Training Clusters

Cortex · TPU · Ascend

Tesla's Texas training compute grew from about 50,000 to more than 230,000 H100-equivalents. Training compute is now the scarce input.

Tesla training compute at Giga Texas in H100-equivalent GPUs: about 50,000 in late 2024; Cortex 1 more than 100,000 in 2026; Cortex 1 and Cortex 2 together more than 230,000 (Tesla Q1 2026 shareholder deck). Cortex 2 is designed for 500 MW, with the first 250 MW online in April 2026. TESLA TRAINING COMPUTE AT GIGA TEXAS (H100-EQUIVALENT GPUs) Tesla, late 2024 ~50,000 Tesla Cortex 1, 2026 100,000+ Tesla Cortex 1 + Cortex 2 230,000+ 500 MWCortex 2 design power; first 250 MW online April 2026
Tesla 2026
230,000+
Cortex 2 power
500 MW
Dojo
Restarted as Dojo 3
Waymo
Google TPU
Hub
Autonomy

Finding

End-to-end driving models made training compute the scarce input, and Tesla's Texas clusters quadrupled in 18 months

Tesla went from about 50,000 Nvidia H100 GPUs at Giga Texas in late 2024 to more than 230,000 H100-equivalents in 2026: Cortex 1 at more than 100,000 and Cortex 2, online since April 2026, at more than 130,000 and still ramping toward 500 MW. The same compute trains FSD and the Optimus robot. Tesla shut its Dojo supercomputer program in August 2025, then restarted work on Dojo 3 in January 2026, and plans to use its own AI5 chips in its data centers.

Other developers follow different models. Waymo trains on Google's TPU pods and runs most of its effective miles in simulation. China's developers face US export limits on Nvidia's top chips: Huawei trains on its own Ascend 910-series chips, while XPeng, Li Auto and NIO built clusters on export-compliant Nvidia parts and are moving toward in-house silicon. Nvidia sells the training and simulation platform (DGX, Omniverse, Cosmos) most others build on. Giga Texas's compute campus is covered on AustinIO.

Clusters

Autonomy training clusters

Chinese cluster sizes are the last figures reported and are likely understated today.

Autonomy training clusters
ClusterOperatorSiliconReported scaleNote
Tesla Cortex 1 and 2TeslaNvidia H100/H200230,000+ H100-equiv. (2026)Giga Texas; FSD and Optimus
WaymoAlphabetGoogle TPUNot disclosedSimulation-heavy
Huawei AscendHuaweiAscend 910B/910CNot disclosedNo US export exposure
Li AutoLi AutoNvidia (export-compliant)~5,000 H800 (last reported)Moving to in-house chips
XPengXPengNvidia H800; Turing in-house~3,000 H800 (2024, last reported)VLA models
NIO AdamNIONvidia A100~1,000 A100 (2021, last reported)Small fleet

Sources

Sources and scope

Scope
AI training infrastructure for driving and robot models
Sources
Tesla Q4 2025 and Q1 2026 shareholder decks; Converge Digest; Not a Tesla App; company disclosures; AustinIO
Updated
October 2026