 10% or more of its revenue. The hyperscaler pays the assembler; the assembler pays Nvidia; Nvidia pays TSMC and the memory makers. Each payment lags the one before it, and the assembler sits in the middle with the thinnest margin and a balance sheet full of other companies' chips.

How fast this layer can grow

Assembly is not a binding step. Lines can be added in months rather than years, and Foxconn expects to more than double its AI rack shipments in 2026. The constraint shows up as a shortage of any single input: a rack missing one memory stack, one cable assembly or one power shelf cannot ship. The rack's growing power and weight also push work out of the assembler's plant and into the data center, where liquid cooling loops, heavier floors and new power distribution have to be ready before the rack arrives. Those are the next three layers.

Networking and optics

Training a frontier model means tens of thousands of accelerators working on one problem, and after every step each of them must share results with the others. The network that carries those results decides how much of the cluster's arithmetic is actually used. A cluster whose network is too slow spends its time waiting. That is why networking, once a modest line in a data-center budget, has become one of the fastest-growing parts of it.

Two networks, not one

AI clusters use two different kinds of network. The scale-up network joins a small group of accelerators so tightly that they can read and write each other's memory directly, as if they were one large chip. Nvidia's NVLink is the leading example: in a GB200 NVL72 rack it joins 72 accelerators with 130 terabytes a second of total bandwidth. Scale-up links must be very fast and have very low delay, so they run over short distances, inside a rack or a few neighbouring racks.

The scale-out network joins those groups to each other across the hall and sometimes across buildings. It carries less traffic per chip but over much longer distances and to many more endpoints. It has been built on two technologies: InfiniBand, a high-performance network Nvidia sells under its Quantum brand, and Ethernet, the standard that runs most of the world's data centers. A third network, the front end, connects the cluster to storage and to the users and applications that send it work.

The difference matters for the supply chain because the two networks use different parts and have different owners. The scale-up network is mostly proprietary today and ships inside the vendor's rack. The scale-out network is a market of switches, network cards, cables and optics from many suppliers. Arista, which sells switches for large AI and cloud data centers, reported revenue of $3.04 billion in the quarter to June 2026, up 38% from a year earlier, in its 10-Q⁷⁷.

Copper or light

Every link has to move electrical signals from one chip to another. Over short distances that is done with copper. Copper is cheap and uses little power, but at today's speeds a signal fades and distorts within a short distance. Two families of parts stretch copper further. Retimers are chips that receive a degraded signal and transmit a clean copy; Astera Labs sells them for the PCIePCI Express, the standard high-speed connection between a server's processor and the other chips and cards inside it. links inside servers, and its 10-Q⁷⁸ describes its products as addressing "critical data, network, and memory bottlenecks." Active electrical cables put similar chips inside the cable itself. Credo's latest 10-Q⁷⁹ attributes more than 90% of its revenue growth to higher shipments of these cables at hyperscale customers.

Beyond a few meters the signal must become light. An optical transceiver contains a laser, a modulator that turns the laser on and off billions of times a second, a photodetector at the other end, and a digital signal processor that cleans up the electrical signal on the way in and out. Lumentum, which makes lasers and optical components, reported revenue up 83% to $3.0 billion in the year to June 2026, according to its annual report⁸⁰. The same report recalls how the optics business behaved in the last cycle: customers "built up inventory in response to supply shortages and then brought down inventories as supply chain constraints eased," and revenue swung accordingly.

Pluggable opticsOptical transceivers built as modules that slot into ports on the front of a switch, so a technician can swap one by hand. use a lot of power. Nvidia says a pluggable transceiver often uses "30W per interface" and that moving the optical engine into the switch package, a design called co-packaged optics, can bring that to "as low as 9W," in an August 2025 technical post⁸¹. The saving comes from shortening the electrical path. In a pluggable design the signal travels across the switch's circuit board to the front panel and loses about 22 decibelsA logarithmic unit for signal strength. Each 10 decibels is a tenfold change in power, so a loss of 22 decibels leaves the signal far weaker than a loss of 4. of strength on the way; beside the switch chip it loses about 4, so the power-hungry signal processor is no longer needed. Nvidia announced co-packaged switches in March 2025, with InfiniBand versions due later that year and Ethernet versions "in 2026," naming TSMC, Coherent, Corning, Fabrinet, Foxconn and Lumentum among its partners, according to its announcement⁸².

The standards

Most of the open competition in this layer is organised through standards bodies, and three matter most. Figure 12 summarises where each stood in 2026.

UALink was ratified in April 2025 by a consortium of "more than 85 member companies," according to its release⁸³, as an open alternative to proprietary scale-up links. The Ultra Ethernet Consortium released its first specification in June 2025, describing it as "open, interoperable standards that prevent vendor lock-in," according to the Linux Foundation's release⁸⁴. And the IEEEThe Institute of Electrical and Electronics Engineers, a professional body whose committees write the Ethernet standards.'s next Ethernet standard, which doubles the speed per electrical lane to 200 gigabits a second, passed its first formal ballot in March 2026, according to the IEEE 802.3 working group⁸⁵.

Both open standards have moved since their first versions. On 7 April 2026 the UALink Consortium ratified a second set of specifications, adding computation inside the network, a definition for building UALink into chiplets, and a separate 200-gigabit physical-layer specification that can be updated without reopening the rest, according to its release⁸⁶. The Ultra Ethernet Consortium published revision 1.0.3 of its specification on 16 July 2026, according to its specification page⁸⁷. Optics has its own standards body. The Optical Internetworking Forum (OIF), which writes interoperability agreements for optical and electrical interfaces, published the first co-packaging standard in 2023: a 3.2-terabit module for switches of 51.2 terabits a second, according to its announcement⁸⁸. An agreement of this kind is what lets a switch maker buy co-packaged modules from more than one supplier.

What happens when an open standard arrives

These documents suggest what to expect. A proprietary technology comes first, because one company can move faster than a committee. Once it is valuable enough, the largest buyers fund an open alternative so they can buy from several suppliers. Nvidia's own product line shows the pressure: it now sells an Ethernet product family, Spectrum-X, alongside InfiniBand. When an open standard is good enough, competition shifts from the protocolThe agreed rules for how data is formatted, addressed and checked as it moves over a link. Two devices that follow the same protocol can talk to each other, whoever made them. to the parts that remain hard to make, such as switch chips, lasers and signal-processing chips, and prices per bit fall. The protocol stops being a source of profit, and the components become one. That reading is an inference from how the scale-out market is now organised. It is not yet evidence that the same will happen to scale-up links, where open products are only beginning to ship.

How fast this layer can grow

Networking parts can usually be added within one to two years. The tighter points come during each move to a faster generation, such as from 800 gigabits to 1.6 terabits a port, when new lasers and signal-processing chips must be qualified by each buyer. Lumentum's account of its own inventory swings shows how quickly a shortage of such parts can turn into a surplus. None of this has been the main limit on the buildout. Networking's bigger effect on the chain is on power and space: optics and switches take a growing share of each rack's power budget, which is why co-packaged optics matter for Section 13 as much as for this one.

The suppliers' own numbers show the growth without a shortage. Cisco took $9.3 billion of orders for AI infrastructure from hyperscalers in its fiscal year to July 2026, $4 billion of them in the final quarter, according to its fourth-quarter release⁸⁹. Coherent, which makes optical transceivers and lasers, reported revenue of $2.05 billion in the quarter to June 2026, up 34%. Its finance chief said the company was prioritising investment "to expand manufacturing capacity so we can efficiently fulfill the ongoing acceleration in customer demand," according to its results release⁹⁰.

Cooling

Almost all the electricity a computer uses ends up as heat. A GB300 rack drawing about 135 kilowatts is, from the building's point of view, a 135-kilowatt heater the size of a wardrobe, running all day. Removing that heat is the cooling layer's job, and the move from air to liquid is reshaping how data centers are designed.

Why air ran out

Air carries heat poorly. Moving a given amount of heat with air requires either a very large flow or a large rise in the air's temperature, and both have limits: fans use power and make noise, and chips must stay below a maximum temperature. Water carries about 4,000 times more heat per unit of volume than air and conducts heat about 30 times better, according to Google's description of its cooling work in an April 2025 post⁹¹. That gap is why AI racks have moved to liquid.

The practical ceiling for air shows up in product specifications. Schneider Electric's rear-door heat exchanger, a radiator fitted to the back of a rack that cools the air leaving it with chilled water, supports "up to 75 kW per rack," according to its September 2025 announcement⁹². An H100 rack at about 41 kilowatts fits under that ceiling. A GB200 rack at about 120 kilowatts does not, and Nvidia's GB200 user guide describes liquid flowing "through the cold plates that are attached to the CPUs and the GPUs," with fans still blowing air over the other components.

How liquid cooling works

Figure 13 follows the heat from the chip to the outside air. A cold plate, a metal block with fine internal channels, is pressed against each processor. Coolant flows through it and out to manifolds running up the rack. From there it goes to a coolant distribution unit (CDU), a cabinet with pumps and a heat exchanger. The CDU keeps the rack's coolant in its own closed loop, separate from the building's water, and passes the heat across to the facility loop. The facility loop carries it to the roof or yard, where dry coolersLarge radiators with fans that cool liquid with outside air, without evaporating any water., chillersRefrigeration machines that cool water mechanically. They work in any weather but use electricity to do it. or cooling towers release it to the air.

Google has run this kind of system at scale for longer than most. Its April 2025 post says it has cooled "more than 2,000 TPU Pods" with liquid "at GigaWatt scale" over seven years with fleet-wide uptime of about 99.999%, and it contributed the design of its fifth-generation CDU, called Project Deschutes, to the Open Compute Project, the industry group that publishes open hardware designs. Super Micro says its second-generation direct liquid cooling captures up to 98% of a rack's heat in liquid, in its annual report. Nvidia specifies a 45°C coolant inlet temperature for the Vera Rubin rack on its product page²⁰. Warm coolant matters because the warmer the water leaving the racks, the more often outside air alone can cool it, without running chillers.

The water question

Cooling towers cool water by evaporating part of it, which is efficient in electricity but consumes water. That has made data centers' water use a local political issue. Microsoft reported a fleet-wide water usage effectiveness of 0.30 liters per kilowatt-hour in its latest fiscal year, down from 0.49 in 2021, in a December 2024 post⁹³. Water usage effectiveness is liters of water consumed per kilowatt-hour of IT energy. The same post says its new designs use chip-level closed-loop cooling with no evaporation, avoiding "more than 125 million liters of water per year per datacenter," with pilot sites in Phoenix and Mount Pleasant, Wisconsin, due to become operational in late 2027.

The trade-off is electricity. Without evaporation the plant relies more on mechanical cooling, which raises power usage effectiveness (PUE), the ratio of total facility energy to the energy used by the IT equipment. Microsoft says the increase is "nominal" because newer chips tolerate warmer coolant. In water-scarce places the trade is worth making; in places where power is the scarce input, the answer may be the reverse. That choice is now part of site selection.

Who supplies it, and how fast it can grow

The main suppliers are the data-center power and cooling specialists. Vertiv reported second-quarter 2026 sales of $3.27 billion, up 24%, and is spending $550–570 million on capital projects in 2026 "to support capacity expansion," according to its 10-Q⁹⁴; it has bought companies in "liquid cooling technologies, heat-exchange and heat rejection technologies." Schneider Electric took a controlling stake in Motivair, a liquid-cooling maker, in February 2025. Server makers sell their own cooling systems as part of the rack.

Cooling equipment can be added in months to a year, and it is not a binding step at the level of the whole buildout. Its constraint is closer to the site. A liquid-cooled hall needs pipework, pumps and heat rejection designed in from the start, and an air-cooled building built three years ago may not be convertible at reasonable cost. The installers who can do that work safely are part of the labour constraint in Section 16. Reliability is the other risk: a leak in a rack loop can damage very expensive equipment, which is why the industry separates the rack loop from building water and why fittings and coolant chemistry receive so much attention.

The electrical path inside the building

Power arrives at a large data center at transmission voltage, tens or hundreds of thousands of volts, and the chips inside it run at less than one volt. Between the two sits a chain of transformers, switches, batteries and converters. Each step changes the voltage or protects against faults, each step loses a little energy as heat, and several of the steps rely on equipment that takes two to four years to arrive.

From substation to chip

Figure 14 shows the path in its usual form. At the substation, large power transformers step transmission voltage down to medium voltage, typically in the tens of thousands of volts. Medium-voltage switchgear, cabinets of heavy circuit breakers, routes that power around the site and cuts off any section that develops a fault. Uninterruptible power supplies (UPS), banks of batteries and power electronics, carry the load for the seconds or minutes it takes backup generators to start if the grid fails. Further transformers step the power down to the few hundred volts used inside the hall, and busway, overhead bars of copper or aluminium, carries it to the rows. In the rack, power shelves convert alternating currentElectricity that reverses direction many times a second. The grid delivers power this way. Chips run on direct current, which flows one way, so the power has to be converted before it reaches them. to direct current at about 54 volts, and converters on each board bring it down to the level the chip uses.

Where energy is lost, and why voltages are rising

Power equals voltage times current. For the same power, a higher voltage means a smaller current, and the energy lost as heat in a conductor rises with the square of the current. So every time power is carried at low voltage, more copper is needed to keep losses down, and every conversion between voltages or between alternating and direct current wastes a few percent. In a rack drawing 120 kilowatts at 54 volts the current is over 2,000 amperes, which takes thick copper bars to carry.

Nvidia's May 2025 post⁷⁴ sets out the problem and its proposed fix. It says 54-volt distribution runs into limits "as racks exceed 200 kilowatts," that a megawatt rack would need "up to 200 kg of copper busbar," and that power shelves would take up so much of the rack that there would be little room left for computers. Its answer is to convert 13.8-kilovolt utility power directly to 800 volts direct current at the edge of the data center and carry that to the racks, which it says cuts copper use by 45% compared with conventional 415-volt alternating current and improves end-to-end efficiency by up to 5%. Its named partners include Vertiv, Eaton, Schneider Electric and Delta. Google, Meta and Microsoft are working on a related design, a ±400-volt direct-current "sidecar" rack that holds the power equipment beside the computing rack, through the Open Compute Project; Google says it improves efficiency by about 3% and borrows components from the electric-vehicle supply chain, according to its April 2025 post⁹¹. ABB has gone further up the chain with a 34.5-kilovolt UPS that lets data centers "connect directly to the grid without voltage conversion," according to its second-quarter 2026 release⁹⁵.

Where the long lead times come from

The waits are longest for the biggest equipment. Wood Mackenzie, an energy research firm, reported in April 2024⁹⁶ that transformer lead times had grown from about 50 weeks in 2021 to an average of 120 weeks, with large substation and generator transformers taking 80 to 210 weeks. In August 2025⁹⁷ it estimated a 2025 US supply deficit of 30% for power transformers and 10% for distribution transformersThe smaller transformers that step power down for local lines and individual buildings. Pad-mounted units, set at ground level on a concrete pad, are one common kind. Power transformers are the large units at substations and power plants., with imports supplying 80% of US power transformers. Cushman & Wakefield's 2026 cost guide, as summarised by CRE Daily⁹⁸, puts pad-mounted transformers at 68 to 113 weeks and generators at 60 to 100 weeks, and says switchgear prices rose 60% between December 2021 and early 2026.

Wood Mackenzie's survey for the second quarter of 2025 put the average wait at 128 weeks for a power transformer and 144 weeks for a generator step-up transformer, the unit that connects a power plant to the transmission grid, according to POWER magazine⁹⁹. The same survey put switchgear at 44 weeks, and Cushman & Wakefield puts medium-voltage switchgear at 38 to 63 weeks. So switchgear is a wait of about a year, and the large transformer is a wait of two and a half to three. The new US transformer plants listed in that POWER report arrive from 2027: Siemens Energy's $150 million plant in Charlotte, North Carolina, early that year, and Eaton's $340 million plant in South Carolina during it.

The reasons are partly physical and partly financial. A large transformer is a custom product built around a core of grain-oriented electrical steel, a specialised steel with magnetic properties tuned for the job, and Wood Mackenzie notes that North America has a single domestic source¹⁰⁰ of it and that its price "doubled since January 2020." Copper windings, insulation, assembly and testing take skilled labour that is also scarce. The financial reason is the one that explains why the shortage has lasted. Wood Mackenzie quotes the industry's view that "it can take decades for transformer manufacturers to break even" on a new plant, so makers are reluctant to build for a demand surge they fear may not last. When the plants do come, they take years. Hitachi Energy's new $1 billion plant in South Boston, Virginia, is due to be completed in 2028, according to pv magazine USA¹⁰¹, which reported in May 2026 that lead times for high-capacity units had stretched to as much as four years.

Who supplies it, and where it stands

The electrical equipment makers are booking orders faster than they can ship them. ABB's electrification orders rose 60% to a record $7.2 billion in the second quarter of 2026, with data-center orders showing "exceptional triple-digit growth" and "no indications of pre-ordering." Siemens Energy's grid business reported a backlog of €51 billion and "substantial growth in the transformer business including data center projects," in its third-quarter results¹⁰². GE Vernova said its electrification data-center orders had passed $5 billion in the first half of 2026, more than double its 2025 total, in its second-quarter release¹⁰³. Eaton bought Fibrebond, a maker of pre-built power enclosures for data centers, for $1.43 billion and a developer of solid-state transformersTransformers that change voltage with power electronics, the kind of circuitry found in chargers and inverters, instead of a large iron core and copper windings., according to its 10-Q¹⁰⁴.

One way the industry saves time is prefabrication: building power rooms and skidsSteel frames on which equipment is mounted, wired and tested in a factory, so it ships and installs as one unit. in a factory while the site is prepared, then shipping them complete. It shortens construction but not the wait for the transformer inside the enclosure. The electrical path is therefore one of the two places in the chain where money spent today buys capacity only years later. The other is the grid itself.

Generation, the grid and connection

On 14 July 2026 PJM Interconnection, which runs the grid for 67 million people in thirteen states and Washington, D.C., announced the results of its auction for power-plant capacity in the year starting June 2028. The price hit the cap of $325 per megawatt-day. The capacity bought was 6,831 megawatts short of what PJM's reliability standard requires. The auction "cleared 525 MW UCAP of new generation and generation uprates," and the forecast peak demand it was buying for had risen by about 2,000 megawatts since the previous auction, with "the continued trend of the addition of large data center loads," according to PJM's release¹⁰⁵. It was the second auction in a row in which the whole region fell short, "the first in PJM history."

What those terms mean

A capacity market pays power plants to be available when demand peaks, separately from the payment for the energy they actually produce. PJM runs it as an auction three years ahead, so that new plants have time to be built. A megawatt-day is one megawatt of capacity held available for one day, so $325 per megawatt-day is about $119,000 per megawatt per year. UCAP, unforced capacity, is a plant's rated output adjusted for how reliably it performs at peak. The 2028/29 auction's total value was $16.4 billion. PJM, its states and FERC had agreed a price cap and floor for four auctions to limit swings, which means the cap, not the shortage, set the price. Without it the price would have been higher.

Only 525 megawatts of new and uprated generation cleared, less than one-tenth of the shortfall. High prices are supposed to call forth new plants, but a plant that clears an auction in 2026 for 2028 must already be well advanced through the interconnection process, and few were.

The other two large markets give different readings. MISOThe Midcontinent Independent System Operator, which runs the grid and the wholesale power market across much of the central United States.'s auction for June 2026 to May 2027 cleared summer capacity at $424.30 per megawatt-day in most of its zones, down from $666.50 a year earlier, and found enough supply for the year, according to the American Public Power Association's report of the results¹⁰⁶. Capacity offered rose by as much as 4% in each season, mostly from new solar plants and delayed retirements. ERCOTThe Electric Reliability Council of Texas, the grid operator for most of Texas. has no capacity auction at all. Texas runs an energy-only market, in which plants are paid only for the electricity they produce, as a March 2026 presentation to an ERCOT working group¹⁰⁷ sets out. In Texas the queue of large loads, not an auction price, is the visible measure of strain.

How a new plant gets connected

Every new power plant in an organised US marketA region where an independent grid operator, such as PJM, MISO or ERCOT, runs the transmission system and the wholesale power market. Elsewhere, utilities run their own systems. must pass through an interconnection queue, in which the grid operator studies whether the grid can absorb its output and what upgrades are needed and who pays for them. Lawrence Berkeley National Laboratory's annual survey, Queued Up 2025¹⁰⁸, counted about 2,290 gigawatts of generation and storage actively seeking connection at the end of 2024. Only about 13% of the capacity that requested connection between 2000 and 2019 had been built by the end of 2024. The typical project completed in 2024 took 55 months from its request to commercial operation, against 36 months in 2015 and 22 in 2008.

Most of what enters the queue is speculative and leaves it without being built, and what remains moves slowly because each project's studies depend on the projects ahead of it. Reforms are under way, including FERC's Order 2023 and the operators' own rule changes, and LBNL reports that more than 700 gigawatts withdrew in 2024 as they took effect. MISO, the Midcontinent operator, created an expedited track called ERAS for a limited number of projects that meet urgent reliability needs. By September 2026 it had drawn 58 projects totalling about 29 gigawatts, with the latest round of 7.3 gigawatts expected online by 2029, according to T&D World¹⁰⁹. PJM is creating a similar expedited track for up to ten state-sponsored projects.

How a large load gets connected

Data centers go through a separate process as loads, customers that take power rather than supply it. Figure 15 shows the usual sequence. The developer asks the local utility for service and pays for studies. The utility checks its local wires and substations. If the load is large enough to affect the transmission system, the grid operator studies it too. The state utility commission sets the terms on which the utility will serve it. Then the developer signs a long-term agreement with financial guarantees, the utility builds the substation and lines, and the grid operator allows the load to switch on, often in phases.

Texas shows how quickly this process can be swamped. ERCOT, the operator of most of the Texas grid, told the state Senate in July 2026 that it was tracking "approximately 474 GW of Large Loads seeking interconnection, of which ~90% are data centers," according to its testimony¹¹⁰. Under a state law known as SB6, ERCOT is moving to study large loads in batches. Of the requests, about 205 gigawatts had studies good enough to qualify for the first batch; 315 projects did not. ERCOT's own generation queue held about 463 gigawatts, mostly solar, wind and batteries, with 78 gigawatts of gas.

Ohio shows the same pattern at the scale of one utility. Figure 16 sets AEP Ohio's actual data-center load against the requests it received. Load grew from 100 megawatts in 2020 to about 600 megawatts in 2024, and AEP projected 5 gigawatts by 2030, but preliminary requests totalled "more than 30,000 MW," according to POWER magazine's report¹¹¹ on the state commission's decision. AEP had stopped signing new data-center service agreements in central Ohio in March 2023 until the new rules were in place. A request of this kind works like an option in finance: it holds a place in the utility's line, and the developer can still decide not to build. Those rules, described below, are meant to make holding that place cost something.

Who pays

When a utility builds lines and substations for a customer who never arrives, the cost is spread across everyone else's bills. The state commissions' answer has been to make large loads commit. The Public Utilities Commission of Ohio approved a tariff in July 2025¹¹² requiring new data centers above 25 megawatts to pay for "at least 85% of their subscribed monthly capacity for up to 12 years, regardless of actual consumption," with exit fees and financial guarantees. Virginia's State Corporation Commission created a new rate class¹¹³ for customers of 25 megawatts or more, effective January 2027, with "14-year commitments to pay at least 85% of their contracted distribution and transmission demand and at least 60% of their generation demand," according to Cardinal News¹¹⁴. The commission said the purpose was "to help insulate ratepayers."

FERC has moved in the same direction for the grid operators it regulates. On 18 June 2026 it issued show-cause orders to all six operators it oversees, requiring them to justify or change their rules for connecting large loads, according to the order issued to PJM¹¹⁵ and a summary by the law firm Sidley Austin¹¹⁶. The orders call for standard cost recovery agreements requiring a minimum level of financial responsibility "if the projected load ultimately fails to materialize," and for flexible, non-firm transmission service for loads willing to cut their use when the grid is stressed. Firm service is delivered at every hour, including the few hours of peak demand the grid is built to meet. Non-firm service can be cut back in those hours, so it can start before the upgrades that firm service would need. PJM's version, which it calls "Connect and Manage," would let data centers connect sooner if they accept being curtailedTold by the grid operator to cut power use, or a plant's output, for a period because the grid is short of capacity. at times.

Co-location and behind-the-meter arrangements

A co-located data center sits beside a power plant and takes some of its output directly. Behind-the-meter means the load is served on the plant's side of the grid meter, so it avoids some grid charges and some grid studies. The most prominent case shows how hard that is to sustain. Talen Energy agreed in June 2025 to supply Amazon with 1,920 megawatts from its Susquehanna nuclear plant through 2042, and said the existing co-located arrangement "will transition to a 'front-of-the-meter' arrangement after the completion of transmission reconfigurations expected in the Spring of 2026," according to its announcement¹¹⁷. In other words, the deal moved back onto the grid. FERC's December 2025 order¹¹⁸ directed PJM to write clear rules for co-located loads, and the June 2026 orders extend the question to every region.

The supply side: turbines, nuclear, contracts

New generation for data centers in the near term is mostly natural gas, and the gas-turbine makers are booked. GE Vernova reported a gas-equipment backlog of 53 gigawatts plus 63 gigawatts of slot reservation agreements, paid deposits that hold a place in its production schedule, at the end of June 2026, in its second-quarter release¹⁰³. It expects "at least 125 GW" under contract by the end of the year, against output of 20 gigawatts a year from the third quarter of 2026, rising to 24 gigawatts in 2028 and 30 gigawatts in 2030. At 2026 output, its current backlog and reservations equal about 5.8 years of production. Siemens Energy's gas business¹⁰² reported a book-to-bill ratio of 2.65, meaning it took orders worth 2.65 times what it shipped in the quarter, citing "large orders related to data centers."

Nuclear plants supply firm power with no emissions, but new reactors are slow to build, so the near-term options are restarts, uprates and contracts with existing plants. Constellation said the Nuclear Regulatory Commission had approved the fuel licence amendment for its Crane Clean Energy Center, formerly Three Mile Island Unit 1, "a major milestone moving us closer to restarting operations in 2027," in its second-quarter release¹¹⁹. Holtec's Palisades plant in Michigan missed a February 2026 restart target and had more than 5,000 work activities still to complete in July, according to the American Nuclear Society¹²⁰. NextEra plans to restart the 610-megawatt Duane Arnold plant in Iowa in early 2029 under a 25-year agreement with Google, according to World Nuclear News¹²¹.

Power purchase agreements (PPAs), long-term contracts to buy a plant's output at an agreed price, are how most of this is financed. They are typically 15 to 25 years. Constellation alone¹¹⁹ signed 920 megawatts of new agreements in the second quarter of 2026 with terms of 15 to 20 years starting between 2029 and 2032. Those start dates show how long new firm power takes to arrive.

Where it stands

The market data say power is now the step that decides when finished buildings fill. CBRE, the property broker, reported a record 7,481 megawatts under construction in North America's eight main data-center markets in the first half of 2026 but a vacancy rate of just 1.4%, and said "power constraints are extending completion timelines," in its September 2026 report¹²². The Energy Information Administration expects¹⁰ the West South Central region, which includes ERCOT, to supply about 40% of US electricity sales growth in 2027, and PJM about 45% of generation growth. The grid is slow to respond because of the sequence in Figure 15. Each party's step is reasonable on its own, and together they take four to six years.

The same CBRE report counts capacity at each stage. In the eight primary markets, 10.9 gigawatts were operating in mid-2026, 33.7% more than a year earlier. Another 7.5 gigawatts were under construction, more than 80% of it already leased, and less than 1.5 gigawatts was free for a new tenant. Requested capacity is a different order of size: AEP Ohio alone received requests for 30 gigawatts, and ERCOT is tracking 474 gigawatts of large loads. Operating capacity grows by gigawatts a year; requests are counted in hundreds of gigawatts.

Developers, neoclouds and financing

At the end of June 2026 Microsoft had signed "additional leases, primarily for datacenters, that had not yet commenced of $329.1 billion," due to start between fiscal 2027 and fiscal 2033 with terms of one to twenty years, according to its annual report²¹. Oracle reported $288 billion of similar commitments, "substantially all related to data center arrangements," starting by fiscal 2029 for terms of fifteen to nineteen years, in its latest 10-Q¹²³. These numbers show who will own and pay for most AI data centers. Much of the building is done by others, and the hyperscalers rent it.

Who owns the buildings

Three groups build and own AI data centers. The hyperscalers build some themselves. Developers and real-estate investment trusts (REITs), companies that own property and pay out most of their income as dividends, build others and lease them out. And a growing set of project companies, often backed by infrastructure funds, build single very large campuses for one tenant. Digital Realty, a data-center REIT, operated 2,515 megawatts of IT load at the end of June 2026 and had 1,402 megawatts under construction, 54% of it already leased, according to its 10-Q¹²⁴. It estimated the stabilised cash yield on that construction at about 11.5%: for each dollar spent, about 11.5 cents a year of rent once the building is full. Equinix also develops large hyperscale buildings through joint ventures it calls xScale, of which it had 23 among 282 data centers, according to its 10-Q¹²⁵.

Land with secured power has become the scarce asset. Digital Realty reported "over 7 gigawatts of additional future development capacity," about 70% of it at sites with more than 100 megawatts of buildable capacity. That is nearly three times what it operates today. Those 7 gigawatts also show the distinction between announced and built: they are land and power positions, not buildings.

Lease terms against asset lives

The structures in Figure 17 match long buildings to long leases, but the equipment inside turns over far faster. Nvidia now releases a new rack generation roughly every year, and each has drawn more power per rack than the last. A building leased for fifteen years will host several generations of chips, each needing more power and cooling than the one before. The lease protects the landlord. The question is who carries the risk that the building, or the chips in it, lose value faster than expected.

Meta's Louisiana campus shows one answer. Meta holds 20% of a venture that is spending about $27 billion to build the campus. Its leases start in 2029 with "an initial four-year lease term and options to renew for a total lease period of up to 20 years," according to its 10-Q¹²⁶. Because four years is much shorter than the building's life, Meta has also provided residual value guarantees, promises to make up any shortfall in the property's value if it does not renew, "with an aggregate threshold of approximately $28 billion that decreases over time." Meta does not consolidateTo consolidate a venture is to include all of its assets and debts in your own financial statements. An owner without control reports only the value of its stake. the venture, because it does not control it. In July 2026 it agreed to a second venture of the same kind in El Paso, Texas, with residual value guarantees of up to about $13 billion. The arrangement keeps the debt off Meta's balance sheet while leaving Meta with much of the economic risk.

The neocloud model

Neoclouds are companies that rent out AI computing capacity and nothing else. CoreWeave's filings show how the model works. It signs multi-year contracts with a few large customers, borrows against those contracts and the hardware they pay for, buys accelerators, and installs them in buildings that developers build and power for it. Its revenue was $2.6 billion in the quarter to June 2026, and its total debt was $35.6 billion, according to its 10-Q¹²⁷. Its three largest customers provided 36%, 26% and 10% of its revenue; in March 2026 Meta committed to pay it up to about $21.0 billion through 2032.

The debt is what the industry calls GPU-backed. CoreWeave's $8.5 billion delayed-draw term loan from March 2026 can be drawn as equipment is bought, and borrowings are limited by "debt-sizing limitations tied to the purchase price of eligible assets," based on "the depreciable cost of computing equipment, projected debt service coverageHow many times a project's cash flow covers the interest and repayments due on its debt over a period. Lenders set a minimum before they will lend. and project-level conditions." In plain terms, lenders advance money against the accelerators and the customer contract that pays for them, and the loan must be repaid out of that contract's cash flow. The interest rates show how the market prices the risk: CoreWeave sold notes due 2031 and 2032 at 9.75% and 9.625% in 2026.

The power for these machines arrives on the developers' schedule. CoreWeave reported 393 megawatts of electrical power still to be delivered at one partner site and 355 megawatts at others, "in phases between 2026 and 2028." Nebius, another listed neocloud, spent $5.66 billion on property and equipment in the quarter to June 2026 against revenue of $582 million, and reported operating cash flow of $2.25 billion, according to its second-quarter release¹²⁸. Spending ten times a quarter's revenue on equipment in a single quarter is only possible with money raised from customers, lenders or shareholders ahead of the revenue it is meant to earn.

Vendor financing

Chip suppliers have started to guarantee their customers' payments. In August 2026 Nvidia agreed to guarantees "capped at a total of $105 billion" to support leases for about 4.25 gigawatts of IT load that an OpenAI affiliate is taking at SB Energy's PORTS campus in Pike County, Ohio, according to its 10-Q²⁹. The guarantees become effective as each of nine phases is completed, the first expected in Nvidia's fiscal 2029, and in exchange the site "will exclusively host NVIDIA AI infrastructure." The same month Nvidia signed memorandums with capital providers to create financing platforms meant to "mobilize more than $500 billion of third-party capital," with Nvidia able to offer "limited residual-value support." Broadcom has built a similar structure, which it calls the AI XPV platform, to fund more than 20 gigawatts of capacity using its custom chips for frontier AI labs through 2028, starting with a $35 billion tranche for more than one gigawatt; it says it "may provide residual value guarantees," according to its 10-Q²⁶.

Broadcom's filing states the purpose plainly: the platform "bridges the gap between the current cash flows of the leading frontier AI labs and significant upfront investments required for their businesses." The buyers whose demand drives the chain cannot all pay for it from cash flow, so the suppliers with the strongest balance sheets are lending them their credit. Section 20 asks what that does to the distribution of risk, and Section 21 recalls the last time equipment makers financed their customers on this scale.

How fast this layer can grow

Money has not been the constraint so far. Buildings take years from land to occupancy, mostly because of power and long-lead equipment, as CoreWeave's phased power deliveries show, while financing such as its March 2026 loan has been arranged quickly. The layer's real limit is that it depends on the creditworthiness of a small number of tenants and on the continued willingness of lenders and suppliers to advance money against contracts that run for years. That limit does not show during a boom, only when a tenant changes its plans.

The ground underneath

On 14 July 2026 the governor of New York signed an executive order directing the state's environmental regulator to pause its review of incomplete permit applications for data centers "that consume or can consume at least 50 megawatts of energy," according to a summary by the law firm Beveridge & Diamond¹²⁹. It was the first statewide pause of its kind. CoreWeave's quarterly report cites it as an example of how "government authorities have in the past sought to restrict data center development." Every layer above this one rests on land, water, permits, workers and trade rules, and each of those is set locally or politically rather than by the market for chips.

Land and permits

A data center needs flat land, a transmission line nearby, fibre routes, water or a cooling alternative, and a community willing to host it. Cushman & Wakefield's head of data-center insights said in August 2026 that "access to power, land, and skilled labor are increasingly determining where growth can occur," in comments on its 2026 cost guide¹³⁰, which found construction costs up 21% per megawatt since late 2024. Land itself is a small share of the cost, about 7% of a greenfieldBuilt from scratch on undeveloped land, with no existing building or connection to reuse. project in that guide as reported by CRE Daily. Land with power is the expensive part, which is why developers hold large banks of it.

Federal policy has moved to speed approvals. Executive Order 14318, signed in July 2025, defines a "Data Center Project" as a facility requiring "greater than 100 megawatts (MW) of new load dedicated to AI," and directs agencies to identify exclusions from environmental review under the National Environmental Policy ActThe federal law that requires agencies to study the environmental effects of projects they approve, fund or carry out. An exclusion lets a project skip the full review., make federal land available and offer financial support, according to the Federal Register text¹³¹. Most data-center permits are local and state, however, and New York's order shows that states can move the other way. Local objections are another brake; CoreWeave's filing notes that authorities have restricted development "based on environmental considerations," but such objections rarely appear in filings except as a general risk.

Water

Section 12 explained the trade between water and electricity in cooling. Its local effect depends on where the building is. In dry regions the water a cooling tower evaporates becomes a local issue; in wet ones it rarely does. Microsoft's switch to closed-loop designs that avoid "more than 125 million liters of water per year per datacenter" was a response to that pressure. The practical result is that water is less a limit on the industry as a whole than a factor in choosing sites, and the choice often comes down to whether water or power is scarcer in a given place.

Skilled labour

Every layer from the electrical yard to the cooling loops is installed by skilled trades. The Bureau of Labor Statistics counted 821,000 electrician jobs in 2025 and projects about 72,700 openings a year over the following decade, according to its Occupational Outlook Handbook¹³². Data centers compete for those workers with factories, transmission projects and housing. TSMC said its Arizona expansion alone would support "40,000 construction jobs over the next four years." Wood Mackenzie names skilled labour among the reasons transformer makers cannot expand faster. Labour is one reason construction costs rose faster than general inflation, and one reason prefabricated power and cooling modules, which move work into factories, are spreading. No filing lists it as a single shortage.

Export controls

The United States controls the export of advanced AI chips and of the equipment to make them, mainly to limit China's access. The rules change often, and they shape where demand in the chain can go. Figure 18 lists the main rules in force or recently changed.

The January 2026 rule is the most revealing for this report. It allows exports to China of chips such as Nvidia's H200 and AMD's MI325X, those with a total processing performanceA score defined in the US export rules that combines how many operations a chip can perform per second with the size of the numbers it works on. A higher score means a more powerful chip. below 21,000 and memory bandwidth below 6,500 gigabytes a second, on a case-by-case basis, provided the exporter certifies "there is sufficient supply of this product in the United States" and that "production of this product for exports to China will not divert global foundry capacity for similar or more advanced products for end users in the United States," according to the rule as filed¹³³. The licence condition treats foundry capacity as the scarce resource to be protected. Nvidia reports that it received such licences beginning in February 2026, that shipments must pass through inspection in the United States and so incur a 25% tariff on import, and that Chinese authorities restricted purchases, so that the shipments it made were less than 1% of its data-center revenue in the latest quarter.

Equipment controls cut the other way for the tool makers. China was 42% of Lam Research's revenue in fiscal 2024 and 34% in fiscal 2026, and its share of Applied Materials' revenue fell from 35% to 28% between the same quarters a year apart. The December 2024 rule⁴¹ also added controls on high-bandwidth memory, which matters for the memory makers' customer lists. On the other side, the July 2026 rule¹³⁴ on the United Arab Emirates gave approved entities there licence-free access to advanced computing items, under a cooperation framework agreed in May 2025. Export controls do not limit the pace of the US buildout directly. They decide which customers the chain may serve, which in a period of scarcity means more of its output stays with US and allied buyers.

Where a dollar of AI spending lands

Most of a dollar of AI capital spending is paid out near the top of the chain. On a campus filled with Nvidia's newest systems, about half becomes the accelerator vendor's gross profit. The grid, the step that takes longest to expand, is barely in the budget at all.

Jensen Huang, Nvidia's chief executive, has put rough numbers on a gigawatt of AI capacity. According to a September 2026 account of his remarks¹³⁵, he puts the cost of building one at roughly $50 billion and Nvidia's share at roughly $18 billion a gigawatt for Hopper, $25 billion for Grace Blackwell and about $40 billion for Vera Rubin. Those are a vendor's own estimates, reported second-hand without a date or venue, and should be read as orders of magnitude. Combined with Cushman & Wakefield's 2026 estimate of $8.9 million to $23.3 million per megawatt for the building and its plant, excluding chips, they allow a rough split of where the money goes. Figure 19 shows it for $100 of spending.

On these assumptions about $73 of every $100 goes on IT equipment and about $27 on the building and its electrical and mechanical plant. The accelerator vendor's share is about $67, of which about $50 is gross profit at Nvidia's reported 75% gross margin and about $17 pays for the vendor's own inputs: wafers, packaging, memory, boards and assembly. The electrical plant inside the fence, the transformers, switchgear and UPS of Section 13, takes about $6. The figure is illustrative. Different chips, different sites and different prices would move every band. But reasonable changes to the inputs do not change the main result: the largest single destination of a dollar of AI capex is the gross profit of the company that designs the accelerator.

A market-wide check gives a smaller share. Nvidia's data-center revenue over the four quarters to July 2026 was about $277.8 billion, from its annual report¹³ and latest 10-Q²⁹. At its 75% company-wide gross margin that is about $208 billion of gross profit, about 36% of the $586 billion the five largest buyers spent on property and equipment over roughly the same months. The two numbers do not cover the same buyers. Nvidia also sells to neoclouds, governments and other companies, and the five buyers' spending includes their own chips, ordinary cloud servers, and land and buildings for racks that arrive later. So about half is the right figure for a hall filled with Nvidia's newest racks, and about a third is the better guide to the market as a whole.

Margins as a map of scarcity

Gross margin, revenue minus the direct cost of what was sold, as a share of revenue, is the simplest public measure of how much pricing power a company has. Figure 20 sets the latest gross margins of eight companies in the chain side by side. They line up with the account of scarcity in Part I. The memory maker, whose product is short, is at the top. The accelerator designer and the leading foundry, each the dominant supplier of a hard-to-replace step, follow. The equipment makers sit in the middle. The power and cooling specialist earns less. The assemblers, whose work can be added in months and done by several firms, earn the least.

Two cautions apply. These are company-wide margins, not margins on AI products alone, and they cover different periods and accounting standards. And margin is not the same as profit captured from AI: Micron's margin reflects scarcity in all of its products, including ordinary DRAM sold into PCs and phones. Still, the ranking matches the mechanism. Where a step has few suppliers and capacity that takes years to add, the supplier can price for scarcity. Where capacity can be added in months, competition takes the margin away.

Money is moving down the chain earlier

A second change is in timing. In a normal supply chain each supplier is paid on delivery. In this one, buyers are paying in advance to secure capacity, and the payments now run several layers down. GE Vernova's cash from operations so far in 2026 was $9.2 billion higher than a year earlier, mostly because of "higher down payments on orders and slot reservation agreements at Power," according to its 10-Q¹³⁶. Oracle reported¹²³ $11.4 billion of "customer prepayments with significant financing component for capital expenditures" in a single quarter, meaning its customers are paying in advance for the data centers Oracle builds to serve them. Micron's take-or-pay agreements come with customer deposits. Nvidia reported²⁹ $2.8 billion of customer advances at the end of July 2026, up from $160 million six months earlier.

Prepayment is how a buyer moves up a queue. It also moves risk. A supplier holding a customer's deposit has less to lose if the customer walks away; the customer has more. Section 20 follows that risk to where it now sits.

The grid is paid for differently

The most important thing about Figure 19 is that it leaves out the grid. Most of the cost of new power plants, transmission lines and grid substations falls outside a data center's capital budget. Utilities and power producers pay it and recover it over decades through regulated electricity rates, capacity payments and long-term power contracts. PJM's 2028/29 capacity auction¹⁰⁵, for example, committed $16.4 billion of payments to power plants for one year of availability; that money comes from everyone's electricity bills, not from the hyperscalers' capex. A utility's prices are set by its regulator, the state commission or, for interstate transmission, FERC. The regulator lets the utility recover the cost of approved lines and substations from customers' bills over the decades the assets last, plus a profit rate the regulator sets. The new state tariffs in Section 14 exist precisely because utilities and regulators do not want the costs of serving data centers to land on households.

This is the underlying reason the slowest step stays slow. The fastest-growing pool of capital in the chain, the hyperscalers' budgets in Figure 1, mostly flows to the parts of the chain that can expand in one to three years and that price for scarcity. The slowest parts, generation and transmission, are financed by a different kind of capital, regulated and long-dated, that is approved by public bodies on their own calendars and earns a regulated return. The equipment makers that serve them, such as transformer manufacturers who believe a new plant can take "decades" to pay back, are cautious for the same reason.

There are signs the two systems are starting to connect. PJM is matching¹⁰⁵ large loads with new generators in bilateral contractsContracts negotiated directly between two parties, here a large power user and a generator, outside the grid operator's auction. of ten years or more. Nvidia's $105 billion guarantee²⁹ in Ohio covers "defined portions of lease and power payments," which pulls power costs into a chip supplier's credit support. And the hyperscalers' long power purchase agreements with nuclear and gas plants are a way of paying for generation directly. If those arrangements spread, the grid will move faster. Until they do, the dollar and the delay will sit in different parts of the chain.

Which step binds, and how the binding step has moved

By 2025 most people in the industry would have said the AI buildout was waiting on power. In 2023 they would have said packaging. That is a story of a single bottleneck that moves from one step to the next. The evidence supports part of it.

Start with a simple rule. A step binds when demand for its output grows faster than its capacity can be added. So what matters is how a step's lead time compares with the warning the industry had, more than how slow it is in absolute terms. A step that takes a year to expand binds only when demand surprises within a year. A step that takes five years binds whenever demand runs above plans for five years. Figure 21 sets out the lead times side by side.

2023–24: packaging

When demand for AI accelerators jumped in 2023, the step with the least spare capacity was TSMC's CoWoS packaging, about 12,000 wafers a month that September. TSMC described the constraint as packaging rather than wafers. Packaging lines were small, and the tools to expand them took about a year to deliver. The shortage lasted about as long as that year of lead time, plus the time it took for demand to keep rising. By 2025 TSMC had expanded CoWoS several times over, and while packaging is still reported to be about 20% short in 2026, it is no longer the single limit. This part of the story holds up well.

2025–26: power

Power became the headline in 2025 for the reason in Figure 21: its lead times, from transformers to turbines to the grid queue, run three to six years, longer than any surge in AI demand had been anticipated. The evidence that it binds is strong. PJM's last two capacity auctions fell short of its reliability requirement for the first time in its history. Only 525 megawatts of new generation cleared the latest one. CBRE reports¹²² that power constraints are "extending completion timelines" for buildings that are already under construction and almost entirely pre-leased. ERCOT faces¹¹⁰ 474 gigawatts of requests. GE Vernova's gas-turbine queue¹⁰³ is nearly six years of output long.

2026: memory joins

Here the single-bottleneck story needs amending. If power were the only binding step, chips would be easy to buy and would pile up waiting for buildings. The filings show the opposite. Nvidia says²⁹ it is "currently experiencing certain supply constraints" and has more than doubled its supply commitments, mostly for memory, in one quarter. Amazon says it cannot meet demand even at $220 billion of capex and raised that figure because memory cost more. Micron's revenue has nearly quintupled in a year, and SK hynix and Samsung describe a market where demand exceeds supply. Micron's gross margin in Figure 20 is higher than the accelerator designer's. Those are the signs of a binding step.

Memory binds for a reason specific to its history. Its lead time, two to three years for a new fab, is not the longest in the chain. But the industry cut capacity in 2023, when Micron's gross margin was negative, so it entered the AI surge with less under construction than demand needed. A step can bind because of what was not started three years ago as much as because of how slow it is.

The better description of 2026 is two binding steps on two clocks. Memory and, to a lesser degree, packaging limit how many accelerators can ship this year. Power limits where and when finished buildings can be switched on. The two interact. A hyperscaler that cannot get power for a site in 2027 will not take all the chips it ordered for it, and a neocloud that has power but no memory will have an empty hall. Which one is felt first depends on the buyer. A company with buildings ready, such as a REIT tenant in a market with power, feels the chip shortage. A company with chips on order and no grid connection feels the power shortage.

What would make a different step bind next

The same rule predicts the next change. Memory capacity now under construction arrives in 2027 and 2028: Micron's Taiwan fab⁶⁷ from mid-2027, its Singapore HBM plant from the first half of 2027, SK hynix's Yongin cleanroom in early 2027. If demand growth slows even slightly while that capacity comes online, memory could flip from shortage to surplus quickly, as it did in 2022–23, and power will be left as the main limit. If demand keeps compounding, three other steps are candidates. Lithography is growing about 30% a year, so leading-edge wafer demand growing faster than that for several years would make EUV tools binding, with Zeiss's optics output as the limit beneath them. Large transformers and switchgear are still on lead times of two to four years. And skilled electrical labour cannot be expanded quickly by any means.

There is also a case in which none of the physical steps binds, because the money does. The buildout is financed by the expectation that AI revenue will grow into it. If that expectation weakens, the binding constraint becomes the buyers' willingness to keep committing capital years ahead, and every step in Figure 21 would find itself with spare capacity at different times, the long-lead steps last. Section 21 shows how earlier buildouts ended in exactly that way.

What breaks when a critical step stalls

An AI data center produces nothing until two separate chains meet: a rack of working accelerators, and a hall that is connected to the grid, cooled and switched on. Figure 22 maps how the layers feed those two chains. The map explains a pattern that is otherwise puzzling, in which some companies have chips and no power while others have power and no chips.

The silicon chain runs from lithography tools to logic wafers, joins the memory chain at the packaging step, and ends in a rack. The power chain runs from turbines and transformers through the grid connection and the site's own electrical plant, and ends in an energized hall. The two meet only at the last step. That has a practical consequence: a stall anywhere upstream does not stop the other chain from spending money. Racks can be built and shipped to a building that cannot yet take them, and buildings can be finished and powered with nothing to put in them. In both cases very expensive assets sit idle and earn nothing.

What happens when each critical step stalls

How critical a step is depends on four things: who controls it, how replaceable it is, how long a replacement would take, and what waits on it. Figure 23 works through the stalls most often discussed in the industry. The table is a synthesis of the layer sections; each entry rests on the sources cited there.

How a stall propagates

The two kinds of stall propagate differently. A silicon-side stall moves fast and shows up in shipments. When packaging was short in 2023 and 2024, it capped how many accelerators could ship, and it eased over about a year as new tools arrived. When memory is short, as now, the effect spreads sideways into other markets: Nvidia reports⁷ that its PC chip sales were held back by "elevated memory and systems prices," as memory makers allocate output under long-term agreements to the largest buyers.

A power-side stall moves slowly and shows up as idle capital. A building whose transformer is late or whose grid connection is phased can sit partly empty for a year or more while its owner pays for land, construction loans and, under the new tariffs, minimum electricity charges. The racks destined for it either wait or go elsewhere. This is where the long-term contracts in Section 15 matter. A tenant with a fifteen-year lease and a utility agreement requiring 85% minimum payments has committed to costs that begin whether or not the building is earning.

Second-order effects

Three consequences follow from the map. First, power pushes the buildout towards places where power can be found fastest, not towards where demand or customers are. ERCOT's queue and AEP Ohio's requests are the visible result, as is CBRE's remark¹²² that "power is the deciding factor in Northern Virginia right now." Second, the silicon chain's response to its own shortages, committing to suppliers years ahead, is making it less flexible. Nvidia's $279 billion of supply commitments and Micron's take-or-pay contracts lock in volumes that will have to be absorbed whether or not the power side keeps pace. Third, the two chains now have different owners of risk. Silicon-side risk sits with chip vendors and their suppliers, who can see orders. Power-side risk sits with utilities, regulators and, through tariffs, ordinary electricity customers, who cannot. That is the subject of the next section.

Who carries the financial and physical risk

Every long-lived asset in this chain was paid for by someone who expects to be repaid over years. If demand for AI computing grows as planned, the question of who carries the risk never comes up. If it does not, the answer decides who loses money. The filings make it possible to trace that answer more precisely than is usual for a boom in progress, because the largest commitments now have to be disclosed.

The principle: risk moves towards the strongest balance sheet

Each contract in the chain is an attempt by one party to make another carry a risk. Utilities and their regulators have made data centers commit to paying at least 85% of their reserved capacity for twelve to fourteen years, so that households do not pay for lines built for customers who never arrive. Equipment makers take deposits and slot reservation payments before they build. Memory makers sign take-or-pay contracts. Landlords sign leases of fifteen to twenty years. Each of these moves risk from the seller to the buyer. At the top of the chain, the buyers with the weakest cash flows, the AI labs and the neoclouds, have in turn found sellers willing to take risk back: Nvidia and Broadcom guarantee their payments, and Meta guarantees the residual value of buildings it leases for only four years.

Figure 24 sets out where the main commitments now sit.

Where the exposure is concentrated

The hyperscalers carry the largest commitments, and they also have the largest operating cash flows to meet them. Even so, the buildout has started to strain cash. Alphabet's free cash flow²⁸, operating cash minus capex, was negative $5.9 billion in the second quarter of 2026, and it raised $49.6 billion of new equity in June. Amazon's free cash flow² over the twelve months to June was an outflow of $7.6 billion. Oracle's free cash flow was negative in every quarter of its last fiscal year and in its latest quarter, and it sold $19.9 billion of shares in the quarter to August. Oracle's $664 billion of contracted future revenue is large, but only about 13% of it is expected to be recognised in the next twelve months, according to its 10-Q¹²³, so most of the cash it is spending now is a bet on revenue several years out.

The AI labs are the least visible link. Their commitments appear in their suppliers' filings: more than $100 billion each to Amazon¹³⁷ from OpenAI and Anthropic, up to $6.5 billion from OpenAI to CoreWeave¹²⁷, 4.25 gigawatts of leases in Ohio backed by Nvidia's guarantee, two gigawatts of AMD accelerators for Anthropic. Their own finances are private. Broadcom's description of the gap between "the current cash flows of the leading frontier AI labs and significant upfront investments" is the clearest public statement that these buyers rely on others' credit. Nvidia's guarantees²⁹ "terminate upon certain events, including OpenAI achieving a satisfactory credit rating," which says the same thing in contract language.

Neoclouds carry concentrated risk. CoreWeave's debt is secured on accelerators and contracts, three customers provide most of its revenue, and the value of its collateral depends on how fast those accelerators lose value as newer generations arrive. Lenders limit this by sizing loans to the "depreciable cost of computing equipment" and to contracted cash flow. The test of that structure would come if a large customer reduced its commitments.

The physical risk

Alongside the financial risk is a physical one: assets built for one set of assumptions that do not fit the next. A building designed for 130-kilowatt racks may not be able to host one-megawatt racks without new electrical and cooling plant. Transmission lines and substations built for a data center that is cancelled are paid for anyway, which is why the new tariffs exist. Gas turbines ordered for data-center load are long-lived and will run for decades whether or not that load arrives. And memory fabs started in 2026 will produce into whatever market exists in 2028.

The ordinary electricity customer is the party with the least say. Utilities recover grid investments through rates. The Virginia commission's decision in Dominion's rate caseThe formal proceeding in which a utility asks its state commission to approve new prices, and the commission decides which costs customers must pay. raises the average residential bill by $13.60 a month over 2026 and 2027, against $19.57 requested, according to Cardinal News¹¹⁴, while creating the new class that makes data centers pay more of their own costs. PJM's capacity payments, $16.4 billion for one delivery year, are recovered from electricity customers across its region, although load hedged through its own supply or bilateral contracts is not exposed to the auction price. Whether the new minimum-payment tariffs are enough to protect households if data-center demand falls short has not yet been tested.

How earlier buildouts ended

Large infrastructure buildouts rarely end because the technology fails. They end when the supply built in anticipation of demand outruns the demand that arrives, and the financing that assumed otherwise breaks. The physical assets usually survive and are used later, by different owners at lower prices. Four earlier cycles show the mechanisms in detail, and each has a counterpart in the chain described in Part I.

Telecom fiber, 1996–2002

The Telecommunications Act of 1996 opened local and long-distance markets to competition just as fiber-optic technology was improving fast and internet use was taking off. Investment in communications equipment grew from about $62 billion a year at the start of 1996 to more than $135 billion by the end of 2000, measured in 1996 dollars, according to a Federal Reserve Bank of Richmond study¹³⁸ by Elise Couper, John Hejkal and Alexander Wolman. A Nasdaq index of telecom stocks rose from 198 in April 1997 to 1,230 in March 2000. New companies such as Global Crossing and Level 3 laid long-haul fiber networks "in the hopes of selling capacity."

Two things ended it. The first was a demand forecast that was wrong. The industry repeated a claim that internet traffic was "doubling every 100 days." Andrew Odlyzko, a researcher at the University of Minnesota, called the claim a myth and estimated that traffic was growing more like 50–60% a year, in a paper on traffic growth¹³⁹. The second was efficiency on the supply side. The Richmond Fed study notes that in 1996 a strand of fiber carried about 2.5 gigabits a second; by 2000 the same strand could carry 100, because new equipment sent 40 wavelengths of light down it at once. Each fiber already in the ground became forty times as capable without being dug up. Supply grew far faster than even honest demand forecasts.

The financing made the fall worse. Equipment makers lent money to their own customers to win orders. Lucent, then a leading US telecom equipment maker, reported that it had provided "substantial long-term financing to some of our customers worldwide as a condition of obtaining or bidding on infrastructure projects," to "established companies as well as start up companies," and that its commitments had reached $8.1 billion in September 2000, according to its 2002 annual report¹⁴⁰. When the start-ups failed, Lucent recorded provisions for bad debts and customer financing of $2.2 billion in fiscal 2001 and $1.3 billion in 2002, much of it from a few customers such as Winstar and One.Tel, and it could not sell the loans to anyone else. By early 2001, the Richmond Fed study says, "it became apparent that massive overinvestment had taken place." Investment in the fourth quarter of 2001 was 69% of a year earlier.

Memory and chip-equipment cycles

Memory follows the same logic in a faster loop, as Section 9 described: makers build in good years, the capacity arrives together, prices collapse, and makers cut until shortage returns. Micron's fiscal 2023, with revenue down 49% and a negative gross margin, is the most recent example. Chip equipment amplifies those swings because it is bought only to add capacity; SEMI's figures show equipment sales fell 46% in 2009. What ends a memory or equipment upswing is the arrival of the capacity that the upswing paid for.

Merchant power, 1998–2005

The closest parallel to today's power build is the merchantIn power, a merchant plant is built without a long-term buyer and sells its output at whatever the wholesale market pays. The chip sense in Section 4 is close: a merchant chip is sold to any buyer, while a custom chip is built for one. gas boom. As electricity markets were opened to competition in the late 1990s, independent power producers built gas plants without long-term contracts, selling into the wholesale market. They favoured gas "because of short construction times and low capital costs," according to the Energy Information Administration¹⁴¹, and about 237 gigawatts of gas capacity was added from 2000 to 2010, 81% of all additions. In 2005 the combined-cycleGas plants that burn gas in a turbine and use the hot exhaust to raise steam for a second turbine. They are the most efficient kind of gas plant. fleet ran at only about 40% of its capacity even during peak hours, according to another EIA analysis¹⁴². Calpine, a leading builder of these plants, filed for bankruptcy in December 2005, according to its 8-K¹⁴³. The plants kept running under new owners. In January 2026 Constellation completed its acquisition of Calpine¹⁴⁴, buying that fleet to serve the demand described in this report.

What carries over, and what does not

Four lessons carry over. First, requests are not demand. The telecom industry built to a traffic forecast that was mostly a sales story, and today's grid queues hold requests many times larger than the load utilities expect: AEP Ohio's 30,000 megawatts against its own 5,000-megawatt projection. Second, efficiency gains on the supply side can create a surplus with no new building. Each fiber became forty times as capable; each new accelerator generation is sold on a lower cost per token, and Epoch AI estimates algorithmic efficiency alone cuts the compute needed for a given result about threefold a year. So far, demand measured in tokens has grown faster than efficiency. If it stopped doing so, capacity built for the old trajectory would be in surplus. Third, vendor financing concentrates the losses on the vendor. Lucent's $8.1 billion was a large sum for its time; Nvidia's guarantees for one campus are capped at $105 billion, and its financing platforms aim at more than $500 billion. Fourth, the assets survive. Fiber laid in 1999 carries traffic today; Calpine's plants are now part of the AI power supply.

Two things are different. The largest buyers today are among the most profitable companies in the world, paying mostly from operating cash flow, which the telecom start-ups were not. And the binding steps now are physical shortages, not just financial exuberance: the power and memory shortages are real, not forecast. The risk is concentrated in a narrower set of borrowers, the AI labs and neoclouds, whose ability to pay depends on revenue that has not yet arrived at the scale of the commitments. That is where an end, if it comes, would start.

The next generations and the physical limits they approach

Each layer of the chain has a published roadmap for the next two or three years, and most of the roadmaps point the same way: more transistors, more memory and more power in each package, each rack and each building. Figure 26 lists the dated milestones the companies and standards bodies have announced. Most of these products will arrive. The question is which physical limits they press against, because those limits decide where the next bottlenecks form.

Chips: lithography and transistors

The logic roadmap continues with one slip. At its April 2026 symposium⁴⁸ TSMC moved mass production of its A16 node from 2026 to 2027 and introduced a further node, A13, for 2029. ASML expects³³ its High-NA EUV machines to "start supporting high-volume manufacturing in 2027," and Intel has qualified them on some layers of its 18A process. The limit here is economic more than physical: each node costs more to develop and each tool more to buy. That favours the largest buyers, who can commit volume years ahead, and it is why Intel says⁴⁴ the pace of its 14A expansion depends on the committed demand it can win.

Memory: taller stacks and the bandwidth wall

The memory roadmap is HBM4E in 2027, with the JEDEC HBM4 standard⁵² already allowing stacks of 16 dies holding 64 gigabytes. Bandwidth per accelerator has risen from 3.35 terabytes a second on the H100³⁰ to up to 22 on Rubin¹⁸, and designers keep adding memory because decode speed depends on it, as Section 3 explained. Taller stacks are harder to keep flat and to cool, which is why each new stacking process is announced as a manufacturing achievement. Heat is the practical limit on how tall and how fast the stacks can go.

Packaging: bigger than a wafer allows

The package is growing because it must hold more processor dies and more memory stacks. The interposer that carries them is cut from a round wafer, which limits its size, so the industry is developing packages built on large rectangular panels. ASE lists such a process⁵⁵, CoPoS, among its technologies. TSMC is reported to plan⁴⁹ CoWoS capacity of about 260,000 wafers a month by the end of 2028, double the end-2026 figure. Larger packages raise the cost of a single failure, because each one carries more good silicon, so packaging yield and testing become more important at each step.

Optics: moving light into the package

Co-packaged optics moves the laser and modulator next to the switch chip and cuts power per port from about 30 watts to about 9, according to Nvidia⁸¹. Nvidia says⁸² its design uses four times fewer lasers. The trade-off is serviceability: a pluggable transceiver can be swapped in seconds, while an optical engine inside a switch package cannot. Whether operators accept that trade at scale will decide how fast co-packaged optics spreads from switches to the accelerators themselves, which would be the larger change.

Racks, power and cooling: the megawatt rack

The largest physical change is at the rack. Nvidia's roadmap⁷⁴ goes to racks of one megawatt from 2027, powered at 800 volts direct current. Google says⁹¹ machine learning will require "more than 500 kW per IT rack before 2030" and is working with Meta and Microsoft on ±400-volt power. A one-megawatt rack needs liquid cooling for essentially everything in it, power distribution at 800 volts direct current, a new standard for data-center equipment, and floors built for much heavier loads. Its arrival means that buildings designed in 2024 for 130-kilowatt racks are likely to need new electrical and cooling plant to host the 2028 generation.

Each new generation moves more of the engineering problem out of the chip and into the package, the rack and the building. That is the reason Section 8 argued the rack is now the product, and the reason the slowest layers of the chain, the electrical plant and the grid, matter more with each generation.

Power over the next five years

The power needed for the AI buildout over the next five years will come from a short list of sources, each with its own dated milestones and its own limit. Figure 27 sets them out. The fastest options add the least, and the largest options take the longest, which is why the next five years will be decided as much by how much existing capacity can be shared as by how much new capacity is built.

Sharing the grid that exists

The fastest source of power for a new data center is spare capacity on the existing grid, which exists for most of the year and disappears at peaks. A data center that agrees to cut its use when the grid is stressed can be connected before the upgrades that would make its supply firm are finished. FERC's June 2026 orders¹¹⁵ push every regional operator towards offering that kind of non-firm service, and PJM plans¹⁰⁵ a "Connect and Manage" framework that would "allow large new loads, such as data centers, to connect to the system but operate flexibly when needed." The open question is how much flexibility AI loads can actually offer. Training jobs can in principle be paused and resumed; inference serving users in real time cannot be paused without moving it elsewhere. The answer will differ by operator and by workload, and no one has yet published operating results at scale.

Gas: the main source of new firm power

Most new firm capacity in the next five years will be gas-fired, and the turbine order books show its pace. GE Vernova signed¹⁰³ 20 gigawatts of new gas equipment contracts in the second quarter of 2026 alone, and expects at least 125 gigawatts under contract by the end of the year, against factory output of 20 gigawatts a year rising to 24 gigawatts in 2028. Siemens Energy's gas business¹⁰² took orders worth 2.65 times its shipments in the same quarter. A turbine reserved today will in many cases not be delivered until the end of the decade. On-site generation, behind the meter, is one way to avoid waiting for the grid, and ABB now markets⁹⁵ data-center power systems built to integrate "battery storage, gas turbines, and renewables." But on-site turbines come from the same factories and need the same air permits and gas pipelines. The federal permitting order¹³¹ of July 2025 covers "natural gas pipelines or laterals" among its covered components, which shows where the government expects the bottlenecks to be.

Nuclear: valuable, scarce, slow

Existing nuclear plants run around the clock and emit no carbon, which makes their output the most sought-after power in the market. Most of the deals so far reallocate output that already reaches the grid, such as Talen's 1,920 megawatts¹¹⁷ for Amazon, rather than adding new supply. Restarts of closed plants add real capacity, but there are very few candidates. Crane's restart¹¹⁹ is targeted for 2027 after its fuel licence approval; Palisades slipped¹²⁰ past its February 2026 target with thousands of tasks remaining; Duane Arnold¹²¹ is planned for early 2029. Duane Arnold's plan runs about three years from agreement to first power. New reactors, including small modular designsReactors much smaller than today's large plants, designed so that much of each one can be built in a factory and shipped to the site., are outside the five-year window on any evidence verified for this report.

Renewables and storage

Solar, wind and batteries dominate the interconnection queues: in ERCOT's queue¹¹⁰ in June 2026 there were 162 gigawatts of solar and 171 gigawatts of batteries against 78 gigawatts of gas. They can be built faster than gas plants once connected, and the IEA expects¹¹ renewables to supply the largest share of the additional electricity data centers use through 2030. Their limit is matching variable output to a load that runs all day, every day. Batteries cover hours, not days, so a data center powered mainly by solar still needs firm capacity behind it, from the grid or from gas.

The realistic picture

Put together, the dated evidence suggests three phases. Through 2027, new AI capacity will be energized mainly where grid headroom exists, under flexible terms, and with gas plants and nuclear restarts already in progress. In 2028 and 2029 the turbines reserved in 2025 and 2026, the first ERAS projects and Duane Arnold arrive. Around 2030 the larger gas and transmission additions land, if the demand they were ordered for is still there. The step that sets the pace of the buildout through the period is the one with the longest queue: firm generation and the grid connection to deliver it.

The open questions that will decide the chain

In February 2025 researchers at Duke University's Nicholas Institute estimated that the 22 largest US balancing areasRegions in which one operator keeps electricity supply and demand matched moment to moment., which carry 95% of the country's electricity load, could absorb 76 gigawatts of new load if that load accepted being cut back for 0.25% of the hours in a year. At 0.5% the figure rose to 98 gigawatts, and at 1% to 126 gigawatts, according to Utility Dive's report of the study¹⁴⁵. The study did not model transmission limits, which could shrink those numbers. Even so, 76 gigawatts is seven times the 10.9 gigawatts operating in CBRE's eight primary markets. It is the strongest evidence against this report's account.

The strongest case against this account

The grid has spare energy for most of the year. What it lacks is rules that let a data center connect on condition that it steps back in the few hours of peak demand, and FERC's June 2026 orders are now writing those rules. On this reading power is a regulatory queue, and queues can be shortened by a commission in a way that a turbine factory cannot.

The second half of the case is about demand. Every step this report calls binding is binding because orders run ahead of capacity, and the orders come from buyers who expect AI revenue to grow into them. Some of the evidence for that expectation is thin. Oracle expects to recognise only about 13% of its $664 billion of contracted revenue in the next twelve months. Alphabet and Amazon had negative free cash flow in their latest periods. Broadcom describes the frontier labs' cash flows as too small for the investments they need. And Micron's largest new contracts carry a ceiling price close to the market price of the second quarter of 2026 and a floor beneath it, according to its May 2026 10-Q⁶⁷. The buyer capped what it will pay and the seller secured a minimum, which is what both sides do when neither is sure how long the shortage will last. In September 2026 CoreWeave raised $4.2 billion of convertible notes at 2.875%, according to its 8-K¹⁴⁶, so money is still available today. The opposing view's point is that it will be available only as long as the revenue keeps arriving.

That view is right that the grid has energy to spare most of the year, and right that money is the constraint behind all the others. It does not explain the 2026 evidence. Memory makers' prices and margins are those of a product in short supply: Micron's gross margin rose from 44.7% to 86.8% in a year. PJM's auction cleared at its cap with only 525 megawatts of new generation, and flexible connection adds no generation to an auction. No AI operator has yet published results from curtailing its load at scale, so the Duke gigawatts are capacity that no one has yet used. The account in this report holds for 2026. The opposing view describes what would have to change for it to stop holding, and Figure 29 lists those signals.

The questions, and what would settle them

Figure 28 sets out six questions that will decide how the chain develops over the next three years. Each has two readings that fit today's evidence, and each will be settled by something that can be observed.

The first question matters most for the next two years, because it decides which of the two binding steps eases first. Memory makers have broken ground on enough capacity to change the market by 2028, and the industry's history says that capacity arriving together turns a shortage into a surplus quickly. The take-or-pay contracts are the new element. If they hold, prices outside the contracts would fall first, and the buyers paying contract floors would be the last to see relief.

The fourth and fifth questions decide who loses money if demand falls short. They are slow to resolve because the commitments run for twelve to twenty years, and the first real test will be the first large customer that tries to walk away from one. Until then the guarantees, tariffs and leases in Sections 15 and 20 are promises whose value depends on the strength of the party making them.

The limits of current understanding

Some of the quantities this subject needs are not disclosed. Companies do not split their capital spending between AI and other uses, so the AI share of the $586 billion in Figure 1 is unknown. TSMC reports that high-performance computing was 58% of its 2025 revenue but does not publish the AI share of its output or its packaging capacity, so the wafer counts in Sections 6 and 7 come from Taiwanese press. Nvidia has not published a power figure for its Rubin rack. The AI labs, whose commitments sit behind much of the financing in Section 15, publish no audited accounts. And no operator has reported how much of its AI load it can actually curtail. Where these gaps touch the argument, this report uses ranges and says so, and none of them carries a conclusion on its own. The list that follows the evidence register records each one.

Evidence register

Every source cited in the text, in order of first use. Tier 1 is an issuer filing or release, a regulator or grid-operator document, a standards body, or a company's own technical documentation. Tier 2 is reputable trade, research-institution or peer-reviewed work. Tier 3 is a single-source estimate or a second-hand account, used only where flagged in the text. Dates are publication or filing dates, or the period a filing covers. Evidence checked through 30 September 2026.

Not verified / not load-bearing

The items below could not be confirmed from a primary document, or rest on an estimate. Each is used in the text only with that caveat, and none carries a conclusion on its own.

• Amazon's 2026 capital spending of about $220 billion. The figure was given on Amazon's July 2026 earnings call and is cited from a secondary report of the call; Amazon's filings do not state capex guidance. The $530–555 billion sum for three companies in the opening takeaways uses it. Without it, the direction of the spending is unchanged.
• TSMC's wafer and CoWoS capacity in wafers a month. TSMC does not publish these figures. The numbers in Sections 6, 7 and 22 come from Taiwanese press as summarised by TrendForce and are used for scale only.
• The AI share of TSMC's output. TSMC reports that high-performance computing was 58% of 2025 revenue. Secondary accounts of its January 2026 call give the AI-accelerator share differently (15% and "high-teens" percent), and no TSMC document read for this report states it.
• Rubin power per accelerator and per rack. Nvidia has not published these. The 1.8–2.3 kW per accelerator and 190–230 kW per rack used in Figure 11 come from a server maker's planning guide; the Rubin point in that figure is an estimate.
• Rubin memory and NVLink bandwidth. Nvidia's two sources differ. Its January 2026 technical blog gives up to 22 TB/s of memory bandwidth per accelerator and 260 TB/s of NVLink bandwidth per rack; its current product page gives 1,400 TB/s per rack (about 19.4 TB/s per accelerator) and 216 TB/s. At the product-page figure, the decode example in Section 3 takes about 15 ms rather than 13 ms. No conclusion changes.
• Jensen Huang's per-gigawatt figures. These come from a second-hand account that gives no date or venue. Nvidia's own call transcript and McKinsey's published split of AI data-center spending could not be retrieved (search results only; page requests timed out). Figure 19 builds its total from parts for this reason, and Section 17 checks it against Nvidia's filed revenue.
• Cushman & Wakefield's 2026 cost guide. Read through CRE Daily's and IREI's summaries, not the guide itself.
• FERC's December 2025 co-location order. Read as described in FERC's June 2026 order to PJM and in a law firm's summary; ferc.gov was not retrieved. The June 2026 order itself was read in full from the copy PJM posts.
• Ohio's tariff terms. The 85%, 12-year terms come from POWER magazine's report of the PUCO order. The PUCO's own minutes confirm the 5–0 approval of Case 24-508-EL-ATA on 9 July 2025; the full opinion was not retrieved.
• MISO's auction results. Read through the American Public Power Association's report; MISO's own results posting returned "access denied".
• The Duke flexible-load study. Read through Utility Dive's report; the study PDF was blocked by a bot filter.
• JEDEC's HBM4 announcement. Read as republished; jedec.org returned HTTP 403.
• Foxconn's margins. From a trade publication's republication of Foxconn's release, not from Foxconn's own filing.
• Quanta and Wistron. Named in Sections 2 and 10 by their role as assemblers. Their own filings, which are in Chinese on the Taiwan exchange, were not opened, and the report draws no conclusion about either company.
• Small modular reactors. No dated commercial milestone within five years was verified, so they are left out of Figure 27.
• The AI labs' finances. OpenAI's and Anthropic's accounts are private. Their commitments appear only in their suppliers' filings.
• How durable take-or-pay memory contracts are. They have not been tested in a downturn, and the contracts themselves are not public.