Nvidia assembled Europe's largest autonomy stack in a single release on October 21: Bolt's robotaxi programme now runs on the Alpamayo reasoning software and the Hyperion compute architecture, Mercedes-Benz's European competitors are joining the same reference design, and Einride becomes the first operator to test the reasoning stack on heavy-duty trucks. What the release does not contain matters more — no fleet order, no operating permit, no revenue line. Europe's autonomy bottleneck is an authorisation, and Nvidia is buying the layer those authorisations are written against.
Nvidia data center revenue, Q2 FY2027
$89.0B
Of $96.2B total revenue; quarter ended July 26, 2026
Nvidia revenue outside the data center
$7.2B
Automotive and everything else, up from $5.6B a year earlier
Lucid vehicles Bolt targets in Europe
25,000
Deployment target, not a disclosed purchase order (Sept. 17, 2026)
Vehicles under Spain's first Level 4 permit
20
Each with an in-car specialist; authorisation issued Sept. 10, 2026
Mobileye revenue, Q2 2026
$508M
Flat year over year; quarter ended June 27, 2026
Lucid production, Q3 2026
2,954
Down from 4,774 in Q2 after the AMP-1 second shift was removed
October 21, 2026
What Nvidia actually announced, and what it deliberately deferred
The October 21 release does three things at once. It puts Bolt's European ride-hailing programme on the Alpamayo reasoning software and the Hyperion compute and sensor architecture, starting with future Lucid vehicles and extending to vehicles from multiple manufacturers. It places Mercedes-Benz, Volkswagen Commercial Vehicles and Volvo Cars on the same stack for L3 and L4 work, plus TIER IV with the Technical University of Munich for autonomous public transport in Germany. And it assembles the certification layer underneath: Aptiv, Valeo and Schaeffler building Hyperion-based computing platforms, and dSPACE, IVEX and TÜV Rheinland collaborating on Halos SAFE evaluation workflows.
- The commercial frame is enormous: at least 25,000 Level 4-ready Lucid vehicles on Hyperion, Germany and France named as priorities, and a stated ambition of 100,000 autonomous vehicles across Europe and the wider UNECE region by 2035.
- What is not committed: no purchase quantity for those 25,000 vehicles is disclosed, and the first Madrid milestone is data collection in 2027 that is explicitly conditional on regulatory approval.
- The first milestone is narrow: Bolt's autonomous-driving unit leads execution, Bolt owns and operates the fleet, and the starting point is a software-and-data programme rather than a service.
- Scale of the prize is not the point: everything Nvidia sells outside the data center, automotive included, was $7.2 billion in the quarter against $96.2 billion of total revenue.
The architecture detail is where the strategy sits. Alpamayo 2 Edge is a two-billion-parameter reasoning system optimised for DRIVE AGX Thor, running at 10Hz for in-vehicle scene understanding and motion planning, and an open blueprint for automating data preparation, post-training and simulation evaluation is due in early 2027. Nvidia is handing the intelligence layer away to make its own silicon the default substrate — because a paid licence keeps operators on somebody else's stack, while an open one makes every approved European vehicle a Nvidia vehicle.
Regulation
Spain's first Level 4 permit covers twenty vehicles — and that is the real market
Spain's Directorate General for Traffic issued the country's first national Level 4 authorisation on September 10, 2026. It went to WeRide's GXR vehicles operated with Uber and AVOMO, covers twenty vehicles that each carry an in-car specialist, and expects commercial operations by the end of the year. That is the unit of scarcity in European autonomy: not a car, but a regulator's signature on a safety case, attached to a named vehicle count and a named operator.
| Operator and stack | Where it operates | What is authorised today | Next milestone |
|---|---|---|---|
| WeRide with Uber and AVOMO | Madrid | Spain's first national Level 4 permit; 20 vehicles, each with an in-car specialist | Commercial operations expected by end-2026 |
| Bolt with Nvidia on Lucid hardware | Madrid, then Germany and France | None yet; 2027 start is data collection only, conditional on approval | Madrid driving-data collection in 2027 |
| Waymo, Alphabet's driverless unit | Madrid | Manual road testing and technical validation only | First riders targeted in 2027 |
| Einride with Nvidia | Germany | First cab-less Level 4 truck in regular public-road service for Lidl | Alpamayo road evaluation in the second half of 2027 |
| Mobileye with MOIA | Hamburg | Public user testing with safety drivers | Own operation: about 100 vehicles in a U.S. city in 2027 |
| Tesla FSD | Netherlands and seven other EU states | Driver-assistance type approvals, not a driverless service permit | Vote across all 27 member states expected |
Every other European programme is queued behind that signature. Waymo drives manually in Madrid while preparing for 2027 riders; Bolt collects data in 2027 if it is approved; Einride's cab-less truck runs a narrow German lane for Lidl while Alpamayo is tested in shadow mode before a road evaluation in the second half of 2027. Tesla's European approvals are a different instrument entirely — type approvals for supervised driving, which unlock a consumer feature rather than a driverless service.
Mechanism
Why the scarce asset is the safety case, not the permit
Nvidia cannot buy a permit; permits go to operators, and the one granted so far in Spain went to a consortium that does not use Hyperion. What Nvidia can buy is the position that makes permits easier to obtain. Because Halos SAFE, the Alpamayo reasoning stack and the Hyperion architecture become the reference design across automakers, tier-1 suppliers, validators and operators, a safety case approved for one European city becomes reusable for the next — and every vehicle inside it carries Nvidia silicon and software.
Nvidia revenue by platform, Q2 fiscal 2027
Total revenue of $96,221M against data center revenue of $89,023M; the $7,198M remainder covers gaming, professional visualization, automotive and OEM work, up from $5,647M a year earlier.
Unit: $ millions
Data center
Up 117% year over year
89,023
Everything else, automotive included
Up from $5,647M a year earlier
7,198
Robotaxis are the first mass-market application of robotics and physical AI, reinventing how people and goods move.
The squeeze
Mobileye's licence model is now written above its own layer
Mobileye's second quarter was flat in a way that flatters nobody. Revenue of $508 million was unchanged from $506 million a year earlier even as systems shipped rose 3%, which is price erosion, and $93 million of the $155 million adjusted operating income came from Israel's statutory research credit rather than from trading. Full-year guidance carries a GAAP operating loss of roughly $4.1 billion, which includes a $3.79 billion non-cash goodwill impairment taken in the first half, and the founder-chief executive is stepping down once a successor is appointed, with Intel still holding about 77% of the company.
| Dimension | Mobileye | Nvidia |
|---|---|---|
| Software model | Licensed Drive and Chauffeur stacks priced per vehicle | Open Alpamayo reasoning stack plus proprietary Level 4 software on a shared compute architecture |
| Safety evidence | Own Drive validation; MOIA public testing with safety drivers in Hamburg | Halos SAFE workflows with dSPACE, IVEX and TÜV Rheinland across OEMs, tier-1s and operators |
| European authorisation path | Public user testing only; no driverless service permit held | Madrid data collection from 2027, subject to approval, across Germany and France priorities |
| Scale of ambition | About 100 vehicles in a U.S. city in 2027, roughly 17,000 over five years; MOIA targets more than 100,000 by 2033 | At least 25,000 European vehicles targeted, 100,000 by 2035 |
| Revenue base | $508M in Q2 2026, flat year over year, $1.3B cash and marketable securities | $96.2B in Q2 FY2027; $26.0B returned to shareholders, $99.0B of buyback authorisation left |
- What Mobileye is doing about it is moving downstream: it has committed to owning and operating fleets, launching in a major U.S. metro in 2027 with about 100 vehicles and scaling toward roughly 17,000 over the following five years.
- That is a demand strategy born of a supply squeeze, and it puts Mobileye in direct competition with the operators it used to supply for the same approvals.
- The installed base is not the problem — more than 250 million vehicles carry its technology across over 1,400 models — but the OEMs it sells into are now announcing L3 and L4 work on Hyperion and Alpamayo.
- Horizon: the chief executive succession and its own 2027 launch city decide whether Mobileye can be supplier and operator at once.
Demand
Uber owns the riders and is paying for every rival stack at once
Uber is the only party able to sell autonomy to all of these stacks simultaneously, and it is doing so with rivals on both sides. Second-quarter gross bookings reached $58.0 billion, revenue $14.2 billion and free cash flow $2.8 billion, with trailing free cash flow passing $10 billion for the first time. It has agreed to deploy Nvidia's own proprietary Level 4 software across 28 cities by 2028, to fund up to 50,000 in-house Rivian R2 robotaxis with $1.25 billion of investment, and to run Nuro software on at least 35,000 Lucid vehicles.
| Stack provider | Vehicles | Timing | Software owner |
|---|---|---|---|
| Nvidia | 28 cities globally | Los Angeles and San Francisco from the first half of 2027, data-collection vehicles first | Nvidia proprietary Level 4 software on Hyperion and Alpamayo |
| Rivian | Up to 50,000 R2 robotaxis, 10,000 in phase one | San Francisco and Miami in 2028; 25 cities by 2031 | Rivian's own stack: two 1,600-TOPS chips, 11 cameras, five radars, one lidar; exclusive to Uber |
| Lucid with Nuro | At least 35,000 Lucid vehicles | San Francisco Bay Area launch, Houston in 2027 | Nuro driving software on Lucid hardware |
| WeRide with AVOMO | 20 vehicles in phase one | Madrid, commercial operations expected by end-2026 | WeRide GXR vehicles, operated with Uber and AVOMO |
The vehicle
Lucid's 25,000-vehicle headline rests on a factory making 2,954 cars a quarter
Lucid is the vehicle story and the weakest balance sheet in it. Second-quarter revenue of $405 million came against cost of revenue of $832 million, a GAAP net loss of $1,035 million, and $3.0 billion of total liquidity that management expects to carry the company into 2027. Third-quarter production then fell to 2,954 vehicles from 4,774 after the second shift at AMP-1 was removed in June, and deliveries of 3,806 exceeded production as inventory was drawn down.
- The programme is a target, not a contract: no purchase quantity for the 25,000 vehicles has been disclosed, the fleet sits on a Midsize platform that has not shipped, and the robotaxi effort has been carved into a separate Lucid Technologies unit with roughly 100 validation vehicles running with Uber and Nuro in the San Francisco Bay Area and Houston.
- The cash arithmetic is unforgiving: trailing free cash outflow of $6.6 billion against $732.6 million of cash and $28.7 million of short-term investments at June 30, so the ramp, if it comes, is funded by Bolt or by fresh capital rather than by Lucid.
- Horizon: the third-quarter call on November 9, 2026 is the first date on which an order book rather than a memorandum would show up.
The counter-case
Freight proves the reasoning stack before cities have to
Einride's thread is the least glamorous and the most informative. On October 21 Reuters reported that Einride will be the first operator to deploy Alpamayo, starting in shadow mode, with the reasoning system running alongside its existing driving software and human drivers without controlling the vehicle, and live road evaluation only in the second half of 2027. The company also put the first cab-less Level 4 truck into regular public-road service for Lidl in Germany in September, runs a fleet of hundreds of electric trucks across the U.S., Europe and the Middle East, and expects 1,500 to 2,000 vehicles by 2028 with roughly 80% of captured demand judged suitable for automation in the medium term.
- Freight is the cheaper regulatory problem: a fixed corridor with a repeatable safety case, validated on a narrow German permit, before anyone argues about mixed traffic in a European capital.
- The training stack is already contracted: Cosmos for camera-data curation and synthetic edge cases, and Blackwell at scale with an Exemplar Cloud partner.
- Horizon: the second half of 2027 road evaluation is the single most informative date on the European calendar, because it decides whether a reasoning stack clears a revenue-generating route at all.
Supply chain
Who captures the money if the permits do come
| Layer | Named companies | Evidence of linkage |
|---|---|---|
| Compute and reference architecture | Nvidia — DRIVE AGX Thor, Hyperion, Cosmos | Alpamayo 2 Edge runs at 10Hz on Thor; Cosmos curates camera data for Einride |
| Tier-1 electronics | Aptiv, Valeo, Schaeffler | Named as building Hyperion-based computing platforms; Aptiv second-quarter revenue $3.27B, gross profit $775M |
| Safety validation and certification | dSPACE, IVEX, TÜV Rheinland; JLR, AVL, Deepen AI, Foretellix, 51Sim, aiMotive | Halos SAFE evaluation workflows covering simulation, hardware-in-the-loop and on-road testing |
| Vehicle platforms | Lucid Midsize platform; Mercedes-Benz, Volkswagen Commercial Vehicles, Volvo Cars, TIER IV with TU Munich | Named in the October 21 announcement as building on Hyperion and Alpamayo |
| Fleet operations | Bolt (more than 850 cities), Einride (hundreds of trucks) | Bolt owns and operates the fleet; Einride's cab-less truck carries Lidl freight |
| Demand aggregation | Uber | 28-city Level 4 network, $58.0B of second-quarter gross bookings, $10B+ trailing free cash flow |
| Authorisation and evidence | Spain's DGT and the Region of Madrid; Waymo, Alphabet | First national Level 4 permit for 20 vehicles; 400M+ autonomous km and 95% fewer serious-injury crashes than human drivers |
If authorisation is the constraint, the scarce layers are the safety case, the reference architecture and the operating permit, and the value sits with the platform that owns the evidence trail, the validators who sign it, the operators who hold the permit and the marketplace that converts demand into vehicles. Lucid, Mercedes-Benz, Volkswagen Commercial Vehicles and Volvo Cars supply the body and the brand; they do not supply the approval. The toll on European autonomy will be collected in silicon and signatures, not in metal — which is why the company that sells the hardware to everyone is also the one trying to write the rules everyone must obey.
Horizons
What to watch, and when
Days to quarters: the next evidence is regulatory and calendared, not financial. The Madrid pilot must convert twenty supervised vehicles into paying rides by the end of 2026, Lucid's third-quarter call on November 9 has to produce an order rather than a target, and Nvidia's third-quarter fiscal 2027 print lands against a $108.0 billion outlook that assumes data center does the work.
- Through 2027: Bolt's Madrid data-collection approval, Waymo's first riders in Madrid, and the EU-wide vote across 27 member states on Tesla's supervised driving approvals.
- 2027 to 2028: whether the Bolt and Lucid programme converts into a purchase order and a German or French operating permit, and whether Uber's 28-city Level 4 launch actually reaches Europe.
- 2028 and beyond: whether the permit queue stays open. Rivian's exclusive 50,000-vehicle deal and Mobileye's own roughly 17,000-vehicle plan are both bets that operators, not platforms, capture the margin.
What would break this reading: a second permit-heavy market opening up, which would turn Europe back into a fleet and cost race and make the certification layer a footnote; an incumbent bundling approvals with hardware, which Waymo's 400 million-kilometre evidence base suggests is possible; or a European operator deciding that safety certification is worth owning outright, which is precisely the move Mobileye has just announced. The thesis holds only while approvals remain scarce and slow, because that is the only condition under which owning the evidence trail is worth more than owning the vehicle.
The autonomy stack, and who owns the queue position
- Owns the certification layer every European Level 4 permit will cite: Hyperion compute, Alpamayo reasoning software and Halos SAFE workflows with dSPACE, IVEX and TÜV Rheinland.
- Horizon: Bolt's Madrid data collection in 2027 and Uber's 28-city launch from the first half of 2027 are the proof points that turn a reference stack into permits.
- The risk to the thesis: operators are already hedging with in-house silicon, so the slot may never be exclusive.
- Distributes every rival stack — Nvidia's proprietary Level 4 software across 28 cities, Rivian's in-house 1,600-TOPS platform, Nuro software on Lucid, WeRide in Madrid — while booking $58.0B of quarterly gross bookings.
- Carries the permit cycle twice: a $1.6B pre-tax equity revaluation gain in Q2 means autonomy marks already move GAAP earnings, and permits determine future supply.
- Horizon: the 28-city rollout from the first half of 2027 and 50,000 Rivian units from 2028 decide whether neutrality still pays.
- Flat $508M of Q2 revenue on 3% more units shipped, with $93M of $155M adjusted operating income coming from a statutory research credit, marks a licence model repricing.
- Writes its own downside: a $3.8B first-half goodwill impairment, a founder-chief executive exit with Intel at roughly 77%, and a pivot to owning about 100 vehicles in 2027.
- Horizon: the successor appointment and its own 2027 launch city decide whether it can be supplier and operator simultaneously.
- A 25,000-vehicle target with no disclosed purchase quantity, on an unshipped platform, against $6.6B of trailing free cash outflow and $732.6M of cash.
- Third-quarter production fell to 2,954 vehicles from 4,774 and deliveries again exceeded production, so inventory is funding the interim with the AMP-1 second shift gone.
- Horizon: the November 9, 2026 call is the first checkpoint for an order book; until then the fleet is a memorandum.
- Waymo enters Madrid with more than 400M autonomous kilometres and 95% fewer serious-injury-or-worse crashes than human drivers, evidence a permit queue rewards incumbents.
- Horizon: first riders in Madrid in 2027 put it ahead of Bolt's data-collection phase and clear of the European reference stack.
- Optionality rather than cost: $445.9B of trailing revenue means Waymo's European build-out does not need external funding.
- Type approvals in the Netherlands and seven other EU states unlock supervised driving, not a driverless service, so the EU-wide vote across 27 states does not solve the permit problem.
- Horizon: an EU-wide approval without a European operating partner would leave the 2027 Madrid race to Waymo, Bolt and WeRide.
- Watch for a European operating partner or city-level authorisation application; neither is announced.
- Holds Europe's first national Level 4 permit alongside Uber and AVOMO, covering 20 supervised vehicles with commercial rides expected by end-2026.
- But it monetises slowly: second-quarter revenue of 232 million yuan against a net loss of 401 million yuan, on trailing revenue of 831 million yuan.
- Horizon: whether the 20-vehicle phase converts into paid rides by year-end, and whether a permit held by a loss-making operator transfers value to shareholders.
- Owns the one stack that needs no Nvidia: two in-house 1,600-TOPS chips, with Uber's option on up to 50,000 R2 robotaxis exclusive to its platform.
- Horizon: Uber's $300M initial tranche, itself conditional on regulatory approval, and 2028 launches in San Francisco and Miami are the checkpoints.
- Offsets the autonomy optionality: trailing free cash outflow of $3.2B on $5.9B of trailing revenue.
