Robotaxi Economics: The Fare Without the Driver

ThesisThe fare without the driver

An average ride-hail fare splits roughly 65/35, and the driver earns the larger share. The robotaxi is a bid to capture it. That share paid for the car, the insurance, the cleaning, and the Friday-night shift. Take the driver out and every one of those costs survives. They move into depots, insurance files, capital budgets and regulatory standing, and the money ends up with whoever can carry all four at once.

A Lyft ride in the second quarter of 2026 cost the passenger $20.98 on average. Lyft kept $7.03. The remaining $13.95 — 66.5% of the fare — went to the driver, plus taxes, tolls and fees. Those are not estimates; they fall straight out of the gross bookings, rides and revenue Lyft reported to the SEC on 7 August 2026¹. Of the $7.03 Lyft kept, $0.68 reached adjusted EBITDA — the company's own measure of operating profit, struck before interest, tax, depreciation and share-based pay. The rest went on insurance, operations, engineering and marketing.

Across Uber's and Lyft's disclosed 2025 mobility volumes that $13.95 amounts to roughly $80 billion a year worldwide, before the taxi trade and before any private-car mileage a cheap driverless fare might eventually take. Whoever cracks autonomy inherits it: that is the consensus one-liner, and it is half right. It reads the wage pool — the share of every fare that has always left with the person driving — as money sitting on a table. It was a payment doing several jobs at once.

One variable payment, made only when a passenger paid, bought all of it: the vehicle, its financing, its insurance deductible in practice if not in law, its cleaning, its charging, its parking, its dispatch to a customer, and its willingness to work Friday at 6pm and not Tuesday at 3am. Remove the human and each of those jobs becomes a line item that somebody has to own, staff, capitalise and insure, around the clock, whether or not anyone is riding.

At Waymo's observed utilisation — the share of each day a vehicle spends earning rather than parked — the vehicle costs about $54 a day to own, the remote-assistance desk about $12, the electricity about $9. Against roughly $469 a day of gross fares, that is 16 cents on the dollar. The other 84 cents has to cover depots, cleaning, maintenance, staging, insurance, mapping, city-by-city validation, whatever commission an aggregator takes, and margin. Almost none of it is publicly disclosed. On these numbers, the hardware argument the industry has run for a decade is about the small end of the cost stack.

The money lands with the operator that holds four things at once: an audited safety record no entrant can reproduce, a per-city operating machine, and the actuarial file and regulatory standing those two buy. The sensor maker, the chip maker and the owner of the model weights are all outside it, and so, in the medium term, is the app. On the evidence available in August 2026, exactly one company holds all four. The second-best position belongs to the aggregator that has spent two years making sure no single stack can hold the rider hostage.

Executive takeaways

1. 1The wage pool, not the fare, is the unit of analysis — and it does not survive as a pool. Of a $20.98 Lyft ride, $13.95 (66.5%) leaves with the driver; removing the driver frees roughly 50 points of fare, of which vehicle capital, remote assistance and energy consume only about 16. The residual 84% lands in undisclosed fleet operations. Known Known
2. 2The audited safety record is the only asset in this industry that compounds and cannot be bought. Waymo has 220.6m rider-only miles — miles carrying passengers with nobody in the driver's seat — through March 2026, 80% of the 275m failure-free miles RAND calculated are needed to demonstrate a better-than-human fatality rate at 95% confidence. No other US operator discloses a driverless denominator — the audited mileage any crash rate has to be divided by — within an order of magnitude, and 96.1% of Tesla's 1,703 Level 2 crash reports to NHTSA (Level 2 is driver assistance that still requires a human watching the road; Waymo's driverless system is Level 4) carry an injury severity of “Unknown” against 1.3% of Waymo's driverless reports. Known Known · Advantaged: Alphabet, High confidence
3. 3Labour did move indoors, and it got 25 times cheaper. At Waymo's disclosed ratio of roughly one on-duty remote agent per 43 vehicles, remote assistance — staff at screens who answer questions the vehicle asks, rather than driving it — costs about 2.6% of gross fares against 66.5% for a human driver, the single largest reason the economics work at all. Estimated
4. 4The regulatory unit is the state, and Washington just confirmed it. NHTSA withdrew the proposed national AV STEP oversight programme on 26 June 2026. California has issued exactly three deployment permits — the authority to run driverless vehicles commercially, as against merely testing them — and six driverless testing permits. A federal exemption, the only route open to a vehicle built without a steering wheel, is capped at 2,500 vehicles a year — below Waymo's current import run-rate. Known Known
5. 5The robotaxi is a base-load asset in a peaking business. A fleet costs the same at 3am as at 6pm; urban ride demand arrives in bursts. A Waymo averages 0.99 rides per hour across the full 24; a human driver online eight hours at the same hourly rate would complete 7.9. The "busier than 99% of drivers" claim measures availability, not productivity — and it means the human network keeps the profitable peak longest. Estimated · Bounds the downside for Uber and Lyft, though Waymo's July 2026 notice ending Uber exclusivity shows the strongest operator no longer needs the partnership
6. 6The supplier layer has already been repriced, and it is integrating forward in response. Mobileye wrote off $3,788m of goodwill — the premium above hard assets that past acquisitions left sitting on its balance sheet — in Q1 2026 on flat revenue, then announced in June 2026 that it would build and operate its own robotaxi fleet. Aurora reported $2.0m of revenue against a $270.0m quarterly loss. Known Known · Exposed: pure-play stacks, High confidence
7. 7The US and China are running the same experiment with opposite inputs, and the fare is winning. A Waymo vehicle grosses roughly $469 a day; Pony.ai booked about $67 of robotaxi services revenue per vehicle per day in Q2 2026 at comparable ride throughput (rides completed per vehicle per day). A 3x vehicle-cost advantage is losing to a 7x revenue disadvantage. Estimated

StakesThe system, quantified

Ride-hail already moves real money. Most published estimates of this market start from a forecast instead.

Sources: Uber 10-K FY2025²; Lyft FY2025 results, 10 Feb 2026³; TechCrunch, 27 Mar 2026⁴. Derived figures shown in the sections below.

Uber's mobility gross bookings — the total dollar value of rides including taxes, tolls and fees — ran $97.5 billion in 2025 (the sum of the four quarters Uber discloses in its 10-K²: $21,182m, $23,762m, $25,111m and $27,442m). Against that, Uber recognised $29,670m of mobility revenue, a take rate of 30.4%. The difference, $67.8 billion, is what left the platform for drivers, taxes and tolls. Lyft's 2025 gross bookings were $18.5 billion on revenue of $6.3 billion, a 34.1% take; $12.2 billion left the platform. Together, roughly $80 billion a year flowed out of these two platforms to drivers and public authorities.

How much of that is American is not disclosed. Uber reports revenue by region — $7,609m from the United States and Canada in Q2 2026 out of $14,191m total — but does not break out US mobility gross bookings, and Lyft now consolidates European operations through FreeNow. Anyone who quotes a precise US ride-hail driver-earnings figure is estimating. The US share of that $80 billion is the largest single national slice, and it is not publicly separable.

The wage-and-salary taxi trade beside it is smaller and better measured. The Bureau of Labor Statistics' May 2025 Occupational Employment and Wage Statistics count 41,050 taxi drivers at a $49,930 annual mean and 248,530 shuttle drivers and chauffeurs at $39,880 — a combined payroll of about $12.0 billion (BLS public API, series OEUN000000000000053305401/04 and 053305301/04⁵). That survey counts employees only. Most of the ride-hail workforce is independent contractors, and the gap between $12 billion of measured payroll and $80 billion of platform outflow is a reminder that the labour being displaced here is largely invisible to the labour statistics.

Waymo was delivering 500,000 paid autonomous rides a week across ten US cities as of 27 March 2026. Lyft alone completed 262.4 million rides in the second quarter of 2026 — 20.2 million a week. The driverless fleet is running at about 2.5% of one of the two US ride-hail networks. Everything in this brief concerns an industry that has proven it works and has barely begun to matter.

01The fare, and who keeps it

New York City publishes the price of a driver. Under Taxi & Limousine Commission rules, a high-volume for-hire driver must be paid at least $1.283 per mile and $0.681 per minute for a non-wheelchair-accessible trip, rates that rose 3.36% on 1 March 2026 in line with the prior year's CPI-W (TLC Industry Notice #26-03, 24 February 2026⁶). At a 12 mph average city speed that is $4.69 a mile before the driver pays for the car, the fuel, the insurance and the depreciation.

Set that beside a robotaxi grossing roughly $2.46 a mile at Waymo's observed miles and utilisation. The regulated minimum paid to a human in New York for the driving alone exceeds the total revenue per mile a driverless vehicle collects in San Francisco. The robotaxi is not cheap; the human wage is simply the largest cost in urban passenger transport, and removing it moves the arithmetic more than any other lever available.

Uber's mobility take rate looks like it collapsed in 2026 — 30.7% in Q2 2025 against 25.4% in Q2 2026 — but the fall is an accounting artefact. Uber changed its business model in the United Kingdom, which moved revenue to a net basis and reduced reported mobility revenue by $1.1 billion in the quarter (Uber Form 10-Q, 5 August 2026⁷). Add it back and the take rate is 29.2%. The same filing records an $813 million decrease in driver payments and incentives on gross bookings that grew 22% — the same accounting change showing up in a second line, and not a wage cut. Most of the analytical work in this industry is making the numbers comparable.

Through all of it, roughly two-thirds of a US ride-hail fare leaves the platform. The platform's own margin on the transaction is thin — Lyft converted 3.2% of gross bookings to adjusted EBITDA in Q2 2026, Uber 4.9% across all segments. Neither company earns its money by taking a large slice of each ride. Both earn what they earn by running enormous volume across a fixed cost base.

02Variable cost, fixed cost, and the airline problem

A driver costs nothing when nobody is riding. That single fact explains ride-hail's balance-sheet elegance and the robotaxi's balance-sheet difficulty.

Uber and Lyft pay drivers per completed trip. When demand falls, the cost falls with it, immediately and completely. Drivers absorb the idle time, the vehicle depreciation, the insurance premium and the fuel, and they do it because the alternative use of their hour is often worth less. The platform's cost of goods sold is almost perfectly variable. It is one of the cleanest operating-leverageOperating leverage is how a business's profit responds when revenue moves. Costs that rise and fall with revenue keep profit steady and unspectacular; heavy fixed costs make it swing violently in both directions. structures in transport.

A robotaxi fleet inverts this. The vehicle depreciates on a Tuesday at 3am exactly as fast as at Friday 6pm. The depot lease, the charging infrastructure, the cleaning crew, the maintenance technicians, the remote-assistance desk staffed around the clock and the insurance premium all accrue continuously, while revenue arrives only when somebody rides. Utilisation therefore governs everything, and the business behaves like an airline: high fixed cost, perishable inventory, and a load factorAn airline's word for the share of available seats actually sold. It is the number that decides whether a flight whose costs are already committed makes money, and it transfers to any business that must buy its capacity before the customer arrives. that decides whether the quarter is good.

Evidence: Uber 10-Q Q2 2026⁷ (driver payments recorded in cost of revenue, variable with gross bookings); Tesla 10-Q Q2 2026⁸, which lists "vehicle cleaning and maintenance, charging, security, teleoperations and fleet management" as dedicated robotaxi infrastructure. Synthesis: The swap from variable to fixed cost is the defining structural change, and it moves the competitive question from software to capital and operating discipline.

Tesla's own filing describes what it is building alongside the robotaxi service: "dedicated infrastructure, including in relation to vehicle cleaning and maintenance, charging, security, teleoperations and fleet management." Every item on that list is a fixed cost that a human driver used to absorb for free. Tesla is describing a transport operating company in its own 10-Q.

The cheapest fleet is the one somebody else buys. Vehicles owned by the households that drive them would put depreciation, insurance and idle time back where ride-hail had them, and an operator running that fleet would carry autonomy on a variable cost base rather than a fixed one. No operator's filings describe such a fleet. Tesla's FY2025 Form 10-K⁹ puts the vehicles on Tesla's own side of the line — the Robotaxi service is “an autonomous ride-hailing platform that harnesses our technology and vehicles” — and the profit it anticipates is “fleet-based profits,” accruing to Tesla. The Q2 2026 10-Q⁸ repeats the phrase and records that Cybercab production has begun. Customer-funded fleets are discussed in public and absent from the disclosure.

Almost all of the robotaxi's cost is fixed, so its cost curve improves sharply with utilisation. Every operator's first strategic objective is therefore ride density before geography. But a fleet sized to serve the evening peak sits idle through the middle of the night, and unlike a human network it cannot send that capacity home. It is the most under-appreciated constraint in the industry, and the strongest reason to doubt the human network disappears quickly; Section 10 takes it up in full.

03The vehicle: three cost curves and a tariff wall

In August 2026 more than 500 Chinese-built robotaxis were photographed together at a Waymo integration facility in Mesa, Arizona. They had arrived at the Port of Los Angeles as stripped base vehicles — no connectivity hardware, no sensors — and would leave Mesa as Waymo Ojais, fitted with US-designed compute and lidarA sensor that measures distance by timing laser pulses off the world around it, assembling a three-dimensional picture of the scene. Unlike a camera it measures range directly instead of inferring it, and it works in the dark — at a hardware cost that has been the industry's central argument for a decade. by Waymo and the Tier 1 supplierIn the car industry, a company that sells finished systems straight to the vehicle maker rather than components to another supplier. Tier 1s do the integration work — mounting, wiring, testing — that turns parts into something a production line can fit. Magna. Bills-of-ladingThe shipping document that travels with a container: who sent what, to whom, in what quantity. US customs filings make them semi-public, which is how an outside reporter can count a company's imports without the company saying a word. data compiled by ImportGenius and reported by Forbes on 2 August 2026¹⁰ put cumulative imports of the Zeekr platform above 3,200 units since 2024, more than 2,600 of them in 2026 alone, at a run-rate of roughly 300 a month.

They arrive over a 127.5% tariff wall. A vehicle that sells for about $39,000 in China lands at roughly $88,725 in the United States, before Waymo adds an autonomy kit that the same reporting puts above $10,000. Call it $99,000 all-in. The Chinese domestic cost base runs the other way: for a comparable purpose-built platform Pony.ai expects to be below RMB 230,000 by 2027, around US$33,000, which the company points out is less than the starting price of a locally produced Tesla Model 3 in China (Pony.ai, 24 April 2026¹¹).

The tariff is a cash cost. Waiting for a domestically built purpose-designed platform is a time cost, and in this industry time compounds into validated miles, which compound into insurance pricing and regulatory standing. Waymo has now begun driving the Hyundai IONIQ 5 with a specialist aboard (Waymo, 8 July 2026¹²), built at Hyundai's Georgia plant and carrying no China tariff at all. The Zeekr fleet is a bridge, bought at a known price, across a gap the domestic supply chain could not yet close.

The retrofit curve — a production passenger vehicle fitted with an autonomy kit — costs more per unit and buys more freedom: it complies with every Federal Motor Vehicle Safety StandardThe federal rulebook every car sold in the United States must meet — brakes, mirrors, lights, airbags, seat belts, controls. Much of it assumes a human driver in a forward-facing seat, which is why a vehicle designed without a steering wheel cannot simply comply with it., so it can be built in unlimited numbers. The purpose-built curve (Zoox's cabin with no steering wheel, Tesla's Cybercab) is cheaper per seat-mile in principle but is legally capped, as Section 9 sets out. The Chinese curve is the cheapest of the three and, for the US market, does not exist except as a base vehicle stripped of everything that made it Chinese.

Which curve is lowest matters less than what happens to margin if the low one clears the wall. If US-built purpose-designed platforms reach $40,000 all-in — plausible on Hyundai and Lucid volumes toward the end of the decade — vehicle capital falls from about $54 a day per vehicle to roughly $22. On $469 of daily gross fares that is a swing of seven points of margin — worth having, and still not where the business is won.

04The stack, and the cost of proof

In 2016 two RAND researchers, Nidhi Kalra and Susan Paddock, asked a question nobody in the field wanted answered: how many miles would an autonomous vehicle have to drive without a fatal crash before you could say, statistically, that it was safer than a human? Their answer¹³ was 275 million failure-free miles, to demonstrate a fatality rate of 1.09 per 100 million miles at 95% confidence. At 100 test vehicles running 24 hours a day at 25 mph, that is 12.5 years. Their conclusion was blunt: developers "cannot drive their way to safety."

A decade on, that paper prices the thing everybody in the industry has been buying. Validation is a capital expenditure denominated in miles, and it is the reason this business has a moat at all.

Camera-only versus sensor fusionAn approach that combines several kinds of sensor — cameras, radar, laser — and cross-checks them against one another, so a failure in one measurement is caught by another running on entirely different physics. The alternative is to rely on cameras alone and solve the harder inference problem in software. is usually staged as a debate about physics — can a photodiode see what a laser sees? The economics of proof decide it, and the two camps are optimising different variables.

Sensor fusion buys evidentiary efficiency. Redundant, physically independent measurement channels mean fewer failure modes have to be excluded empirically, so fewer miles are required per unit of demonstrated safety. Those miles can also be driven driverless, and driverless miles are the only kind regulators and insurers count. The cost is a bill of materialsThe parts list for a product, priced. Engineers use it as shorthand for what a thing costs to build before labour, overhead and margin. above $10,000 and a per-city mapping and validation obligation.

Camera-only buys mileage cheaply. A Tesla with FSD costs the customer, not Tesla, and the fleet generates billions of supervised miles a year at near-zero marginal cost to the manufacturer. The wager is that supervised consumer miles can substitute for driverless validation miles.

NHTSA's Office of Defects Investigation opened Engineering AnalysisThe second and more serious stage of a US defect investigation. A Preliminary Evaluation asks whether there may be a problem; an Engineering Analysis establishes its scope and cause, and it is the step from which a recall demand can follow. EA26002¹⁵ on 18 March 2026 covering an estimated 3,203,754 Tesla vehicles, examining the failure of FSD's degradation-detection system to recognise reduced roadway visibility and warn the driver. Nine crashes, one fatal, two injury. The economically important passage is the one about Tesla's own records:

Tesla also described internal data and labeling limitations that prevented a uniform identification and analysis of crash events with the subject system engaged. ODI believes this limitation could have led to under-reporting of subject crashes.

NHTSA ODI, Engineering Analysis EA26002, 18 March 2026

If a manufacturer cannot reliably identify which of its own crashes occurred with the system engaged, its supervised mileage stops working as a denominator at all. The problem is the bookkeeping, not the size of the number. A safety case is a numerator over a denominator, audited: the argument an operator documents for a regulator and an insurer that its vehicle is safe enough to be let out. Here is the exposure we accumulated, here is every bad outcome inside it, here is how we know the two match. So far the camera-only thesis has failed on bookkeeping rather than perception, and the bookkeeping has not been fixed.

NHTSA publishes a second Standing General OrderA standing instruction from the federal road-safety regulator requiring companies to report certain crashes involving automated systems, on a fixed schedule and in a fixed format, whether or not anyone asks. The resulting public file is the closest thing this industry has to a common ledger. file covering Level 2 driver-assistance systems, and its injury-severity field tests that finding directly. Of Tesla's 1,703 unique Level 2 reports through July 2026, 1,636 — 96.1% — record the highest injury severity as "Unknown." The comparable share is 68.3% for General Motors, 35.0% for Ford and zero for Subaru. In the driverless file, Waymo's unknown share across 1,117 reports is 1.3%.

A second, narrower NHTSA investigation runs alongside it. PE25012¹⁸, opened 7 October 2025, covers 2,882,566 vehicles and 58 incidents — 14 crashes, 23 injuries — in which FSD allegedly ran red lights or changed lanes into oncoming traffic. Both files are open as of this brief's cutoff.

05What the audited data actually says

Every operator of an automated driving system on US public roads must report certain crashes to NHTSA under Standing General Order 2021-01. That file is the only common public record the industry has. As published on 27 August 2026, it contains 1,324 unique incident reports covering April 2025 through July 2026, and its distribution is not what the headlines suggest.

Evidence: NHTSA SGO 2021-01 ADS incident report file¹⁷, last modified 27 August 2026. Method: 1,365 rows deduplicated to 1,324 unique Report IDs by retaining the highest Report Version; severity from the "Highest Injury Severity Alleged" field; rows with "No Injured Reported" or "Unknown" omitted from the injury columns, so rows do not sum to reports. Synthesis: Report counts track exposure, not fault or safety; Waymo files most reports because it drives the most driverless miles by a wide margin.

Of Waymo's 1,117 reports, 977 — 87.5% — were property damage with no injury reported. Two were serious or fatal, 0.18% of the total. Those two are set out below.

The fatality occurred in Tempe, Arizona at 1:21am in September 2025. A Waymo travelling in the rightmost lane slowed with its right indicator on, yielding to a pedestrian before turning into a parking lot. A motorcycle struck it from behind; the rider fell into the adjacent lane and was then hit by a passing car that did not stop. The Waymo was carrying no passengers. The June 2026 serious-injury case in Scottsdale is structurally identical in the opposite direction: an SUV crossed two lanes at speed and struck a Waymo stopped at a red light, and the SUV's driver was hospitalised.

Both of the worst outcomes in the audited record involved a third party striking a stationary or correctly signalling autonomous vehicle. Read that as a caution about what the SGO file measures — involvement, not causation — rather than as a defence of the technology. It also carries a commercial implication, taken up in Section 8. The liability tail of a driverless fleet may be dominated by other people's driving. An insurer already knows how to price that.

In July 2026 the field received its strongest external validation. The Insurance Institute for Highway Safety published a study on 23 July 2026¹⁹ comparing about 50 million Waymo driverless miles from 2021 to 2024 against roughly 222 billion human miles in the same four cities. Waymo's crash-involvement rate was 68% lower overall and its injury-crash rate 81% lower.

That does not settle the safety question, and the study's own numbers show why. The reduction was 76% in Phoenix, 71% in Los Angeles, 35% in San Francisco — and 4% higher in Austin, the newest and smallest sample. IIHS also had to discard roughly a quarter of reported crashes in cleaning, found that only 22% of 736 public-road crashes met police-reportable thresholds, and cautioned that around half of human crashes and a third of human injury crashes go unreported at all. Its lead author's conclusion was that "the present data collection system isn't good enough to allow continuous monitoring of a large-scale expansion."

The spread across cities matters more than the average. An operational design domain — ODD, in the industry's shorthand — is the envelope a driverless system is built and validated to handle: these streets, these speeds, this weather, these hours of the day. Evidence accumulated inside one envelope says little about a different one. Safety performance is therefore a property of an operational design domain, not of a company, and it does not transfer for free when the company enters a new city. Every expansion resets part of the evidence base. The safety asset therefore compounds slowly, an entrant cannot short-cut it, and a rapid multi-city expansion carries a risk the aggregate statistic hides.

Waymo's own published figures, drawn from the same SGO reporting and benchmarked in peer-reviewed work by Kusano and Scanlon in Traffic Injury Prevention, are stronger than the IIHS numbers: at 220.6 million rider-only miles through March 2026 the company claims 94% fewer serious-injury-or-worse crashes, 82% fewer any-injury crashes and 93% fewer pedestrian injuries than a human benchmark in the same areas (Waymo Safety Impact¹⁴). The direction agrees with IIHS; the magnitude does not, and the gap is the benchmark construction. An operator's own benchmark should be read as a well-documented advocacy figure; the insurer's is the one that will price the policy.

06Remote assistance: the labour that moved indoors

In July 2025 a Tesla robotaxi in Austin stopped and would not proceed. The safety monitor in the car requested support. A teleoperator took control, accelerated the vehicle, turned it left, drove it up a curb and into a metal fence. Someone was hurt badly enough to be taken to hospital. In January 2026 a second Tesla teleoperator took control of a stopped vehicle and drove it into a construction barricade at 9 mph.

Both narratives are in Tesla's own filings to NHTSA. Of the 24 Tesla ADS crash reports in the current SGO file, 18 state that a safety monitor was present in the vehicle, and two — 8.3% — were caused by the remote operator rather than by the automated driving system. Two crashes is a small sample and should not be over-read; it is also the only publicly documented instance of remote-assistance error causing crashes in the audited record. The human in the loop is a source of risk as well as a mitigant.

How often any of this happens is the load-bearing undisclosed number in the whole industry, and in February 2026 someone finally asked. Senator Edward Markey wrote to seven operators — Aurora, May Mobility, Motional, Nuro, Tesla, Waymo and Zoox — with fourteen questions about their remote-assistance operations. The resulting report²⁰ is the first systematic public account of the role, and its first finding is the one that matters:

Every AV company refused to disclose how frequently their RAOs intervene to help their self-driving cars.

Office of Senator Edward J. Markey, Remote Back Seat Operators, 2026

Seven companies gave the same answer. When an entire industry declines the same question, the number is either embarrassing or competitively valuable, and both readings stay open. One-way latencyThe delay between something happening and someone elsewhere seeing it. In remote operation it sets how stale the picture on the screen is at the moment a person acts on it, which is why a few hundred milliseconds is the difference between advice and hazard. between vehicle and operator ranged from 100 milliseconds to at least 300 across the seven. Waymo's overseas operators run at 250ms against 150ms domestically — two-thirds higher, in a system where published research suggests 300ms can make remote driving dangerous. Zoox's worst case is 333ms; May Mobility's is 500ms. There is no federal standard, so each company has set its own safety threshold and none has had to defend it.

Waymo is the only operator using overseas staff, and the only one at which a substantial share of remote assistants do not hold a US driver's licence: half its remote-assistance workforce sits in the Philippines and holds local licences. Every company assured the Senate that its automated system would ignore dangerous advice. An NTSB investigation cited in the report found that a Waymo vehicle, acting on incorrect guidance from a remote assistant, passed a school bus with its stop sign extended.

Then the regulator moved. When NHTSA granted Zoox its commercial exemption on 31 July 2026, it attached a condition that had not previously existed in US AV regulation:

Zoox shall ensure that all personnel involved in safety-critical tasks (e.g., remote assistants) are located in jurisdictions within the Continental United States such that the location of remote personnel does not degrade system safety, data integrity, or regulatory accountability.

NHTSA, Grant of Temporary Exemption No. 2026-01²¹, 31 July 2026

Five months separate the Senate report from that condition. The labour that migrated from the driver's seat to a screen has just been told it may not migrate abroad, at least for vehicles running under a federal exemption. Waymo's current architecture, half offshore, would not satisfy that condition. Waymo does not currently need an exemption, because its vehicles have steering wheels — a distinction Section 9 takes up.

The same grant notice records that the Transport Workers Union of America asked NHTSA to require reporting of remote-operations headcount, the number of "floaters" who can be reassigned to remote duties, median hourly compensation, benefits and annualised turnover. NHTSA declined most of it, stating that "details like these are necessary for the agency's safety-focused operational oversight" only if the agency later deems them so. The ratio remains undisclosed by regulatory design, not merely by corporate preference.

One credible number survives in public. In February 2026 Waymo disclosed roughly 70 remote-assistance agents on duty worldwide at any moment against about 3,000 vehicles — one on-duty agent per 43 vehicles. Taken at face value, that ratio carries the whole labour argument.

Remote assistance is widely expected to be a hidden human workforce that eats the savings. At Waymo's disclosed ratio that expectation is wrong by roughly an order of magnitude, and it stays wrong at a ratio four times worse and a wage cost half again as high. Remote assistance takes a small share of the wage pool; the rest goes elsewhere.

07The hidden operating company

Somebody has to clean the car. No operator has ever published what it costs, and the one public company to describe the work in a filing did so as a risk factor.

Lyft's most recent 10-Q discloses that Flexdrive, its vehicle-fleet subsidiary, "has also begun providing autonomous vehicle fleet management services where we operate fleets of AV vehicles for AV manufacturer partners," and enumerates the obligations: "vehicle availability, maintenance, charging, cleaning, staging and depot operations" (Lyft Form 10-Q, 7 August 2026¹). That is the most precise public description of the robotaxi operating company anywhere in the SEC filing corpus, and it exists because Lyft is contractually exposed to failing at it.

Waymo's Mesa, Arizona facility is the physical form of layer four. Base vehicles arrive from China; Waymo and Magna fit US-designed compute, connectivity and sensors; finished Ojais leave for eleven cities. It is a factory, staffed and capitalised, that exists because the vehicle a robotaxi needs cannot be bought finished. Tesla's 10-Q describes building the same thing, and Lyft is selling it as a service.

Anything Lyft will sell you as a service has stopped being a moat. Fleet operations is hard, capital-hungry and real-estate constrained, and still only a margin. Operating excellence here shows up as two or three points of cost. That matters enormously to an operator running at scale, and not at all to whether a competitor can enter.

08Underwriting a mile with nobody in it

In February 2025 Lyft's captive insurer paid Riverstone $85.1 million to take on a legacy portfolio of auto liability policies written between October 2020 and September 2022, with an aggregate limit of $120.5 million (Lyft 10-Q, 7 August 2026¹). That is seventy-one cents of premium to cap a dollar of tail — the market's price for the uncertainty in three years of human-driven ride-hail claims, and the cleanest available measure of what the human driver's liability profile costs.

Uber carried $3,758 million of short-term and $9,528 million of long-term insurance reservesMoney set aside on the balance sheet for claims already incurred but not yet paid — including accidents that have happened and not been reported, and lawsuits that will take years to settle. It is an estimate, revised as the claims come in. at 30 June 2026 — $13.3 billion, or 35.5% of its total liabilities — backed by $9,486 million of restricted investments. Lyft's insurance reserves of $2,308 million are 25.3% of its entire balance sheet. Read the liability side and neither company looks like a marketplace; each is a captive insurer with an app attached.

Evidence: Uber 10-Q⁷ and Lyft 10-Q¹, both for the quarter ended 30 June 2026. Method: Uber $3,758m + $9,528m = $13,286m ÷ $37,402m total liabilities = 35.5%. Lyft $2,307.7m ÷ $9,107.5m total assets = 25.3%. Riverstone loss portfolio transferA deal in which one insurer pays another to take over a closed book of old claims, together with the uncertainty about what they will finally cost. The premium paid, set against the cover bought, is a market price for that uncertainty.: $85.1m premium ÷ $120.5m limit = 70.6%. Synthesis: The incumbent aggregators already own the actuarial infrastructure a driverless fleet needs — a genuinely underweighted asset in their favour.

Uber's FY2025 10-K attributes an $851 million increase in insurance expense to "an increase in insurance rate per mile and miles driven." Per mile. The cost of insuring a human driver is rising faster than the miles they drive, and Uber has put that in writing.

Against that, the driverless mile is being priced on evidence. Waymo's collaboration with Swiss Re examined liability claims across 25.3 million rider-only miles and found nine property-damage and two bodily-injury claims, against human baselines of 78 and 26 for the same distance — reductions of 88% and 92% against the overall driving population, and 86% and 90% against 2018–2021 vehicles with modern driver assistance (Waymo, 19 December 2024²⁵). The human baseline came from Swiss Re's file of over 500,000 claims across more than 200 billion miles of exposure.

The data now sits with the reinsurer. Swiss Re built a rating basis for a risk class that did not previously exist, using one operator's miles. Every subsequent entrant will be quoted against a curve derived from Waymo's experience, at a credibility loading reflecting that it is not Waymo. Insurance is priced from experience, and experience means a file: enough miles, enough claims and enough settled years for the losses to form a curve an underwriter will stand behind. A credibility loading is what an insurer charges for the part of that file a newcomer has not built yet. The mile leader's actuarial file is a cost advantage that compounds with every additional mile and cannot be bought or reverse-engineered. It is the clearest example in this industry of an asset that gets more valuable precisely because the technology gets cheaper and more competitors arrive.

The Uber–Nuro–Lucid arrangement is, underneath, a decision about who carries the liability. Uber licenses Nuro's driving system and, with third-party fleet partners, owns and operates the vehicles. That places the operating risk — and the insurable event — with the party that already has $13.3 billion of reserves and thirty years of claims machinery, rather than with a venture-funded software company. It is a rational allocation, and it shows Uber has thought harder about the liability layer than about the software layer.

The scenario that haunts this field is a single at-fault fatality, on camera, with a child. The audited record says the first fatality has already happened — Tempe, September 2025 — and it repriced nothing, because responsibility fell elsewhere. What repriced the field was Cruise, where the vehicle's conduct was survivable and the company's conduct was not — Section 15 takes that apart. Insurers price the crashes; regulators price the candour that follows, and it is the second of those that has ended a company.

09The permit map is the market map

On 26 June 2026 NHTSA withdrew the rulemakingThe formal process by which a US agency makes law: publish a proposal, take public comment, then issue a final rule — or withdraw. It runs for years, it can be abandoned at any stage, and nothing binds anyone until the final rule appears. that would have created a national framework for overseeing automated vehicles. AV STEP — the Automated Driving System-Equipped Vehicle Safety, Transparency, and Evaluation Program, proposed in January 2025 — would have offered manufacturers and fleet operators a voluntary federal review in exchange for transparency. The agency killed it²⁶ "based on its reevaluation of the proposal, consideration of the comments received, and subsequent progress on other ADS initiatives."

There is now no federal licence to operate a robotaxi in the United States, and there is not going to be one soon. The unit of market access is the state, and a city permit is not a market.

California requires two separate authorities: a DMV permit to operate the vehicle and a California Public Utilities Commission authorisation to carry passengers for money. The CPUC's quarterly reporting²⁷ lists exactly one carrier in the Deployment programme — Waymo — with Aurora, Tensor, WeRide and Zoox in the Pilot programme, which does not permit charging. The most recent reporting period is January to March 2026, and both Waymo and Zoox have claimed confidentiality over portions of their submissions.

Texas ran the opposite experiment and has now partially closed it. Before September 2025 an operator could deploy in Texas with no state authorisation at all, which is why Austin became the industry's proving ground. SB 2807 changed that: the law took effect 1 September 2025, final rules on 27 February 2026, and authorisation requirements became enforceable on 28 May 2026. Operators must now certify traffic-law compliance and recording equipment, prove liability coverage, and file an emergency-services interaction plan with the Department of Public Safety. There is no fee and authorisations do not expire (Texas DMV, Automated Vehicles Regulatory Program²⁸). Texas remains the most permissive large state, but it is no longer unregulated, and the direction of travel is one-way.

A vehicle with no steering wheel, no mirrors and no manual brake cannot comply with the Federal Motor Vehicle Safety Standards, so it needs an exemption under 49 U.S.C. 30113 — and the statute caps that exemption at 2,500 vehicles introduced into commerce in any twelve-month period. That ceiling shapes the industry and is routinely overlooked. NHTSA granted Zoox exactly that on 31 July 2026: Temporary Exemption No. 2026-01, effective to 31 July 2028, covering portions of eight FMVSS, with a new "Operational Authorization" attached that NHTSA can modify at will.

The cap sits below the run-rate. Waymo imported more than 2,600 Zeekr platforms in 2026 alone, roughly 300 a month, or about 3,600 a year. A purpose-built no-controls vehicle could not legally be deployed at that rate. Waymo's fleet is uncapped precisely because its vehicles keep their steering wheels and self-certify to the standards like any other car. Self-certification is the foundation of the American arrangement. No agency inspects and approves a model before it goes on sale; the manufacturer certifies for itself that the vehicle meets every applicable standard, and the regulator polices that claim afterwards, through investigations and recalls. So a vehicle that can meet the standards faces no volume limit at all, while one that cannot — because the standards assume a driver, a wheel and a mirror — has to ask permission, and permission arrives with a ceiling. Under current law the retrofit — which looks like an expensive interim compromise — is the only architecture that scales.

Nevada shows a third model, and its paperwork is public. The state separates vehicle certification, handled by the DMV under NRS Chapter 482A on a self-certification basis, from the commercial authority to carry paying passengers, handled by the Nevada Transportation Authority under NAC Chapter 706B as an "autonomous vehicle network company" permit. On 20–21 August 2026 the Authority approved AVNC applications from Tesla Robotaxi, Waymo and Aviari Services — an Uber subsidiary — with twelve-month fleet ceilings of 5,000, 1,000 and 1,000 vehicles respectively across Clark County including Harry Reid International Airport, subject to separate airport authorisation. The orders provide that "the permits will be issued upon verification by NTA staff that the respective applicants have satisfied the conditions established by the Authority" (Nevada Department of Business and Industry, 21 August 2026²⁹). Zoox already held the first such permit and launched paid service on 10 August with 64 vehicles.

An approved ceiling is not an operative permit, and the difference is worth several thousand vehicles. The interim order and permit actually issued to Tesla Robotaxi LLC — AVNC Permit 002, Docket 26-05015, dated 27 July 2026 — restricts operations to "a maximum fleet of ten (10) fully autonomous vehicles," to the Las Vegas Strip corridor within Clark County, and prohibits passenger service "on Nevada roadways with posted speed limits exceeding forty-five (45) miles per hour" and pickups within a quarter-mile of Harry Reid International. It also requires that "Operations shall include appropriate human supervision in accordance with applicable Society of Automotive Engineers (SAE) standards," and that any accident, security incident, system failure or communication interruption be reported within five working days (Nevada Transportation Authority, 27 July 2026³⁰).

Ten vehicles on the Strip, under human supervision, off the airport and off the freeway: that is what a state regulator with no ideological objection to autonomy — Nevada has permitted every level of automation on public roads since 2011 — was willing to grant in mid-2026 to the company with the largest supervised-mileage base on earth. Arizona, by contrast, requires only a law-enforcement interaction plan filed with ADOT and DPS plus a written self-certification of FMVSS compliance, minimal-risk-conditionWhat an automated vehicle is required to do when it can no longer safely continue: pull over, stop, put on hazard lights and wait for help, rather than hand a confused situation back to someone who was not driving. capability and traffic-law compliance, with ADOT empowered to "immediately issue a cease and desist letter" if the filings are absent (Arizona DOT, A.R.S. § 28-9702³¹). Georgia is looser still: O.C.G.A. § 40-8-11 allows a fully autonomous vehicle to operate without a human driver if it is FMVSS-compliant, can achieve a minimal-risk condition, is insured, and is registered, and the statute bars rules that would narrow that authority (Georgia Code § 40-8-11³²). Same country, same technology, four different answers.

Evidence: CA DMV permit holders²²; CPUC quarterly reporting²⁷; TxDMV AV programme²⁸; ADOT³¹; Georgia Code § 40-8-11³²; NTA AVNC Permit 002³⁰; Nevada B&I, 21 Aug 2026²⁹; Las Vegas Review-Journal, 21 Aug 2026³³ (Zoox launch fleet of 64). Method: Service levels are as disclosed by regulators and operators; Texas city coverage for Tesla is inferred from the cities appearing in its NHTSA SGO filings, since Tesla publishes no service map. Synthesis: No two of the four states impose the same gate. The permissive-jurisdiction map is the competitive terrain itself, and the operator with the most audited miles clears the highest gates first.

The federal standard itself is now moving: on the same day it withdrew AV STEP, NHTSA proposed modernising FMVSS No. 135³⁴, the light-vehicle braking standard, to "distinguish how regulations apply to vehicles with and without manually operated driving controls" while keeping stopping-distance requirements for all. On 31 July 2026 it published interim guidance on commercial deployment exemptions³⁵ for comment. Read together, these say the agency intends to make it possible to build a no-controls vehicle that self-certifies instead of needing an exemption. That would remove the 2,500 cap entirely.

This rulemaking is the highest-leverage regulatory event in this industry, and it is getting very little attention. If the FMVSS modernisation completes, the purpose-built cost curve is unlocked at unlimited volume and every calculation in Section 3 changes in favour of Tesla's Cybercab, Zoox's cabin, and whatever Chinese-designed platform eventually clears customs. If it stalls, the retrofit incumbent keeps a legal monopoly on scale. The rulemaking is in comment as of this brief's cutoff and no final rule has issued.

10The shape of a city's day

Uber's CEO told investors in 2025 that the average Waymo in Austin was busier than 99% of Uber's human drivers, measured by completed trips per day. The claim is true, and it compares two things that are not comparable.

Waymo runs about 500,000 paid rides a week across roughly 3,000 vehicles: 23.8 rides per vehicle per day, or 0.99 rides per hour of a 24-hour day. A human driver online for eight hours, completing rides at exactly the same hourly rate, would finish 7.9 in a day. Most Uber drivers are part-time; the median works far fewer hours than a machine that never sleeps. The robotaxi's advantage in that comparison is availability, not productivity.

Urban ride demand is not flat and never has been. It concentrates in the weekday commute, in Friday and Saturday nights, in weather events, and at airports — Uber generated 15% of its 2025 mobility gross bookings from trips that started or ended at an airport, a concentration it discloses as a risk. A human network handles this by having 30-hour-a-week drivers log on for the peak and log off afterwards; surge pricing is the mechanism that recruits them and it costs the platform nothing when demand recedes.

A fleet cannot do that: size it for the 6pm peak and it idles at 3am; size it for the average and it turns riders away exactly when fares are highest. Because almost all of its cost is fixed, the robotaxi's economics improve with base-load demand: night shifts, airport runs, hospital trips, the flat hours a human driver does not want. They improve least at the peak, where an extra vehicle is needed for two hours and paid for over twenty-four.

Base load is the demand that is always there, served by an asset that runs continuously; the peak is the spike that has to be met by capacity standing idle the rest of the time. Power generation coined both words and both transfer intact. Peak concurrency — how many vehicles a city needs on the road at the same moment on a Friday evening — is what sets how large a fleet must be, and every vehicle bought for that hour is paid for through all the others.

The robotaxi is structurally strongest in the hours human drivers least want to work and structurally weakest in the hours they most want to work. The two networks are more complementary than substitutable at current fleet sizes. This is why Waymo's expansion has run to airports — SFO from January 2026, San José Mineta, Phoenix Sky Harbor — and to freeways, which opened to public riders in the Bay Area, Phoenix and Los Angeles on 12 November 2025³⁶. Longer trips at higher speeds mean more revenue per vehicle-hour — the only variable that fixes a fixed-cost business.

And it bounds the aggregators' downside. If the robotaxi takes the base load and leaves the peak, Uber and Lyft lose their least profitable hours first and keep their most profitable ones longest. The erosion is slower and shallower than the substitution narrative assumes.

Price is converging, slowly: Obi, a firm that aggregates real-time ride-hail pricing, simulated more than 94,000 ride requests in the San Francisco Bay Area between 27 November 2025 and 1 January 2026 and found an average Waymo fare of $19.69 against $17.47 for Uber and $15.47 for Lyft (TechCrunch, 27 January 2026³⁷). Against April 2025 the Waymo fare had fallen 3.6% while Uber's rose 12% and Lyft's 7%. Waymo still carries a 13% premium to Uber, down from 30–40% a year earlier. It collects that premium in a single high-fare market, so the figure does not generalise. Zoox launched paid service in Las Vegas on 10 August 2026 at what it describes as comfort-tier pricing, roughly 20–40% above standard UberX.

Nobody is competing on price yet. In a fixed-cost business with excess demand, that is the correct behaviour, and it is also the reason the industry's margin structure is untested. The first operator to face a genuine competitor in the same city on the same day will discover what its pricing power actually is.

11The arithmetic of one driverless mile

Seven things must all hold for a single paid driverless mile to happen, and four of them have no substitute if they fail.

The build-up below runs per vehicle-day rather than per mile, because the fixed-cost structure makes the day the natural unit, and converts to a per-mile figure at the end. Every assumption is dated, with its provenance stated.

Evidence: TechCrunch, 27 Mar 2026⁴ (rides/week, fleet from NHTSA data); Waymo blog, Aug 2026³⁸ (4m miles/week); TechCrunch/Obi, 27 Jan 2026³⁷ (fare); Forbes, 2 Aug 2026¹⁰ (vehicle cost); Markey RAO report²⁰ (remote assistance structure). Method: Every derived line is shown. Vehicle life, remote-assistance compensation, energy rate and consumption are AIR assumptions and are flagged Tier 3. Depot, cleaning, maintenance, insurance, mapping, city launch amortisation, aggregator commission and R&D are not estimated because no operator discloses them. Synthesis: The residual is the analysis. Everything the industry argues about publicly is 16% of the cost stack.

A human ride-hail passenger paying $20.98 sends $13.95 to the driver plus taxes and tolls, $6.35 to platform costs and $0.68 to platform profit. A robotaxi passenger paying $19.69 sends about $2.27 to vehicle capital, $0.51 to the remote desk, $0.36 to electricity — and $16.55 to a category nobody discloses.

The arithmetic needs two sanity checks, because a clean number is a suspicious number.

First, the fare. The $19.69 comes from San Francisco, the highest-fare US market and the one where Waymo has been operating longest. In Phoenix, Austin or Nashville the realised fare is lower and unknown. Push the average fare down to $15 and gross fares per vehicle-day fall to $357, the disclosed cost stack rises to 21% of revenue and the residual falls to $282. The conclusion holds; the cushion narrows.

Second, the residual is not profit. It has to fund depots and their leases, cleaning and charging labour, maintenance and tyres, incident response and towing, insurance premiums or self-insured reserves, HD mappingA centimetre-level survey of a road network — lane edges, kerb heights, signal positions, stop lines — that a driverless vehicle carries as prior knowledge, so it need not deduce everything from what it sees. It has to be built before service starts and kept current afterwards, which makes it a per-city cost rather than a one-off. and its upkeep, the amortised cost of every city launch, whatever commission Uber charges on the rides it sources, and the R&D that produced the stack in the first place. Alphabet's Other Bets segment — which contains Waymo and is described as generating revenue "primarily from the sale of autonomous transportation services and internet services" — lost $3,899 million in the first half of 2026, an annualised $7.8 billion (Alphabet Form 10-Q, 23 July 2026³⁹). Whatever the vehicle-level economics look like, the enterprise is a long way from covering its own cost.

The gap is an absorption problem: R&D and city-launch capital spread across too few vehicles, rather than unit economics that do not work. At $469 of gross fares a day, a vehicle covers its own capital and remote-assistance cost after 3.4 rides, 14% of its observed utilisation. Waymo is a business waiting for volume, which is a materially better place to be than a business waiting for a technology.

12What a city costs

Waymo raised $16 billion in February 2026 at a $126 billion post-money valuationWhat a company is deemed to be worth immediately after an investment round closes, counting the new money as part of the total. It is a price struck between a buyer and a seller for a slice, not an appraisal of the whole., led by Dragoneer, DST Global and Sequoia, with Alphabet funding the majority and a long list of crossover investors alongside (Waymo, 2 February 2026⁴⁰). The stated use of funds was fleet growth, hiring and entry into more than twenty additional cities in 2026 including Tokyo and London.

Twenty-plus cities against $16 billion implies several hundred million dollars per market, and that is before the fleet. The division is crude, and the company publishes no per-city budget. The order of magnitude is still right for what a launch involves: HD mapping and its maintenance, a depot with charging and cleaning capacity, a local operations and incident-response team, state and municipal permitting, law-enforcement and first-responder training, a validation programme in the new operational design domain, and the vehicles themselves at roughly $99,000 each.

The vehicles are the smaller half of it. Five hundred vehicles at $99,000 is $49.5 million — well under a fifth of the implied per-city figure. Autonomy scales by operational design domain rather than by software release, and most of the capital that scales with it goes into things other than silicon.

Alphabet does not name Waymo once in its Q2 2026 Form 10-Q. Other Bets revenue was $382 million in the quarter against $373 million a year earlier, and Other Bets contains Google Fiber — now classified as held for sale ahead of a transaction expected to close in late 2026, though its results stay in the segment until then — alongside Verily and the rest. One quarter earlier the number was stranger: Other Bets revenue was $411 million in Q1 2026 against $450 million in Q1 2025⁴¹ — down 8.7%. Across the half, segment revenue fell from $823 million to $793 million while Waymo's weekly ride volume roughly doubled. That $382 million is a hard ceiling on Waymo's recognised quarterly revenue. Against a $126 billion valuation it implies a multiple above 82 times an annualised upper bound, and the true multiple is higher because Waymo is only part of the segment.

The implied gross fare run-rate — 500,000 rides a week at $19.69 is $128 million a quarter — does not reconcile with that ceiling. A segment whose revenue is falling cannot contain a business whose gross bookings are doubling unless very little of those bookings is being recognised as revenue. Waymo's implied gross fares are a third of a segment that also contains a fibre business and a life-sciences business. The most likely explanations are that fares outside San Francisco are materially lower, that rides sourced through Uber are sold at UberX rates rather than Waymo's own, and that Waymo's share of them is recognised net rather than gross. Gross and net are accounting positions, not opinions. A company treated as the principal in a transaction books the whole fare as revenue and the payments it makes out of it as costs; a company treated as an agent books only its own cut. The same rides can therefore appear several times larger or smaller depending on which side of that line an arrangement falls. The Atlanta pricing term supports the second directly; the first and third are plausible and undisclosed. Waymo's recognised revenue is well below its gross bookings, and anyone modelling this business from ride counts and San Francisco fares is over-counting.

Evidence: Waymo, 2 Feb 2026⁴⁰; Alphabet 10-Q Q2 2026³⁹, Note 15 and Note 2. Method: Other Bets H1 2026 operating loss $3,899m × 2 = $7,798m. Ceiling logic: Waymo's recognised revenue cannot exceed total Other Bets revenue. Synthesis: The best-capitalised position in the industry is also the one that most conclusively demonstrates how much capital this takes.

13How the aggregators actually earn

Uber's second-quarter 2026 press release contains the sentence "we are... building the world's largest platform for autonomous vehicles." It contains no autonomous ride count, no AV gross bookings, no AV take rate and no AV margin. The strategy is asserted; its scale is not disclosed.

Uber ran $58.0 billion of gross bookings in the quarter across 3.9 billion trips and 208 million monthly active consumers, converting to $2.8 billion of adjusted EBITDA — 4.9% of gross bookings. Mobility alone did $28,988 million of gross bookings and $2,215 million of segment operating income. Lyft did $5.5 billion of gross bookings, 262.4 million rides and $177.2 million of adjusted EBITDA.

Those numbers describe a business built on two-sided liquidityHaving enough of both sides of a marketplace present at the same moment that neither has to wait — enough drivers that a rider gets a car in minutes, enough riders that a driver is rarely idle. It is hard to start and hard to dislodge once it exists. — enough of both sides present that a rider opens the app expecting a vehicle within four minutes. Airport concessions, corporate accounts, payment rails, insurance apparatus and regulatory relationships in fifteen thousand cities are what keep that expectation reliable. An autonomy stack buys none of them.

Uber's risk factors say so with unusual precision. The company warns that competitors "could introduce autonomous vehicle offerings earlier than we will be able to," that partnerships "could expire," that partners "could experience organizational failures of their own or otherwise remove their vehicles from our platform," and — the sentence that matters — that autonomous vehicles "could substantially reduce the cost of providing ridesharing... which could allow competitors to offer such services at a substantially lower price."

Uber's response has been to make itself indispensable to everyone and dependent on no one. The partner list disclosed across filings and partner releases now includes Waymo in Austin, Atlanta and Phoenix; Nuro, with Uber licensing the Nuro Driver and owning and operating the vehicles; Pony.ai, with a contracted deployment of more than 2,000 robotaxis across five European cities announced 14 August 2026; WeRide in Abu Dhabi, Dubai, Riyadh, Madrid and Zurich; Baidu's Apollo Go in Dubai and in London testing; and Wayve with Nissan for a Tokyo pilot.

In Atlanta, Uber riders requesting UberX, Comfort or Comfort Electric “could be matched with a Waymo fully autonomous all-electric Jaguar I-PACE vehicle – at no additional cost” (Uber Newsroom, 24 June 2025⁴²). On the aggregator's channel the robotaxi is sold at the human price, not at the premium Waymo commands on its own app. Whatever Waymo receives per Uber-sourced ride is a share of an UberX fare rather than of a $19.69 one. Most of the gap between Alphabet's Other Bets revenue and Waymo's implied gross bookings sits right there.

Uber has bought or contracted supply from at least six independent autonomy stacks on three continents. If any one of them succeeds spectacularly, Uber still sources rides. If any one fails, Uber loses a vendor. And through the Lucid arrangement it has gone further: Uber and its designated fleet operators have committed to a minimum purchase of 25,000 Lucid Midsize Plus robotaxis over six years from a targeted late-2028 start of production, bringing the aggregate minimum across both vehicle production agreements to at least 35,000 units (Lucid Form 10-Q, 4 August 2026⁴³).

An aggregator is underwriting a vehicle supply chain to guarantee the fleet is never scarce. It is the most consequential competitive move any company in this industry has made in two years, and it appears in a third party's filing rather than Uber's own.

Lyft has taken the same view at one-fifth the scale and with a different instrument. Rather than buy fleets, it is offering to run them. Flexdrive's AV fleet-management service positions Lyft as the operations layer for manufacturers who have a stack and no depots — a smaller, cheaper and more defensible position than trying to out-buy Uber. Its CEO's framing on the FY2025 call was that "2026 will be the year of the AV with deployments in the U.S. and overseas."

Both companies' best outcome is the same: many competing fleets, none dominant, all needing demand. Their worst outcome is one vertically integrated winner with its own app and enough vehicles to fill the peak. Everything Uber has done since 2024 is an attempt to make the first outcome more likely, and it has been effective.

14China: the other experiment

Two economies are running the same trial with the inputs reversed. In the United States, vehicles are expensive and fares are high. In China, vehicles are cheap and fares are low. Both sides have now been operating driverless at commercial scale long enough to read the result.

The Chinese operators disclose more than their American counterparts. Pony.ai reported robotaxi services revenue of $12.073 million in the second quarter of 2026, up 691% year over year, with fare-charging revenue up more than 800%, on a global fleet of 1,975 robotaxis as of 30 June and a target above 3,500 by year end. Total revenue was $36.2 million; gross margin 17.5%; loss from operations $65.7 million; net loss $45.4 million (Pony AI 6-K, 18 August 2026⁴⁴). WeRide's fleet reached roughly 3,400 L4 vehicles including more than 1,800 robotaxis as of 31 July 2026, on first-half revenue of RMB 345.9 million ($51.0 million), gross margin 36.6% and EBITDA of negative RMB 667.0 million (WeRide 6-K, 12 August 2026⁴⁵).

WeRide disclosed that average daily completed rides per vehicle in China rose 24% quarter over quarter in Q2 2026 to more than 21, with a peak of 28; Waymo's implied figure is 23.8. Ride throughput per vehicle is, within the precision of these disclosures, the same on both sides of the Pacific.

The Chinese operators have solved the problem the American industry has been agonising over — cheap purpose-built vehicles, mass-produced with automotive-grade components, BOM down 70% on the Gen-7 autonomous driving kit and a 2027 total vehicle cost heading below RMB 230,000, around US$33,000 — less, Pony.ai points out, than a locally built Tesla Model 3 costs in China (Pony.ai, 24 April 2026¹¹). Pony.ai's fleet went from 270 vehicles to more than 1,400 in a year, and to 1,975 by 30 June 2026. They have comparable ride throughput, and they still lose money at a rate that would be alarming if it were not being subsidised: Pony.ai's operating loss margin was 181.5% of revenue in Q2 2026, improved from 285.6% a year earlier.

Cheap vehicles are not the constraint; the fare is. The fare depends on which markets you are permitted to operate in. That permission depends on the validation and permitting you have completed, and that is what costs money and time. The Chinese operators cannot export their advantage to the US — the tariff wall and the Commerce Department's connected-vehicle rulesUS restrictions on vehicles containing Chinese hardware or software that links them to networks, imposed on national-security grounds. They bite on the parts of a car that talk to the outside world — which is most of what makes an autonomous vehicle work. see to that, and NVIDIA's own 10-Q flags the same restrictions as a risk to its automotive business. So they are exporting to the places where they can: the Gulf, Singapore, Europe, and increasingly through Uber's app rather than their own.

Pony.ai has contracted with Uber for more than 2,000 robotaxis across five European cities, bringing vehicles under international agreement to over 4,000. WeRide runs joint deployments with Uber in Abu Dhabi, Dubai, Riyadh, Madrid and Zurich, with roughly 400 robotaxis in the Middle East as of 31 July 2026, and describes its model as "asset-light." Baidu's Apollo Go went fully driverless commercially in Dubai with rides available through Uber, is testing in London with both Uber and Lyft, and holds Hong Kong's first fully driverless testing permits (Baidu 6-K, 18 August 2026⁴⁶).

The Chinese stacks have become supply for an American aggregator in third markets. The outcome is rational for everyone and poor for the Chinese operators' bargaining power: they bring the cheap vehicle and the working stack, Uber brings the rider, and the layer that owns the rider sets the split. Whether that is a durable arrangement is Section 28's question.

Three Baidu releases in a row are worth reading in sequence. For the fourth quarter of 2025, Baidu reported on 26 February 2026⁴⁷ that Apollo Go delivered 3.4 million fully driverless rides in the quarter, with weekly rides peaking above 300,000 and total rides up more than 200% year on year. For the first quarter of 2026, it reported on 18 May⁴⁸ 3.2 million rides, weekly peaking above 350,000 in March, growth decelerating to 120%. For the second quarter, the 18 August release⁴⁶ gives 28 cities and cumulative kilometres — over 350 million autonomous, over 240 million fully driverless — and no ride count, no weekly peak and no revenue.

Quarterly rides fell sequentially, from 3.4 million to 3.2 million, in the quarter before the number was withdrawn. And the replacement metric is cumulative distance, which can only ever go up. Apollo Go's fully driverless kilometres rose from over 220 million at the Q1 report to over 240 million at the Q2 report — roughly 20 million kilometres in a quarter, which at Q1's implied intensity is consistent with flat-to-lower ride volume rather than growth. The inference is from the company's own disclosures and is labelled as one, and it is why Apollo Go carries lower confidence in this brief than its city count and cumulative mileage would otherwise justify.

15The graveyard, read for the precondition each casualty failed

The most useful documents in this industry are the obituaries. Four significant robotaxi programmes have closed, each failing a different precondition. The base rate for replication is poor for that reason, and the survivors' positions are worth more than their technology suggests.

Cruise is the case that matters, because it is the only one where a company with a working product, a permissive regulator and a wealthy parent was killed inside fourteen months, and because the mechanism is entirely reproducible.

GM's own 10-K sets out the sequence without editorial. In October 2023 a Cruise robotaxi was involved in an accident. Over the following year the company settled with three separate authorities: a $112,500 fine and reporting obligations from the CPUC in July 2024; a Consent Order and $1.5 million fine from NHTSA in September 2024, with enhanced reporting for two years and an optional third at NHTSA's discretion; and, in November 2024, a Deferred Prosecution AgreementA settlement in which prosecutors bring charges but agree to drop them if the company admits specified conduct, pays, and behaves for an agreed period. It is not an acquittal — the admission is on the record and enforceable. with the US Attorney for the Northern District of California in which "Cruise admitted to one count of submitting a false report to a federal agency." In December 2024 GM announced it would no longer fund Cruise's robotaxi development and would acquire the minority interests, having reached approximately 97% ownership.

The crash did not end Cruise. The false report did. And the cost is measurable: across 2022 to 2024, GM's Cruise segment produced $461 million of revenue against $6,286 million of EBIT-adjustedGeneral Motors' own measure of segment profit: earnings before interest and tax, with items the company judges non-recurring stripped out. Useful for comparing one GM segment against another, and not comparable to anybody else's numbers. losses and $2,642 million of restructuring and compensation charges — roughly $8.9 billion of recognised cost, five cents of revenue per dollar spent, and that covers only the three years GM reported Cruise as a segment. The programme's full lifetime cost is larger and not separately disclosed.

Roughly $9 billion and a fourteen-month destruction window: that is the replacement cost of a safety culture and a working relationship with a regulator. Both are hard assets, and an incumbent's regulatory standing therefore belongs on the durable side of the layer stack in Figure 6, while an entrant's willingness to move fast counts against it in the underwriting.

16The conclusion, stated first

Whoever holds an audited safety record, a per-city operating machine and the insurance and regulatory standing those two buy keeps the profit once competitors arrive — and the second-best position belongs to the aggregator that has made certain no single operator can hold the rider hostage.

Autonomy is usually described as a technology race whose winner inherits urban transport. Three things are wrong with that description.

The first is that the technology has largely stopped being scarce. Wayve licenses its driver to automakers and raised $1.2 billion in February 2026 at an $8.6 billion valuation with Microsoft, Nvidia, Uber, Mercedes-Benz, Nissan and Stellantis participating. Nuro licenses its driver to Uber. Mobileye licenses Mobileye Drive and has now decided to operate a fleet itself. Baidu, Pony.ai and WeRide all license or joint-deploy. Six or more independent stacks can move a vehicle without a driver in at least one city. What remains scarce is the permission to do it, the record that earns the permission, and the operating apparatus that makes it repeatable.

The second is that the wage pool does not survive as a pool. Two-thirds of a fare left the platform with the driver; after autonomy, vehicle capital, remote assistance and energy account for roughly a sixth of gross fares, and the remainder disperses into depot operations, insurance, city launch amortisation and — where an aggregator sources the ride — commission. There is no single layer to buy in order to own the released money. There are four, and only one company currently holds all of them.

The third is that the aggregators were mispriced as roadkill. Uber has contracted supply from at least six stacks on three continents, underwritten a minimum of 35,000 vehicles through Lucid to guarantee the fleet is never scarce, and holds $13.3 billion of insurance reserves that any fleet operator will eventually need to rent or rebuild. Its exposure is real, and its position is far stronger than the substitution narrative allows.

Where this lands as verdicts:

Evidence: Aurora 10-Q Q2 2026²⁴; Mobileye 10-Q Q2 2026²³; Uber 10-Q Q2 2026⁷. Synthesis: Grades are on the robotaxi position only. Section 20 states explicitly how little Waymo moves Alphabet.

17Where the wage pool lands

The 66.5 points of fare that used to go to the driver are traceable, layer by layer.

Evidence: shares from the Section 11 build-up; layer ownership from the filings cited throughout. Method: "Does it stay?" answers one question — can the holder resist a price cut after a competent competitor arrives? Synthesis: Only two lines survive the test, and both are accumulated rather than purchased. Everything a chequebook buys, a chequebook competes away.

Two second-order consequences follow.

The first: the actuarial moat strengthens precisely as the technology gets cheaper. When autonomy hardware halves in cost, entry becomes easier for everyone — more stacks, more fleets, more competition on price. But every one of those entrants must be insured, and none of them has miles. They will be quoted off a curve built from the mile leader's experience, with a credibility loading for being an unknown. Cheaper technology therefore widens the incumbent's cost advantage in the one line item that cannot be engineered away. This is the opposite of how commoditisation usually works and it is the strongest single argument for Waymo's durability.

The second: one at-fault catastrophe reprices the whole field at once, not just the operator responsible. Insurers underwrite a class, not a company, and there are not yet enough operators for individual experience to dominate class pricing. Cruise's collapse cost Cruise; a fatal at-fault incident with a child, on camera, would raise every operator's premium and every state's permitting friction simultaneously. That is a correlated risk sitting under a sector that is otherwise diversified by geography and stack.

18Capture and commoditisation, layer by layer

Each scarce node gets the same four questions: who controls it, how durable that control is, whether it can be replicated, and what happens to the rent if capacity doubles. The industry sorts cleanly into two halves.

Sensors and compute. Nvidia's automotive line no longer appears as a separate segment; the Q2 FY2027 10-Q folds "automotive platforms and autonomous and electric vehicle solutions" into Compute & Networking, and the earnings release breaks revenue only into Data Center ($89.0 billion) and Edge Computing ($7.2 billion) of a $96.2 billion total. When a supplier stops disclosing a line, that line has stopped being material. Innoviz supplies lidar into Mobileye Drive; Magna does the integration for Waymo. Every one of these nodes has at least two qualified alternatives. Double the capacity and the price falls to the buyer. No durable rent.

Vehicle platforms. Zeekr, Hyundai, Lucid, Toyota, GAC, BAIC, Geely Farizon and Nissan are all building or supplying robotaxi platforms. Lucid's arrangement is instructive: it secured a minimum 25,000-unit commitment and $200 million of equity from a Uber-linked investor, and still recorded a $1.0 billion quarterly net loss with three workforce reductions in eighteen months. Being the chosen supplier to the aggregator is not the same as capturing value. No durable rent.

The autonomy stack sold as a product. Aurora — $2.0 million of revenue against a $270.0 million quarterly loss. Mobileye — $508 million of quarterly revenue, flat year on year, after a $3,788 million goodwill impairment. Wayve, Nuro and the Chinese operators all license. Double the number of licensable stacks and the licence fee falls. No durable rent. The participants know it, which is why they are integrating forward.

Fleet operations. Hard, capital-hungry, real-estate constrained, and now purchasable as a service from Lyft. Excellence here is worth two or three points of margin to an operator at scale. It keeps nobody out. A margin, not a moat.

Insurance and the actuarial file. Built from audited miles that took a decade and roughly $9 billion of comparable spending to accumulate, priced by reinsurers who have now published a rating basis derived from one operator's experience. Doubling the number of operators does not dilute it; it increases the number of parties who must be priced against it. Durable, and it compounds.

Regulatory standing. Three California deployment permits. Six driverless testing permits. A federal exemption capped at 2,500 vehicles a year and revocable at NHTSA's discretion. AV STEP withdrawn, so no national licence is coming. Destroyed in fourteen months by one false report to a federal agency. Durable, and asymmetric — slow to build, fast to lose.

Demand. Contested, and honestly so. Uber and Lyft hold liquidity, airport access, corporate accounts and payment rails across fifteen thousand cities. Waymo holds its own app in ten. Rider switching cost is near zero in both directions. This is the one layer where the outcome is genuinely undetermined, and Section 28 argues it turns on peak concurrency rather than on brand.

The incentive map

Several of the field's puzzles dissolve once the incentives are written down side by side.

Evidence: incentives inferred from each participant's own disclosures cited throughout Parts I and II — including Tesla's CEO milestone schedule⁸, Uber's AV risk factors⁷, and the Zoox exemption docket²¹. Method: Each row states the incentive as the participant's disclosures reveal it, then the structural consequence. These are AIR readings, not company statements. Synthesis: No participant's incentive points toward sharing rent with the layer below it, which is why every layer that can integrate forward is trying to.

19The preconditions, and how they were derived

A self-sustaining robotaxi business needs six things. Each is drawn from the evidence in Part I rather than asserted, and each is the precondition some casualty in Section 15 failed.

1. 01
   A licensable safety case

   Audited driverless miles the regulator and the insurer both accept. RAND's threshold is 275m; the record is reconciled or it is not a record.
2. 02
   Vehicle cost compatible with the fare

   At ~$99,000 landed and 24 rides a day, capital is 11.5% of gross fares — comfortable. At half the utilisation it is not.
3. 03
   An operating stack that undercuts the human

   Depots, cleaning, charging, staging, incident response and a 24/7 remote desk — the 84% residual nobody discloses.
4. 04
   Demand density to hit utilisation

   Fixed costs are absorbed only by ride volume, which is why density precedes geography in every operator's plan.
5. 05
   Bearable insurance

   Priced off the mile leader's file. Lyft paid 71 cents to cap a dollar of human-driver legacy tail; the driverless class is barely three years old.
6. 06
   Patient capital, and the governance to keep it

   ~$9bn and a decade, on the Cruise base rate — and a governance structure that survives the first bad incident.

Evidence: derived from Sections 4, 5, 8, 10, 11 and 15 and their underlying sources. Method: Each precondition corresponds to a documented failure mode in the graveyard: Uber ATG failed 1; Argo and Drive.ai failed 6; Cruise failed 6 through governance rather than money. Synthesis: Five of six are economic or institutional. Only the first is partly technical, and even it is mostly a question of accumulated exposure.

Each per-name verdict applies these six. It states what is deployed, permitted or contracted today as against what is demonstrated or promised; the evidence; the name's own dated falsifierA specific, observable event that would show a claim to be wrong. Naming one in advance is what separates a thesis from an opinion: it fixes the terms on which the author agrees to be judged.; and a confidence sized to how much of the underlying data is independently verifiable. Where the robotaxi position sits inside a conglomerate, the dilution is stated.

20Alphabet — Waymo

Deployed today. 500,000 paid autonomous rides a week across ten US cities as of 27 March 2026, on roughly 3,000 vehicles. Public rider-only service in Phoenix, the San Francisco Bay Area, Los Angeles, Austin, Atlanta, Miami, Nashville, Dallas, Houston, San Antonio and Orlando — the last four opened to first public riders simultaneously on 24 February 2026⁵¹, taking commercial metro areas to ten. Freeway operation opened to public riders in three metros in November 2025; SFO airport service from January 2026, plus San José Mineta and Phoenix Sky Harbor. Fully autonomous operations preparing in Denver, Las Vegas, San Diego and Tampa (Waymo, 8 July 2026¹²). More than 4 million autonomous miles a week. The only holder of a California driverless passenger deployment permit and the only carrier in the CPUC Deployment programme.

Demonstrated or promised. Germany, Tokyo and London are announced, not operating. Munich is targeted for late 2027, with mapping and specialist-supervised testing beginning and approvals to be secured from Germany's Federal Motor Transport Authority; London testing began in April 2026 with safety drivers, and Transport for London has said publicly that no autonomous vehicle currently meets its standards. Twenty-plus additional cities in 2026 is a plan. The Hyundai IONIQ 5 is being driven with a specialist aboard, not in rider-only service. Profitability at the enterprise level is unproven and, on the Other Bets numbers, distant.

The evidence for the grade. Precondition 1: 220.6 million rider-only miles, independently corroborated by IIHS at 68% fewer crashes and 81% fewer injury crashes across 50 million miles. Precondition 2: roughly $99,000 landed per vehicle at 11.5% of gross fares. Precondition 3: an owned integration facility in Mesa with Magna, depots in eleven cities, and a remote-assistance ratio of about 1:43. Precondition 4: ten metros with airport and freeway access, the highest-value demand in each. Precondition 5: a published Swiss Re rating basis derived from its own miles. Precondition 6: $16 billion raised in February 2026 at $126 billion post-money, with Alphabet funding the majority.

The conglomerate dilution, stated plainly. Other Bets — which contains Waymo, Verily, Google Fiber and the rest — generated $382 million of revenue in Q2 2026 against Alphabet's $119.8 billion, or 0.32%. The segment's $1,799 million quarterly operating loss is 4.4% of Alphabet's $40,770 million of operating income. Alphabet's Q2 2026 Form 10-Q does not contain the word "Waymo." An investor buying Alphabet for its robotaxi position is buying a rounding error attached to a search and cloud business; the grade here is on the position, and the position does not move the parent.

What runs against it. Three qualifications matter here. The IIHS study found Waymo's crash rate 4% higher than humans in Austin, its newest large market — evidence that the safety asset is domain-specific and does not port free. Half its remote-assistance workforce sits in the Philippines with local rather than US licences, which the Senate has flagged and which NHTSA has just prohibited for exempted vehicles. And expansion into twenty-plus cities in a year is precisely the kind of pace that produced Cruise's governance failure; the safety record is an asset that a single mishandled incident can impair faster than a decade of miles built it.

The strongest counter-evidence arrived in July 2026 and it cuts in Waymo's favour on strategy and against it on operations. Waymo gave notice ending its Uber exclusivity in Austin and Atlanta, will launch its own app in both cities in January 2028, and has added no new Uber city since Atlanta — an operator confident enough to drop the largest demand aggregator on earth. Bloomberg reported that Uber's side of the disagreement included “persistent safety problems, and a failure by Waymo to be forthcoming about incidents,” naming vehicles driving into flooded roads despite a software recall and illegally passing school buses in Austin. Those attributions are Bloomberg's and unconfirmed here. But the school-bus incident is independently corroborated: the Markey investigation records an NTSB finding that a Waymo passed a school bus with its stop sign extended after acting on incorrect remote-assistance guidance. That is the Cruise mechanism in miniature — an operational failure plus a disclosure question — and it is the specific risk that a High-confidence grade on this name has to live with.

21Tesla

Deployed today. A robotaxi service in Austin since June 2025, extended to Dallas and Houston on the evidence of the SGO file, which contains 24 Tesla ADS crash reports across those three cities. Eighteen of the 24 narratives state a safety monitor was present in the vehicle. Cybercab production began in the first half of 2026 (Tesla Form 10-Q, 23 July 2026⁸). Tesla holds a California testing-with-driver permit under the entity "Tesla Robotaxi LLC" and appears on neither the driverless-testing nor the deployment list.

Demonstrated or promised. Everything else. Tesla discloses no robotaxi revenue, no ride count, no fleet size and no city-level service metrics in its 10-Q. The scale of the Austin operation is known only through third-party observation and the SGO file. The service is described in the filing as having been "expanded and refined" since launch, without quantification.

What one regulator was willing to write down. Nevada is the only jurisdiction that has published the operative terms of a Tesla robotaxi authority, and they are narrow. AVNC Permit 002, issued 27 July 2026, restricts Tesla Robotaxi LLC to "a maximum fleet of ten (10) fully autonomous vehicles" in the Las Vegas Strip corridor, bars passenger service on roads posted above 45 mph, prohibits pickups within a quarter-mile of Harry Reid International, and requires that operations "include appropriate human supervision." On 21 August 2026 the Authority approved Tesla's application for a twelve-month ceiling of up to 5,000 vehicles, with the permit to issue on verification of conditions. The gap between ten and five thousand is the whole Tesla question in two numbers, and only the first of them is currently operative.

The evidence. Precondition 1 is unmet and contested. NHTSA opened Engineering Analysis EA26002 on 18 March 2026 across 3,203,754 vehicles, and recorded that Tesla "described internal data and labeling limitations that prevented a uniform identification and analysis of crash events with the subject system engaged," with ODI's own view that this "could have led to under-reporting." A parallel preliminary evaluation, PE25012, covers 58 incidents of red-light running and wrong-way lane changes. Precondition 2 is Tesla's genuine strength: it manufactures its own vehicle at automotive scale, and the Cybercab should undercut every retrofit. Precondition 3 is early — the 10-Q describes building cleaning, charging, security, teleoperation and fleet-management infrastructure, in the present tense. Precondition 5 is untested: no published insurance study, no actuarial basis, no disclosed claims experience. Precondition 6 is unambiguously satisfied.

The evidence-quality problem, quantified. Of Tesla's 1,703 unique Level 2 crash reports in NHTSA's Standing General Order file through July 2026, 96.1% record the highest injury severity as “Unknown.” Waymo's equivalent share, in the driverless file, is 1.3%. The mechanism is structural rather than evasive: an operator that dispatches an incident team learns the outcome, and a manufacturer notified by telematics of a customer's crash often cannot. The supervised-mileage advantage has therefore not converted into a safety case. Cheap miles that cannot be adjudicated are not evidence.

The remote-operations disclosure. Tesla told the Markey investigation that its remote operators can take control of a vehicle and drive it at up to 10 mph — the only operator of the seven to confirm remote driving rather than remote assistance. Two of the 24 SGO reports describe crashes caused by exactly that. Tesla declined, with the other six, to disclose intervention frequency.

Who would carry the car. If a fleet is ever assembled from customer-owned vehicles, the household buys the depreciating asset while Tesla keeps both prices that decide whether it earns: the FSD subscription the vehicle needs in order to operate, and the platform's share of each fare. Neither is contractually fixed, and neither is disclosed for robotaxi use. An owner in that position holds the layer Section 18 grades as commoditising — the vehicle, where falling cost accrues to the buyer — while the two variables that set the return sit one layer above. Tesla reports two segments, automotive and energy generation and storage; there is no mobility segment in which such an arrangement would surface.

The regulatory pivot that matters most to Tesla. If NHTSA's proposed FMVSS 135 modernisation completes, a Cybercab without manual controls could be self-certified rather than requiring a Part 555 exemption — removing the 2,500-vehicle annual cap that would otherwise bind it. Tesla has not filed an exemption petition. On the public record it is waiting for the standard to move rather than asking for relief from it, which is a coherent strategy and a slow one.

Tesla holds the industry's best answer to precondition 2 and its weakest answer to precondition 1. Precondition 1 is the one that governs market access. A camera-only stack may well prove licensable; the argument against it has never been about photons. It is that a safety case requires a reconciled numerator and denominator, and the federal regulator has stated on the record that Tesla could not supply one. Until that changes, the vehicle-cost advantage has nowhere to be spent.

22Uber

Contracted today. Waymo on the Uber app in Austin, Atlanta and Phoenix. Nuro's driver licensed, with Uber and third-party fleet partners owning and operating Lucid Gravity vehicles; Nuro holds a California driverless testing permit. Pony.ai contracted for more than 2,000 robotaxis across five European cities (Pony AI 6-K, 14 August 2026⁵²). WeRide joint deployments in Abu Dhabi, Dubai, Riyadh, Madrid and Zurich. Baidu's Apollo Go live in Dubai and testing in London. Wayve with Nissan for Tokyo. Two vehicle production agreements with Lucid totalling a minimum of 35,000 units.

What is not disclosed. Any of it, in numbers. Uber's Q2 2026 press release and 10-Q contain no AV ride count, gross bookings, take rate or margin. The claim to be "building the world's largest platform for autonomous vehicles" is unaudited. Investors are asked to take the strategy on the partner list.

Why the grade is advantaged despite that. Three assets, all disclosed and all underweighted.

The first is the insurance apparatus. $13.3 billion of reserves at 30 June 2026, 35.5% of total liabilities, backed by $9.5 billion of restricted investments, run through a captive insurer with thirty years of claims machinery. A fleet operator entering a new state must either buy that capability or rent it. Uber already owns it, at scale, and the Nuro arrangement — where Uber owns and operates the vehicles — places the insurable event exactly where the reserves are.

The second is that Uber's supply is genuinely diversified. Six or more independent stacks across the US, Europe, the Gulf and Asia. Any single partnership expiring is a vendor loss, not an existential one. The Lucid minimum-quantity guarantee is the sharpest expression of the strategy: rather than depend on a fleet owner's goodwill, Uber has contracted the vehicles into existence.

The third is peak concurrency, argued in Section 10 and worth restating here. A robotaxi averages about one ride per hour of a 24-hour day. Human supply flexes to the peak and disappears afterwards, at no cost to the platform. As long as urban demand is peaked, the marginal Friday-evening ride is served by a human, and Uber loses its lowest-value hours first.

The fourth asset is newer and changes the shape of the position. Uber is no longer only an aggregator: through Aviari Services, LLC, a Nevada subsidiary, it applied for and on 20–21 August 2026 obtained approval from the Nevada Transportation Authority to operate as an autonomous vehicle network company, with a twelve-month ceiling of up to 1,000 fully autonomous vehicles in Clark County on the same conditions imposed on Tesla and Waymo (Nevada Department of Business and Industry, 21 August 2026²⁹). Uber had already launched a robotaxi service with Motional in Las Vegas in March 2026. The aggregator has begun taking out its own operating authority — what a company that has read the layer stack in Figure 6 would do.

What runs against it — and it is serious. On 24 July 2026 Uber confirmed that Waymo had given notice ending their exclusivity in Austin and Atlanta. Waymo will launch its own app in both cities in January 2028 and remain available on Uber through at least May 2028 under the existing contract (reported 25 July 2026⁵³, on disclosures confirmed by Uber to CNBC and Bloomberg). Bloomberg reported Uber's frustrations as “unsustainable economics, persistent safety problems, and a failure by Waymo to be forthcoming about incidents,” citing robotaxis driving into flooded roads despite a software recall and illegally passing school buses in Austin — the second of which matches the NTSB finding in the Markey report about a Waymo acting on incorrect remote-assistance guidance. Those attributions are Bloomberg’s and were not independently confirmed for this brief; the notice itself was confirmed by Uber. Waymo has also announced no new Uber city since Atlanta, expanding instead through its own app.

The rest of the bear case still stands. Uber’s take rate is more at risk than its volume: an operator with fleet scale and demand density in one city can price below Uber’s all-in cost — Uber’s own risk factor says exactly this — and need not share revenue with anyone. Mobility segment operating income of $2,215 million on $28,988 million of gross bookings is a 7.6% margin, which leaves little room to buy loyalty from either side of the market.

Why the grade stays advantaged, stated explicitly. A reader is entitled to ask why an aggregator qualifies at all under a criterion that hands the lasting profit to accumulated operating and regulatory assets. Uber holds two of those assets outright — a $13.3 billion insurance balance sheet built over a decade, and now its own state operating authority — and the Waymo split is evidence about Waymo rather than about aggregation. Waymo is the single operator that has satisfied all six preconditions; it does not need a demand partner, and its economics with one are, on Uber’s account, poor. The other five stacks on Uber’s platform have satisfied none of the six in the United States. If the price of aggregation is losing the one partner strong enough to leave, that is a take-rate problem rather than an existence problem. Confidence is cut to Moderate accordingly, and the first falsifier below has been partially triggered on a dated schedule.

23Lyft

Deployed today. A profitable ride-hail network: $5,504.2 million of gross bookings, 262.4 million rides, 30.5 million active riders and $177.2 million of adjusted EBITDA in Q2 2026, with FY2025 gross bookings of $18.5 billion and free cash flow above $1.1 billion. Flexdrive provides AV fleet-management services to manufacturer partners, with contractual commitments on "vehicle availability, maintenance, charging, cleaning, staging and depot operations." Lyft also disclosed that in Nashville, in partnership with Waymo, fleet operations officially began in June and are running smoothly as the company gears up for the opening of its 80,000-square-foot purpose-built AV depot in October; Lyft and Curb expanded their strategic partnership to New York City, the largest taxi market in the U.S. (Lyft Q2 2026 earnings release⁵⁴).

Demonstrated or promised. The AV deployments themselves. Lyft's CEO called 2026 "the year of the AV with deployments in the U.S. and overseas"; the 10-Q names Alphabet, Amazon, Baidu, Bolt and Tesla as competitors and does not name a single AV partner in the risk factors, which is unusual and not encouraging. Baidu's own release confirms Apollo Go is testing in London with Lyft as well as Uber. The company has previously written off an AV partner: "in the past we have incurred impairment charges in connection with the wind-down of an AV partner."

The Flexdrive bet. Flexdrive is the only public expression anywhere of the thesis in Section 7 — that fleet operations is a service business, not a moat. If depot operations can be bought, the layer commoditises, and the buyer of that service is any manufacturer with a stack and no interest in leasing real estate in eleven cities. Lyft is small enough that a fleet-management contract materially moves its economics and large enough to have the depots, the cleaning crews and the insurance reserves already. The position is a genuinely sensible one for a company that cannot win the capital contest.

Fleet management makes no claim on the wage pool, though. It is cost-plus, with a margin set by the customer's alternative — doing it themselves. Lyft's grade is neutral because its core marketplace is healthy and improving and its AV position is a hedge rather than a thesis.

24Amazon — Zoox

Deployed today. Paid commercial service in Las Vegas from 10 August 2026 with a launch fleet of 64 robotaxis, holding the first Nevada autonomous-vehicle-network-company permit and operating under NHTSA Temporary Exemption No. 2026-01, granted 31 July 2026 and effective to 31 July 2028, exempting portions of eight FMVSS. Free service in San Francisco, Austin and Miami. A California driverless testing permit but no deployment permit and no CPUC deployment authority, which is why San Francisco remains free. 61 SGO crash reports across the window, none serious or fatal.

What the exemption actually gives and costs. It permits Zoox to introduce up to 2,500 exempted vehicles into commerce in any twelve-month period — the statutory ceiling under 49 U.S.C. 30113(d). NHTSA attached a new "Operational Authorization" that it may modify or expand at will, a requirement that all safety-critical remote personnel be located in the Continental United States, an obligation to publish operating-area maps updated within fourteen days of any substantive change, and annual VIN-level reporting including the geographic location of every remote-assistance centre. This is a materially more intrusive supervision regime than a self-certified vehicle faces, and it is the price of the cabin.

The structural bind. Zoox has the industry's most elegant answer to precondition 2 — a purpose-built vehicle with no driver's cabin, no steering column and a campfire seating arrangement, which should be cheaper per seat-mile than any retrofit. And it cannot scale that answer past 2,500 vehicles a year without a change in federal law. Waymo, whose vehicles keep their steering wheels, imported more than 2,600 in 2026 alone. The elegant design is capped; the inelegant one is not.

The conglomerate dilution. Amazon's Q2 2026 10-Q does not break out Zoox. The only reference is in the technology-and-infrastructure cost discussion, which mentions "the development of a satellite network for global broadband service and autonomous vehicles for ride-hailing services." Against $200.6 billion of quarterly revenue, Zoox is not a line item. As with Waymo, the position may be structurally interesting and the parent exposure is immaterial — with the added consideration, which Amazon has never confirmed, that a purpose-built driverless platform has obvious value in last-mile logistics regardless of how ride-hail goes.

25Baidu, Pony.ai and WeRide

Evidence: Baidu 6-K, 18 Aug 2026⁴⁶; Pony AI 6-K, 18 Aug 2026⁴⁴; WeRide 6-K, 12 Aug 2026⁴⁵. Synthesis: All three have the operating stack, the cheap vehicle and comparable ride throughput. None can bring any of it to the United States, and all three now monetise substantially through Uber's app in third markets.

These three are graded on the same evidence and separate on one variable: whether the parent can carry the losses.

Pony.ai and WeRide are exposed because they are standalone companies burning cash against small revenue in a business whose defining requirement is patience. Pony.ai lost $65.7 million from operations in a quarter on $36.2 million of total revenue and took a $25.0 million impairment on prepayments for long-term investments "determined to be unrecoverable." WeRide's first-half EBITDA was negative RMB 667.0 million on RMB 345.9 million of revenue. Both are executing well — Pony.ai's fleet nearly doubled to 1,975 with a 3,500 year-end target, WeRide's China rides per vehicle rose 24% quarter over quarter to more than 21. Both are running the Cruise experiment without a General Motors behind them, and that is the grade.

Baidu is neutral because Apollo Go sits inside a company with an AI cloud business growing 50% and positive operating cash flow for four consecutive quarters, so the patience precondition is satisfied. It carries lower confidence than the other two for the opposite reason to Waymo. Baidu reported 3.4 million fully driverless rides for Q4 2025 and 3.2 million for Q1 2026 — a sequential decline — and then, for Q2 2026, reported no ride count at all, substituting cumulative kilometres. A metric that can only rise has replaced one that had begun to fall. The substitution is the most informative item in Baidu's robotaxi disclosure, and it is an absence.

The common structural fact. All three are locked out of the highest-fare market on earth by tariff and by connected-vehicle rules — restrictions NVIDIA's own 10-Q flags as a risk to its automotive customers. Their response has been to license, joint-deploy and ride on Uber's demand in the Gulf, Europe and Southeast Asia, a rational choice that hands the pricing decision to the aggregator. Their competitive advantage is a US$33,000 vehicle. The fare in the markets open to them is a fraction of the US level, and the demand arrives through someone else's app. Nobody captures rent from a position like that.

26Mobileye, Nvidia and the supplier squeeze

Two events in the first half of 2026 tell the supplier story. In the first quarter, Mobileye wrote off $3,788 million of goodwill against its own reporting unit, producing a $3,839 million half-year net loss on revenue that grew 0.4% year over year in the second quarter — $508 million against $506 million. In June, it announced it would build and operate its own robotaxi fleet.

A company whose entire thesis was selling autonomy to everyone marked down the value of that thesis by $3.8 billion, and three months later concluded it needed to own the customer instead. Its own 10-Q describes the move in the risk factors: a plan "to establish a vertically integrated robotaxi business that would expand our role beyond supplying autonomous-driving technology to owning and operating an autonomous ride-hailing service," which "is in its initial stages, and may require significant management attention, Company resources, capital investment and operating expenses." The 16 June 2026 release⁵⁵ puts an initial fleet of roughly 100 vehicles in a major US metro in 2027, scaling toward roughly 17,000 over the following five years, integrating Mobileye Drive with Moovit's platform, fleet management and teleoperation into a single division. Its CEO framed it as "an extension" of the supplier business rather than a replacement. The Q2 results release is more revealing than the announcement was: Mobileye reports being "engaged with potential vendors for vehicle platform supply, self-driving-system installation and upfit, and fleet logistics support," while accelerating Moovit's work on "customer-facing app, fleet optimization, and rider engagement applications" (Mobileye Q2 2026 results, 23 July 2026⁵⁶). That is a supplier assembling, one procurement at a time, every layer in Figure 6 it does not already own — the clearest available confirmation that the supplier layer does not capture the rent.

The grade is exposed, and the reason should be explicit: 100 vehicles in 2027 places Mobileye roughly where Waymo was in 2019, into a market where the leader will have five years and two hundred million more audited miles, and the capital to fund it must come from a business whose revenue is flat and whose goodwill has just been halved. This grade runs against the natural instinct that a technology leader with an installed base is advantaged. The defence is the thesis of this brief: value accrues to accumulated operating and regulatory assets, and on that criterion Mobileye starts from zero in 2027.

Nvidia is the cleanest case in the industry of activity without rent. Automotive is no longer a reported segment: the Q2 FY2027 10-Q folds "automotive platforms and autonomous and electric vehicle solutions including software" into Compute & Networking, and the earnings release splits $96.2 billion of revenue into Data Center at $89.0 billion and Edge Computing at $7.2 billion, with no automotive line. The company also warns that US "Connected Vehicle" restrictions "could impact our ability to develop and supply solutions for our automotive customers." Nvidia will sell compute into every robotaxi built, and it will not matter to Nvidia. The operators will have at least two qualified alternatives. Neither side has pricing power over the other: a commoditised layer, seen from both ends.

Evidence: Mobileye Form 10-Q, 23 July 2026²³; NVIDIA Form 10-Q, 26 August 2026⁵⁷; NVIDIA Q2 FY2027 results, 26 August 2026⁵⁸. Synthesis: Both grades are on the robotaxi position. NVIDIA's overall business is not the subject of this brief and no view is expressed on it.

27The system actors

Four private or peripheral participants shape the economics without being investable on their robotaxi position, and one public company shows what the stack alone is worth.

Nuro has executed the most commercially astute pivot in the field. Having abandoned its own delivery-vehicle deployment, it now licenses the Nuro Driver as a service and lets someone else carry the fleet risk — specifically Uber, which “has licensed the Nuro Driver self-driving AI system and will own and operate the vehicles along with its third-party fleet partners,” on a Lucid Gravity platform integrated on Lucid's Casa Grande line and running NVIDIA DRIVE AGX Thor (Nuro⁵⁹). Nuro holds both a California driverless testing permit and a California deployment permit, one of only three deployment permits issued. It is also the counter-example to the licensing thesis: licensing works precisely because Nuro is not trying to capture the rent, only a fee.

Lucid is the cautionary case. A minimum 25,000-unit commitment from Uber, $200 million of equity from a Uber-linked investor, a jointly developed robotaxi unveiled at CES 2026, 75 engineering vehicles delivered — and a $1.0 billion quarterly net loss, three workforce reductions in eighteen months and start of production for the Midsize Plus not targeted until late 2028. Being the chosen vehicle supplier to the largest aggregator on earth has not made Lucid a good business, which is the vehicle layer's verdict in a single company.

Wayve raised $1.2 billion in a Series D announced 25 February 2026⁶⁰ at an $8.6 billion post-money valuation, within $1.5 billion of total capital secured including milestone-based investment from Uber. The round was led by Eclipse, Balderton and SoftBank Vision Fund 2, with Microsoft, Nvidia, Uber, Mercedes-Benz, Nissan and Stellantis participating. Wayve licenses its AI Driver to automakers without HD maps and claims to be the only AV company to have driven zero-shotA machine-learning term for performing in a situation the system was never specifically trained on. Driving zero-shot in a city means arriving without that city's maps or tuning and coping anyway. in more than 500 cities across Europe, North America and Japan in a single year. It is preparing a Tokyo pilot with Nissan and Uber and London trials. Its significance to this analysis is as supply: a sixth independent, credible, licensable stack, which is exactly what makes the autonomy layer commoditise. Wayve's own economics are not publicly assessable and no grade is offered.

May Mobility, Motional and Avride appear in the SGO file at 16, 11 and 72 reports respectively and in the Markey investigation. They matter here as evidence that the number of operators running automated vehicles on US roads is larger than the two or three that dominate coverage — and as a reminder that operating an ADS and running a self-sustaining robotaxi business are different achievements.

Aurora is the proof. A public company, a competent stack, driverless trucks on Texas highways — a new International LT fleet launched in July 2026 with nobody behind the wheel, and an allocation to exit the year with 200 driverless trucks (Q2 2026 shareholder letter, 29 July 2026⁶¹) — and $2.0 million of revenue against a $270.0 million net loss in the quarter ended 30 June 2026, with the filing stating it "will need to opportunistically raise additional capital." Trucking is a different market with different economics, and the point transfers exactly: an autonomy stack without a demand franchise, an operating machine and an insurance position is a research programme with a ticker.

Magna, Hyundai, Geely and Zeekr are the supply chain. Magna co-operates Waymo's Mesa integration facility; Hyundai builds the IONIQ 5 at its Georgia plant, free of China tariff; Zeekr ships stripped base vehicles that are Chinese in origin and American in everything that touches the road. None of them is graded here because none has a disclosed robotaxi-specific position that can be separated from its parent, and the analysis in Section 3 assigns the vehicle layer no durable rent in any case.

28Aggregation versus integration, resolved as far as the evidence allows

On 24 July 2026 Waymo told Uber it was ending their exclusivity in Austin and Atlanta. Waymo's own app arrives in both cities in January 2028; its vehicles stay on Uber through at least May 2028 under the existing contract. That is the most direct evidence anyone has produced on the industry's most consequential open question, and it points toward integration.

The steelmanThe strongest version of an argument you disagree with, put as its best advocate would put it. Beating the weak version proves nothing. for integration is strong and should be stated at full strength. A network effect that arises from matching many part-time drivers to many riders does not obviously survive when supply becomes a fleet of company-owned vehicles dispatched from a depot. In that world liquidity is a capital-allocation decision, not an emergent property of a marketplace, and the operator who owns the vehicles can dispatch them from its own app for free rather than paying an aggregator a commission. Waymo has the highest customer satisfaction scores in the category, an app already installed in ten metros, and no reason to share revenue with a middleman once density is sufficient. On this reading Uber is a transitional demand channel that will be dropped city by city as fleets thicken. Uber's own risk factors concede the possibility in terms.

And the notice is not the only support for it. Waymo has announced no new Uber city since Atlanta, expanding instead through its own app into six further markets, and Bloomberg reported Uber citing “unsustainable economics” in the arrangement. An aggregator that cannot make the numbers work with the best operator in the field, and an operator that would rather build its own demand channel, is a strong prima facie case that the app layer loses.

Three things the notice does not settle run the other way.

The first is peak concurrency, and it is decisive as long as urban demand stays peaked. A robotaxi averages 0.99 rides per hour across a 24-hour day. To serve the Friday 6pm surge from an owned fleet, an operator must own vehicles that will be idle at 3am on Tuesday and pay for them anyway. A human network solves this for nothing, because drivers self-select into the peak. The aggregator's demand is worth most in precisely the hours the fleet cannot serve alone.

The second is the multi-partner structure Uber has built. It works as a bargaining position, not merely a hedge. When six independent stacks can supply autonomous rides — Waymo, Nuro, Pony.ai, WeRide, Apollo Go, Wayve — no single one of them can threaten to withdraw and hurt Uber more than it hurts itself. And the Lucid minimum-quantity guarantee goes further: it means Uber can create fleet supply rather than negotiate for it. An aggregator that can commission 35,000 vehicles into existence is not a price-taker.

The third is the insurance layer, which almost every analysis of this question omits. Uber holds $13.3 billion of reserves and a captive insurer. Any operator scaling into a new state must underwrite the driverless mile there, and the cheapest way to do that at the margin is to plug into an existing balance sheet rather than build one — a reason to stay on the platform that has nothing to do with demand.

Evidence: utilisation from the Section 11 build-up; supply structure from Lucid 10-Q Q2 2026⁴³ and the Pony.ai and WeRide releases; reserves from Uber 10-Q Q2 2026⁷. Synthesis: The resolution is partial. Volume is defensible; price is not. Expect the aggregator to keep the rides and lose the take rate.

The resolution has moved since the Waymo notice. Aggregation loses to integration for the one operator that has satisfied all six preconditions, and that is exactly what the January 2028 date records. It does not follow that aggregation loses generally, because no other operator is close to Waymo’s position, and because Uber’s answer to the notice was not to plead but to take out its own operating authority through Aviari and to underwrite 35,000 vehicles through Lucid. Uber’s volume in autonomous rides remains more defensible than consensus allows, because the peak cannot be served economically by a fleet alone and because five of the six stacks on its platform have no alternative demand channel. Uber’s take rate is not defensible, because the counterparty is a handful of sophisticated operators rather than millions of atomised drivers. The likely equilibrium is that Uber keeps most of the rides and earns materially less per ride than the roughly 30% it takes from a human trip — and that the layer it actually monetises shifts from matching toward insurance, fleet operations and, increasingly, ownership.

What would falsify what remains of it: sustained rides per vehicle above roughly 40 a day in a major metro would show a fleet can economically serve the peak, and the complementarity argument would collapse with it. A second operator giving Uber notice would show the January 2028 date was about aggregation after all, rather than about Waymo.

29The debates, argued fairly

Can a camera-only stack carry a licensable safety case? Argued as physics this is unresolvable and boring. Argued as validation economics it is tractable. A safety case is a reconciled numerator over an audited denominator. Sensor fusion produces fewer, more expensive miles that all count as driverless. Camera-only produces vastly more, far cheaper, supervised miles whose evidentiary value depends entirely on being able to say which of them involved the system and which did not. NHTSA has stated on the record that Tesla could not do that reliably. Bookkeeping failures are fixable, and until this one is fixed the cheap miles do not convert into a safety case. Resolving evidence: closure of EA26002 without a recall, plus disclosed crash-labelling remediation and a peer-reviewed rate comparison on unsupervised miles. Known Unknown

Do the aggregators' network effects survive? Partly, and for a reason that has nothing to do with brand — the peaked demand, substitutable supply and insurance balance sheet Section 28 sets out. The network effect as classically described — more drivers attract more riders attract more drivers — genuinely weakens. What replaces it is a different asset: being the only party with enough demand to fill the hours a fleet cannot fill alone. Resolving evidence: partially in. Waymo gave notice on 24 July 2026 ending Uber exclusivity in Austin and Atlanta from January 2028, which settles the question for the strongest operator and leaves it open for the other five. A second operator giving notice would settle it generally. Known Unknown

Do cheap purpose-built vehicles or falling retrofit costs set the cost curve? Neither, currently — the tariff wall and the 2,500-vehicle exemption cap set it. A Chinese purpose-built platform at about US$33,000 cannot enter; an American purpose-built platform without manual controls cannot scale past 2,500 units a year. The retrofit wins by default, at $99,000. This is the clearest case in the brief of a temporary condition masquerading as a structure, and it has a specific expiry: the FMVSS 135 rulemaking. Resolving evidence: a final rule permitting self-certification of vehicles without manual controls. Known Known

Can remote-assistance ratios fall fast enough? They already have. At 1:43 the desk costs roughly 2.6% of gross fares against the driver's 66.5%, and even a punitive 1:20 ratio at $70,000 a head is 8.6%. The interesting question is no longer cost but jurisdiction: NHTSA now requires remote assistants for exempted vehicles to sit in the Continental United States, which forecloses the offshoring that half of Waymo's desk currently uses. Resolving evidence: extension of the Continental-US condition beyond exempted vehicles, or an operator disclosing intervention frequency. Estimated

Does peak concurrency cap the robotaxi's share of the profitable hours? Yes, structurally, until fleets are large enough to serve the peak and dense enough to absorb the resulting off-peak idleness. That is a capital problem, not a technology problem, and the reason this is the most durable constraint in the industry. Resolving evidence: rides per vehicle per day sustained above ~40 in a major metro. Estimated

Was Cruise a technology failure? No, and treating it as one is the most expensive misreading in the field. GM's 10-K records a $1.5 million NHTSA fine, a $112,500 CPUC fine, and a Deferred Prosecution Agreement under which Cruise "admitted to one count of submitting a false report to a federal agency." The vehicle's conduct was survivable. The company's was not. Known Known

30Alternatives rejected, and what runs downstream

Alternatives considered and rejected

"The vehicle and sensor supply chain captures the value, as it did in electric vehicles." Rejected. There are at least eight vehicle platforms and multiple qualified suppliers at every hardware node; falling BOM accrues to the buyer, and the layer is 11.5% of gross fares and shrinking. Reopens if: a single sensor or compute architecture becomes a regulatory requirement, creating a licensed chokepoint where none exists today.

"The autonomy stack is the product, and licensing is the business model." Rejected on the numbers. Aurora: $2.0m of revenue against a $270.0m quarterly loss. Mobileye: $3,788m written off, then forward integration. Six credible licensable stacks exist. Reopens if: validated stacks consolidate to two, which would restore pricing power to the survivors.

"Fleet operations is the moat because depots are scarce." Rejected. Lyft is already selling depot operations as a service, which is the definitive test of whether a capability is a barrier. Reopens if: a major city imposes binding limits on commercial vehicle depots or curb access, converting an operating cost into a licensed scarcity.

"China's cost advantage eventually wins in the US." Rejected under current law. The tariff, the connected-vehicle rules and the absence of any US permitting path make it unavailable, and Section 14 shows the fare gap dominates the cost gap anyway. Reopens if: tariffs are lifted and connected-vehicle restrictions relaxed — at which point the vehicle layer gets cheaper for everyone and still captures nothing.

"Personal ownership of autonomous vehicles disrupts the fleet model." Out of scope on current evidence, which is not the same as rejected. No operator sells a consumer vehicle capable of unsupervised operation; Tesla's milestone framework contemplates a million robotaxis in commercial operation, not a million privately owned driverless cars. Reopens if: an FMVSS-compliant vehicle is certified for unsupervised consumer operation.

Second- and third-order effects

A fare collapse reprices car ownership before it reprices transit. The marginal urban household keeping a second car for occasional use is the first to drop it, and that decision runs straight into auto finance — loan books written against residual values that assume a stable used-car market — and into parking, which is the highest-and-best-use question sitting under a large fraction of urban land. Neither effect is visible at 2.5% penetration. Both are large and slow.

The actuarial moat compounds into a cost advantage precisely as technology cheapens. Argued in Section 17 and worth repeating because it inverts the usual logic: more entrants means more parties who must be priced against the leader's curve at a credibility loading. Falling hardware costs widen rather than narrow this gap.

One at-fault catastrophe reprices the entire class at once. Insurers underwrite classes. With so few operators, no individual experience dominates. A fatal at-fault incident at any operator raises everyone's premium and everyone's permitting friction simultaneously — a correlated risk under a sector that looks diversified.

The labour displacement is politically quiet and will not stay quiet. The Transport Workers Union already intervened in the Zoox exemption docket asking NHTSA to require disclosure of remote-operations headcount, compensation and turnover. NHTSA declined. That is the opening move in a fight over a workforce that BLS does not measure, at a scale — roughly $80 billion of platform outflow a year — that no city government can ignore once it starts falling.

Waymo is worth more inside Alphabet than outside it, and that is a problem in waiting. Other Bets lost an annualised $7.8 billion in the first half of 2026 against 0.32% of group revenue. Alphabet has funded this because it can. The February 2026 round brought in $16 billion of outside capital at a $126 billion valuation, which is a partial answer to the capital-allocation question and also a step toward a structure where outside holders will eventually want a listing, a return, and a spending discipline that a decade-long city-by-city rollout does not accommodate.

31Falsifiers and dated predictions

Dated, testable predictions

Method: Each prediction states the observable that settles it and the source where the observable will appear. Confidence reflects the quality of the evidence behind the underlying reasoning, not the desirability of the outcome. Synthesis: Taken together these describe a decade in which the technology works, the deployment compounds, and the displacement of the wage pool remains slower and more concentrated than the discourse expects.

AEvidence register

All sources below were opened and read in the preparation of this brief. Tier 1 = issuer filings and regulator or statutory publications; Tier 2 = peer-reviewed research, authoritative institutional studies and reputable wire or trade reporting; Tier 3 = single-source estimates, teardowns and company marketing claims, used only where flagged.

BNot verified — and carrying no conclusion

The following could not be confirmed against a primary source in the preparation of this brief. None supports a conclusion above, and each is listed so a reader can see the shape of what is missing.

This is an internal structural assessment of industry economics and competitive position. It is not investment advice, not a recommendation, and not a solicitation. No price targets, entry or exit levels, or position sizing are expressed or implied. Grades apply to the robotaxi position of each named entity, not to the security or to the parent company's overall valuation. Evidence cutoff: 30 August 2026.