The Bet That Broke and the Bet That Spread: Tesla and Nvidia, Three Years Later

When I first wrote about the split between Tesla and Nvidia, the argument was simple enough to fit in a sentence. Tesla wanted to own the whole stack. Nvidia wanted to sell that stack to everyone else. Three years on, both companies have been tested against reality, and the interesting part is not that one was right. It is that each has been proven right about a different thing.

The vertical bet ran into physics and payroll

Tesla's position was always that autonomy is a compute problem before it is a driving problem. Train a large enough model on enough real-world video, and the driving behavior falls out of it. That logic pushed Tesla toward building its own training silicon, because renting compute from Nvidia meant paying a margin to a supplier on the single input that mattered most.

That was the case for Dojo. It did not survive contact with the schedule. In August 2025 Tesla disbanded the Dojo team, its lead Peter Bannon left, and roughly twenty engineers walked out to found a startup of their own. Musk's explanation was that every development path had converged on the AI6 chip, which made the second-generation Dojo design an evolutionary dead end.¹ The reversal came weeks after he had projected a second Dojo cluster running at scale in 2026, and days after Tesla signed a 16.5 billion dollar agreement with Samsung for AI6 production.²

The plan then reassembled itself around a different axis. In January 2026 Musk revived the concept as "Dojo 3," essentially a large number of AI6 chips on a single board rather than a purpose-built training architecture.³ The AI5 chip taped out in late March 2026, dual-sourced between Samsung's Taylor fab in Texas and TSMC in Arizona, with volume production somewhere between late 2026 and 2027.⁴ Worth noting for calibration: Musk said in June 2024 that AI5 would be in vehicles in the second half of 2025.⁵

So the vertical strategy did not fail. It arrived late and in a different shape, and during the delay Tesla went back to buying Nvidia and AMD compute. That is the honest summary.

The horizontal bet found its customers

Nvidia spent the same three years doing what a supplier does when the market is uncertain: selling to everyone. The current product is DRIVE AGX Hyperion 10, a reference architecture pairing two Blackwell-based DRIVE AGX Thor computers with a validated sensor suite of fourteen cameras, nine radars, one lidar, and twelve ultrasonics. Each Thor delivers over 2,000 FP4 teraflops.⁶

The customer list is the strategy made visible. Stellantis, Lucid, and Mercedes-Benz signed on for Level 4-ready passenger vehicles; Aurora, Volvo Autonomous Solutions, and Waabi for freight.⁷ At GTC 2026 Nvidia added BYD, Hyundai, Nissan, and Geely, and announced an Uber partnership targeting robotaxis in twenty-eight cities across four continents by 2028, beginning in Los Angeles and San Francisco.⁸ By GTC Taipei in June, Foxconn was building Level 4 fleets in Kaohsiung, and Uber and Autobrains were preparing a Munich robotaxi program.⁹

Note the sensor line in that spec: one lidar, nine radars. Nvidia's platform assumes fusion. Tesla still assumes cameras alone. The two architectures have not converged, and the reason matters more now than it did in 2023.

Germany's answer, three years later

The manuscript I wrote in 2023 asked whether a European incumbent could build Tesla-grade AI infrastructure. Volkswagen has now answered, and the answer is that it decided not to try alone.

In April 2026 MOIA America — the renamed Volkswagen ADMT — began on-road validation in Los Angeles with about ten autonomous ID. Buzz vehicles, scaling toward a hundred-plus, with commercial rides on Uber's platform planned for late 2026 and driverless operation in 2027.¹⁰ The vehicles run Mobileye's sensor suite: cameras, radar, and lidar.¹¹

That is not a company matching Tesla's compute spend. It is a company buying perception from a supplier, buying demand from Uber, and keeping the vehicle. Whether that reads as capitulation or as sound capital allocation depends on how you score the next five years — and given how the Dojo story went, the question is no longer rhetorical.

Where the scoreboard actually stands

Tesla crossed ten billion cumulative FSD miles in May 2026, and the fleet now generates roughly a million miles of data per day across more than a million active FSD subscribers.¹² Nothing else in the industry produces data at that rate. The flywheel argument is intact.

The driverless numbers tell a different story. Waymo runs roughly 3,000 robotaxis across eleven metros, about 500,000 paid rides a week, near 200 million fully autonomous miles. Tesla's unsupervised fleet is roughly twenty vehicles across Austin, Dallas, and Houston, with about 1.7 million cumulative paid robotaxi miles as of Q1 2026, plus recent launches in Miami, Orlando, and Tampa.¹³ Musk has tied real scale to an FSD v15 rewrite expected late 2026 or early 2027.

And the sensor question is now a regulatory one. In March 2026 NHTSA escalated its FSD investigation to an Engineering Analysis — the last formal step before a mandatory recall — covering about 3.2 million vehicles, finding that the system failed to detect or warn appropriately under degraded visibility such as glare and airborne obscurants.¹⁴ In nine reviewed crashes, one fatal, the system lost track of or never detected a lead vehicle. Those are exactly the conditions redundant sensing is meant to cover.

What the split means now

The 2023 framing was vertical versus horizontal. That still holds, but the sharper version is this: Tesla is betting that generality wins, Nvidia is betting that standardization wins.

Tesla's thesis is that one model, trained on enough data, generalizes to any road — and that when it works, deployment is a software push rather than a city-by-city hardware campaign. Nvidia's thesis is that autonomy will be a certified, validated, sensor-redundant product that most companies buy rather than build, and that the supplier of the common foundation captures the industry regardless of which brand wins.

Nvidia's position is structurally more comfortable. When Tesla builds its own chip, Nvidia loses a customer. When Tesla's schedule slips, Nvidia gets that customer back. Dojo's collapse and revival happened entirely inside Nvidia's win condition.

But the original point in my manuscript survives the update, in inverted form. If Tesla — with the largest real-world driving dataset ever assembled, a decade of in-house silicon, and full control of hardware and software — still needed two extra years and an external foundry to ship a training chip, then the difficulty facing everyone else is not a matter of will. It is structural. Volkswagen's decision to buy perception rather than build it is not timidity. On the current evidence, it is arithmetic.

Watching this from Korea, the uncomfortable question is not which strategy is better. It is which one is available to us. Building the Tesla stack requires capital most companies cannot commit and a tolerance for multi-year schedule failure most boards will not grant. Buying the Nvidia stack is achievable, and it means competing on integration and manufacturing rather than on the intelligence layer. Neither of those is a comfortable answer. Pretending there is a third one is worse.

For research inquiries or collaboration, contact: ceo@technorns.com

Endnotes

1. TechCrunch, "Elon Musk confirms shutdown of Tesla Dojo, 'an evolutionary dead end'" (Aug 11, 2025). https://techcrunch.com/2025/08/11/elon-musk-confirms-shutdown-of-tesla-dojo-an-evolutionary-dead-end
2. TechCrunch, "Tesla Dojo: The rise and fall of Elon Musk's AI supercomputer" (Sep 2025). https://finance.yahoo.com/news/tesla-dojo-rise-fall-elon-161846832.html
3. Reporting on Musk's January 2026 Dojo 3 restart. https://www.mexc.com/news/508091
4. Tom's Hardware, "Tesla's AI5 with 2nm-class node tapes out at Samsung Foundry." https://www.tomshardware.com/tech-industry/artificial-intelligence/teslas-ai5-with-2nm-class-node-tapes-out-at-samsung-foundry-production-starts-soon-months-after-tsmc-tape-out
5. Electrek, "Tesla taped out AI5 chip, Musk says — nearly 2 years behind schedule" (Apr 15, 2026). https://electrek.co/2026/04/15/tesla-ai5-chip-taped-out-musk-ai6-dojo3/
6. NVIDIA Newsroom, DRIVE AGX Hyperion 10 / Thor specifications. https://nvidianews.nvidia.com/news/nvidia-uber-robotaxi
7. NVIDIA investor release, "NVIDIA Makes the World Robotaxi-Ready With Uber Partnership" (Oct 28, 2025). https://investor.nvidia.com/news/press-release-details/2025/NVIDIA-Makes-the-World-Robotaxi-Ready-With-Uber-Partnership-to-Support-Global-Expansion/default.aspx
8. GTC 2026 AV program expansion. https://winbuzzer.com/2026/03/17/nvidia-gtc-2026-uber-robotaxi-physical-ai-drive-hyperion-xcxwbn/ and https://www.electrive.com/2026/03/17/nvidia-partners-with-byd-geely-hyundai-isuzu-nissan-and-uber/
9. NVIDIA, "DRIVE Hyperion Becomes the Global Platform for a Robotaxi-Ready World" (Jun 1, 2026). https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-DRIVE-Hyperion-Becomes-the-Global-Platform-for-a-Robotaxi-Ready-World/default.aspx
10. TechCrunch, "Volkswagen begins testing its self-driving microbuses in Los Angeles" (Apr 8, 2026). https://techcrunch.com/2026/04/08/volkswagen-moia-uber-los-angeles-testing-self-driving-microbuses-id-buzz/ ; Volkswagen Group release: https://www.volkswagen-group.com/en/articles/moia-america-to-deploy-autonomous-id-buzz-vehicles-on-the-uber-platform-in-los-angeles-by-the-end-of-2026-20306
11. ID. Buzz AD sensor configuration (Mobileye). https://malaymail.com/news/life/2025/06/22/volkswagen-challenges-waymo-with-launch-of-electric-id-buzz-autonomous-robotaxi-fleet-in-los-angeles/181272
12. FSD 10 billion mile milestone and daily data rate. https://www.tesery.com/blogs/news/fsd-hits-1-million-miles-per-day-how-teslas-global-fleet-is-training-austins-robotaxis ; subscriber and fleet figures: https://thechargeport.com/robotaxi-tracker
13. Robotaxi fleet and mileage comparison. https://thechargeport.com/robotaxi-tracker ; https://newmarketpitch.com/blogs/news/autonomous-vehicle-tesla-losing-robotaxi-race
14. NHTSA Engineering Analysis EA26002. https://www.techtimes.com/articles/320170/20260711/waymo-driverless-four-more-cities-fleet-14-times-teslas-sensor-debate-hits-road.htm