2026-06-18 👁️ — views

Physical AI Compute Infrastructure — Tesla Dojo vs Google TPU vs NVIDIA H100 in the AV Training Arms Race

Tesla bets on Dojo custom silicon at $1/FLOP target while Waymo inherits Google TPU scale; both outpace NVIDIA-dependent rivals on training iteration speed.

Article 130 in the Physical AI Benchmark Series — Physical AI Compute Infrastructure: Tesla Dojo vs Google TPU vs NVIDIA H100/H200 in the AI Training Arms Race Behind FSD, Waymo’s Neural Nets, and Optimus Policy Learning

The AI models that power FSD, Waymo’s perception system, and Optimus policy learning are trained on massive compute clusters. Compute infrastructure determines how fast each company can iterate — how quickly it can train a new model, run ablation experiments, and deploy improvements to the fleet. Tesla bet on custom silicon (Dojo D1 chips and ExaPOD clusters). Waymo inherits Google’s world-class TPU infrastructure. Every other AV and robotics company rents NVIDIA H100/H200 clusters. This article maps compute infrastructure as a Physical AI benchmark dimension.

All figures labeled “(est.)” are derived from public market information, company disclosures, and analyst estimates rather than verified primary data.

Section 1 — Tesla Dojo: The Custom Silicon Bet

Tesla’s Dojo program is the most ambitious custom-silicon bet in the autonomous-vehicle industry. Rather than renting NVIDIA compute from cloud providers, Tesla designed its own training chip (the Dojo D1) and assembled it into ExaPOD clusters dedicated entirely to FSD training, auto-labeling, and Optimus policy learning.

The Dojo thesis rests on a straightforward cost equation: if Tesla can train a model for $1/FLOP rather than $3-4/FLOP on rented H100 clusters, each iteration of FSD costs three to four times less than the equivalent run at a competitor. Over hundreds of training runs per year, that cost advantage compounds into an iteration-speed advantage — more experiments per dollar means faster convergence to a better model. The critical uncertainty is whether Dojo meets its utilization and yield targets; custom silicon programs historically run 12-24 months late and incur higher-than-expected per-chip costs in early production (est.).

Section 2 — Waymo and Google TPU: Infrastructure Advantage by Parentage

Waymo’s compute advantage is structural rather than earned: as an Alphabet subsidiary, Waymo inherits access to Google’s TPU infrastructure — the same custom silicon that powers Google Search, YouTube recommendations, and Gemini training runs. No AV startup can match this without comparable capital investment.

The simulation multiplier is Waymo’s most underappreciated compute amplifier. A single mile of real-world driving data can be converted into thousands of simulation variants — different lighting conditions, different pedestrian behaviors, different traffic patterns, sensor noise variations. Every simulation variant is a new training example. When those examples are generated and processed by Google-scale TPU infrastructure, Waymo’s effective training data volume far exceeds its physical fleet miles. This is the structural reason why Waymo’s perception models have historically led on long-tail edge cases: the training distribution is wider than the raw mile count suggests.

Section 3 — NVIDIA: The Incumbent Every Other AV Company Depends On

For any AV or robotics company that is not Tesla and not Waymo, NVIDIA is the only viable path to scale compute. This creates a structural dependency that limits training iteration speed to whatever H100/H200 capacity the company can afford or negotiate access to.

The NVIDIA dependency creates a strategic asymmetry that compounds over time. Aurora, Zoox, Mobileye, and Wayve are all running their training workloads on rented H100 clusters. When NVIDIA announces H200 or the next Blackwell-generation chip, every one of these companies benefits equally — no one gets an advantage from hardware access. The training compute floor rises for everyone, but the ceiling remains the same: whatever the cloud market will sell you at market price. Tesla and Waymo, by contrast, control their own compute ceilings and can expand capacity through internal investment rather than cloud procurement.

The in-vehicle story is equally concentrated. NVIDIA DRIVE Orin (254 TOPS) is the dominant AV compute platform for every major AV program except Tesla (which uses its own HW4 chip at 288 TOPS) and Waymo (which uses custom in-vehicle TPU). NVIDIA DRIVE Thor at 2,000 TOPS is the announced successor, but until it ships at volume, Orin is the standard. This means NVIDIA controls not just the training-side compute dependency but also the inference-side compute dependency — creating a double leverage point over the entire non-Tesla/non-Waymo AV industry.

Section 4 — Compute as a Ramp Multiplier: Training Iteration Speed

The practical consequence of compute infrastructure differences is training iteration speed — how many model improvement cycles each company can run per month. More iterations mean faster disengagement rate improvement, faster edge-case coverage, faster adaptation to new driving environments.

Zoox deserves particular attention in this table because it is the most underappreciated compute-advantaged AV company after Tesla and Waymo. Amazon’s ownership of Zoox gives it access to AWS infrastructure at internal transfer prices — a structural advantage essentially parallel to Waymo’s TPU access. This advantage has not translated into visible commercial traction yet, but it means Zoox’s training iteration ceiling is not limited by the same budget constraints that cap Aurora’s experiments.

Section 5 — Compute Infrastructure Benchmark Scorecard

The compute infrastructure scorecard reveals a two-tier Physical AI industry. Tier 1 consists of Tesla and Waymo — both with custom silicon for training and inference, both with dedicated clusters that can expand independently of cloud market pricing, and both with in-vehicle compute that does not depend on NVIDIA. Tier 2 consists of every other AV and robotics company, all of which are structurally dependent on NVIDIA for both training and in-vehicle inference, with training iteration rates limited by cloud budgets rather than infrastructure ceiling.

The long-run implication is a training iteration gap that compounds over time. If Tesla can run 3x as many FSD training experiments per month as Aurora because Dojo costs one-third per FLOP compared to rented H100, and if Waymo can run 5x as many experiments because Google infrastructure has no capacity ceiling, then the neural network quality gap between Tier 1 and Tier 2 widens with every passing month — independent of the quality of the research teams involved. Compute infrastructure is not a sufficient condition for Physical AI leadership, but it is increasingly a necessary one.

Note: All figures labeled “(est.)” are derived from public market information, company disclosures, analyst estimates, and industry reports as of mid-2026. This article does not constitute investment advice.

Sources

• Tesla Dojo supercomputer — Tesla AI Day ↗
• Google TPU infrastructure — Google Cloud ↗
• NVIDIA H100 specifications — NVIDIA ↗
• NVIDIA DRIVE Orin automotive — NVIDIA ↗
• Waymo research and simulation infrastructure — Waymo ↗

Tags