Giga Texas Cortex Training Cluster

Tesla · Cortex 1 and 2 · Giga Texas, Travis County

More than 230,000 Nvidia GPUs in two clusters train FSD, Robotaxi and Optimus on the campus that builds the cars and robots. About 260 MW today, nearly 400 MW by the end of 2026.

  • Cortex 1 and 2
  • Other Giga Texas programs
Schematic of the Giga Texas campus, not to scale: the main factory, cell and cathode lines, Cortex 1 and 2, the Optimus Factory, the Terafab Research Fab, RFTX and the Cybercab test track COLORADO RIVER SH 130 SCHEMATIC, NOT TO SCALE Giga Texas main factory: Model Y, Model Y L, Cybertruck and CybercabMAIN FACTORY 30 MW ROOFTOP SOLAR Tesla 4680 cell line4680 CELLS Cathode plant for 4680 cells, operating at scale since August 2026CATHODE PLANT Roadster Factory Texas, with a looping vertical trackRFTX Roadster looping vertical track LOOP Optimus Factory, building Optimus 4OPTIMUS FACTORY Cortex 1 AI training clusterCORTEX 1 Cortex 2 AI training clusterCORTEX 2 Megapack and switchyard beside Cortex 2, toward about 250 grid-buffer units MEGAPACK YARD Terafab Research Fab, a joint Tesla and SpaceX fab on the east side of the campusTERAFAB RESEARCH FAB TESLA + SPACEX Cybercab test track, a multi-lane asphalt oval CYBERCAB TEST OVAL Oncor Giga TX substation, 138 kV and upONCOR
GPUs
230,000+ H100e
Power now
~260 MW
End-2026 target
~400 MW
Full design
~630 MW
Cortex 2 online
April 2026
Sector
AI and data centers

Overview

Two Clusters Training FSD and Optimus

Cortex is Tesla's AI training program at Giga Texas. Two Nvidia GPU clusters, Cortex 1 and Cortex 2, train the networks behind Full Self-Driving, the Tesla Robotaxi service and Optimus, on the same campus that builds the vehicles and robots that run them.

Tesla reports its Texas training capacity in megawatts: about 115 MW at the end of 2025, 135 MW at the end of Q1 2026 and 260 MW at the end of Q2 2026, most of the increase from Cortex 2. It targets nearly 400 MW by the end of 2026.

200 400 600 End 2025: about 115 MW115 MWEnd 2025 End Q1 2026: about 135 MW135 MWEnd Q1 2026 End Q2 2026: about 260 MW260 MWEnd Q2 2026 Target, end 2026: about 400 MW400 MWTarget, end 2026 Full design: about 630 MW630 MWFull design

Reported   Target   Full design

The clusters at the end of Q2 2026
ClusterLocationGPUsPowerWorkloadStatus
Cortex 1South end, beside the main factoryMore than 100,000 H100eAbout 130 MWFSD and robotaxi trainingIn production since Q4 2024
Cortex 2North endMore than 130,000 H100eAbout 130 MW active of 500 MW designVehicle and humanoid autonomyOnline April 2026; ramping
CombinedGiga TexasMore than 230,000 H100eAbout 260 MW; nearly 400 MW targeted by end of 2026FSD, robotaxi and OptimusRamping

H100e counts the mix of H100 and H200 GPUs in H100 units.

Cortex 1

FSD and Robotaxi Training

Cortex 1 came online in Q4 2024 with 50,000 H100s and grew through 2025 to more than 100,000 H100e. Tesla credits it with a fivefold increase in training compute, which carried FSD from hand-written C++ in v11 to the single end-to-end network of v12. It now trains v14 and v15. It draws about 130 MW, cooled by six chilled-water pipes, four second-floor tanks and the rooftop fans drone observers film.

FSD versions
VersionArchitectureStatus, October 2026
v11Modular C++ codeSuperseded
v12End-to-end neural network, first trained at Cortex scaleSuperseded
v14About 1 billion parametersConsumer vehicles on v14.3.x; v14 Lite on HW3
v15Rebuilt, about 10 billion parametersTesting in the Robotaxi fleet; HW4 release targeted late 2026 or early 2027

v15 runs on current HW4 computers with 16 GB of memory, so owners will not need AI5. Tesla tied round-the-clock robotaxi operation to v15 validation. Robotaxi launched in Austin in June 2025 and has spread to Dallas, Houston, Miami, Orlando, Tampa and the Bay Area, with 2.4 million paid miles through Q2 2026; the Cybercab began paid rides in Austin on September 4, 2026. See Texas Robotaxi Deployments.

Cortex 2

Vehicle and Humanoid Autonomy

Cortex 2, at the north end, was permitted as a "Central Campus Support Facility" and redesignated in mid-2025 as one large data center. It came online in April 2026 with more than 130,000 H100e; its first phase is 250 MW and its full design 500 MW. Beside it, Tesla's Megapack and switchyard installation is growing toward about 250 units.

Its heaviest workload is the neural world simulator Tesla AI vice president Ashok Elluswamy described at ICCV 2025: a model trained on fleet video that generates training environments in place of a hand-coded physics engine, producing thousands of synthetic variations a day.

The loops

Two Training Loops, One Campus

FSD loop, Cortex 1

  1. About 4 million vehicles upload selected clips and telemetry
  2. Cortex 1 trains in cycles of about 70,000 GPU hours
  3. New weights go to the fleet over the air
  4. Edge cases from the road flow back

Optimus loop, Cortex 2

  1. Human demonstration video, fleet footage and Optimus telemetry flow in
  2. Cortex 2 trains with neural world simulator output
  3. New weights go to Optimus units over the air
  4. Units at Fremont and Giga Texas send manipulation data back

Both loops share Tesla's fleet-data ingestion and over-the-air distribution. Fleet video trains Optimus's spatial understanding, and world simulator environments serve FSD.

After Dojo

Nvidia for Training, Tesla Chips for Inference

Tesla disbanded Dojo, its four-year effort to train on its own D1 chip, in August 2025; Musk called it an "evolutionary dead end" because Tesla could not keep pace with Nvidia's release cadence or performance per watt. Its people moved to Cortex, and Tesla's own chip work focused on inference: AI5 from Samsung Taylor in volume from mid-2027, and later generations through the Terafab Research Fab.

Training and inference silicon
TierSiliconWorkload
Training, FSDNvidia H100 and H200 at Cortex 1Vehicle autonomy
Training, vehicle and humanoidNvidia H100 and H200 at Cortex 2Vehicle and humanoid models
Inference, vehiclesTesla AI4 and AI4.5 today; AI5 in volume from mid-2027On-board FSD
Inference, OptimusTesla AI5, then AI6 from the Terafab pipelineOn-robot inference

Power

The Campus's Largest Load

At nearly 400 MW, AI training would draw more than the vehicle factory at full production. Full design for both clusters is about 630 MW. The campus offsets part of it with a 30 MW rooftop solar array and the Cortex 2 Megapack yard.

On August 3, 2026, Governor Abbott told the Public Utility Commission and ERCOT to stop advancing data center interconnections until pending projects are audited; Cortex 2 was already operating. On September 21 he halted new TCEQ data center permits until the audit ends, expected mid-December. See Texas Data Center Moratorium.

Record

Dataset Record

Sources
Tesla Q1 and Q2 2026 shareholder updates and Form 8-K; earnings calls; ICCV 2025; Texas DMV registrations via the FSD Database; drone footage
Power
Tesla-reported onsite training capacity; targets as stated
Last verified
Q4 2026