The clusters
Two Nvidia Clusters at Giga Texas, and No Cortex 3 Yet
Both Cortex clusters sit on the Giga Texas campus in Austin. Tesla reports their power in its quarterly decks; it does not report facility power, cooling design or the GPU model in Cortex 2.
- Cortex 1
- Cortex 2
- Other sites
| Cluster | MW in production | Compute | Status |
|---|---|---|---|
| Cortex 1South end of Giga Texas, Austin | More than 105 | More than 100,000 H100-equivalents; Nvidia H100 | In production since late 2024 |
| Cortex 2North side of Giga Texas, Austin | More than 120 | More than 130,000 H100-equivalents (April 2026); GPU model not disclosed | Early ramp April 2026; in production by October 2026 |
| Other training computeSites not named | About 40 | Not disclosed | In production |
Cortex 1 began as about 50,000 Nvidia H100s, completed in late 2024; the 2024 plan also called for 20,000 units of Tesla hardware. Musk's August 2024 video tour showed large fans and Supermicro liquid-cooled racks.
Tesla has not announced a Cortex 3. Its plan to reach about 400 MW by the end of 2026 does not say where the added capacity goes. Fan sites report a 200 MW permit, temporary chillers and onsite Megapacks for Cortex 2, and a 500 MW build-out; Tesla's own Q3 figure of about 120 MW for Cortex 2 does not support the larger numbers. See Tesla Giga Texas on AustinIO.
Growth
Texas Training Power More Than Doubled in the First Half of 2026
| Date | MW | Notes |
|---|---|---|
| End of 2025 | About 115 | Tesla reported compute in H100-equivalents: more than 100,000 at Cortex 1 |
| End of Q1 2026 | About 135 | Cortex 2 in early ramp at more than 130,000 H100-equivalents |
| End of Q2 2026 | About 260 | Tesla switches to reporting megawatts |
| Q3 2026 | About 265 | Cortex 1 more than 105, Cortex 2 more than 120, other sites about 40 |
| End of 2026 (target) | About 400 | Location of the added capacity not stated |
Tesla cut its 2026 capital spending guidance from more than $25B to about $20B, with $13.0B spent through the first nine months. It does not break out AI compute; research and development costs rose partly from depreciation on the new compute.
Cybercab and Optimus
Tesla Does Not Publish How Compute Splits Between Cybercab and Optimus
Tesla says Cortex 2 "supports the development of both vehicle and humanoid robot autonomy software." It has given no share by cluster or by program.
| Cluster | Cybercab and FSD | Optimus |
|---|---|---|
| Cortex 1 | FSD training, the job it was built for in 2024 | Not disclosed |
| Cortex 2 | Yes, Tesla says | Yes, Tesla says |
| Share of compute | Not disclosed | Not disclosed |
- Cybercab and FSD share one model line: HW4 cars run distilled versions of the models trained for Cybercab.
- Optimus gets the AI5 computer around April 2027, Musk says.
- Headlines calling Cortex 2 an Optimus supercomputer trace to news aggregators, not to Tesla.
What happened to Dojo
Dojo Ended in 2025; the Name Moved to an Orbital Chip
- 2021: Tesla unveils Dojo and its D1 chip at AI Day.
- 2023: Dojo production begins in July.
- January 2024: Tesla pledges $500M for a Dojo cluster at Giga New York in Buffalo.
- August 2025: Tesla shuts Dojo down. Musk: "all paths converged to AI6," and Dojo 3 "lives on in the form of a large number of AI6."
- January 2026: Musk restarts Dojo 3, saying "AI7/Dojo3 will be for space-based AI compute."
- April 2026: Tesla's Q1 deck lists Dojo 3 custom silicon under development to cut training cost.
- October 2026: Dojo is not mentioned on the Q3 call. No ground Dojo cluster runs.
| Element | Specification | Role |
|---|---|---|
| D1 chip | 7 nm TSMC, 362 TFLOPS, 354 training nodes | Tesla's own training accelerator |
| Training tile | 25 D1 chips, about 9 PFLOPS | Building block for cabinets |
| Cabinet | 12 tiles with power and liquid cooling | Rack-scale unit |
| ExaPOD | 10 cabinets, about 1.1 EFLOPS | Cluster-scale unit |
Chips and Terafab
Tesla Plans Its Own Chips, but Not Yet for Cortex
- AI5: Tesla's next computer for cars and Optimus, made by TSMC and Samsung. Design finished April 2026; volume around 2027. Tesla claims 50 times the performance of AI4.
- AI6: made by Samsung under a $16.5B contract, for cars, Optimus and data-center training.
- D3 (Dojo 3, also called AI7): a chip for orbital compute, shown on Terafab slides powering 100 kW "AI Sat Mini" satellites launched by Starship. No published specifications.
- Terafab: a chip-fab project of Tesla and SpaceX, launched March 2026. A research fab next to Giga Texas commissions tools in the first half of 2027, with a first test chip around mid-2027 and volume production late 2028, Musk says. SpaceX plans a larger plant in Grimes County, Texas. See Giga Texas Terafab and Terafab Grimes County on AustinIO.
Musk says about 20% of Terafab output would serve ground uses such as FSD and Optimus, and the rest space and orbital data centers. Tesla has not said Cortex will run Tesla-made chips; AI6 is the only Tesla chip it has named for data-center training, and it comes from Samsung, not Terafab.
Inference
Driving Runs in the Car; Grok Runs on SpaceX Servers
- FSD and Cybercab: inference runs on the car's own computer, AI4 today. Tesla has published no figures for data-center inference serving the robotaxi fleet.
- Fleet compute: on using idle cars and Megapacks as distributed compute, Musk said it is "still early days," with more clarity in the first half of 2027.
- Grok in Tesla vehicles: runs on SpaceXSI servers, not in the car. See SpaceX Colossus.
See Vehicles and Devices.
Buffalo and SpaceX
The Buffalo Dojo Pledge Is Unresolved; No Shared Clusters With SpaceX
- Buffalo: New York's economic development agency called Dojo central to extending the Giga New York lease. As of November 2025 no new lease was signed, and the state still wanted computing placed in Buffalo.
- SpaceX: Tesla invested about $2B in xAI in January 2026; SpaceX bought xAI in February, turning that into a SpaceX stake. Tesla and SpaceX share Terafab but no training clusters have been reported as shared. Musk said on October 20, 2026 that combining Tesla and SpaceX has merit and special committees are studying it.
Record
Dataset Record
- Dataset
- Tesla training clusters and compute by quarter; Cortex 1 and 2 also listed under live capacity on Largest Data Centers in the US
- Sources
- Tesla quarterly updates and earnings calls, SEC filings, TechCrunch, Interesting Engineering, TechRadar, Not a Tesla App, TeslaNorth, Converge Digest, Manufacturing Dive, Spectrum News Buffalo, WRVO
- Last verified
- October 2026