Tesla technical article 012
Cortex AI Training Clusters
What Tesla has disclosed about Cortex 1 and Cortex 2 at Gigafactory Texas, how large GPU training clusters operate, and which architectural details remain unpublished.01
What it is
Cortex is Tesla’s onsite GPU training infrastructure at Gigafactory Texas. It is used to train and improve neural networks for Tesla’s AI products and services, rather than to run those networks inside vehicles or robots.
The programme currently comprises Cortex 1, which Tesla lists in production, and Cortex 2, which came online and began running training workloads in the first quarter of 2026. Tesla reports capacity in H100-equivalent GPUs: a normalised compute measure, not a literal count of one accelerator model.
OwnerTeslaOn-premises AI infrastructure
LocationGiga TexasAustin, Texas
Cortex 1>100k H100eProduction
Cortex 2>130k H100eEarly ramp
02
Capacity and disclosed milestones
| Disclosure | Cortex 1 | Cortex 2 |
|---|---|---|
| Q4 2024 | Tesla completed a roughly 50,000-H100 cluster. | Not yet disclosed. |
| Q4 2025 | Listed at more than 100,000 H100 equivalents; production. | Under construction; Tesla planned to more than double onsite Texas compute in H1 2026. |
| Q1 2026 | More than 100,000 H100 equivalents; production. | More than 130,000 H100 equivalents; online, running training workloads and in early ramp. |
| Combined published capacity | More than 230,000 H100 equivalents, subject to Tesla’s capacity and ramp definitions. | |
Read the unit carefully. “H100e” allows unlike generations or configurations to be expressed against an H100 baseline. It does not establish the installed number of H100 GPUs, their precision, utilisation or sustained model throughput.
03
System architecture
04
How a training job uses the cluster
- Prepare the dataset.Training pipelines select and transform real-world video and other recorded signals into examples suitable for model optimisation.
- Request resources.A job describes the required GPU, CPU, memory and time allocation; Tesla identifies Slurm as the scheduler at the heart of its AI HPC infrastructure.
- Place workers.The scheduler assigns nodes while considering topology, availability, priority, reservations and fair-share policy.
- Run distributed training.Workers calculate forward and backward passes, then exchange gradients or model partitions across the cluster network.
- Checkpoint state.Periodic checkpoints preserve progress so a long job can recover after hardware or software failure.
- Evaluate and iterate.Engineers test the resulting model, alter data or training configuration and submit the next run.
05
Software and cluster control
| Layer | Confirmed role | Why it matters |
|---|---|---|
| Linux and containers | Tesla recruitment material names Linux, Docker, Kubernetes and Pyxis/Enroot experience. | Packages reproducible jobs and isolates dependencies across shared nodes. |
| Slurm | Allocates GPUs and governs scheduling, pre-emption, backfill, reservations and accounting. | A large cluster is useful only when scarce accelerators are placed and kept busy efficiently. |
| Observability | Tracks utilisation, throughput, failures and queue behaviour. | Identifies idle devices, fabric bottlenecks and repeated job failures. |
| Training framework | Tesla has not published one complete Cortex software stack. | Framework, parallelism and compiler choices materially affect scaling. |
06
What Cortex trains
Tesla links Cortex to the development of its AI products and services. Public filings connect the original cluster to FSD (Supervised) development, while current infrastructure roles describe the clusters as central to Robotaxi and Optimus work. These are training workloads: the resulting model weights later run on separate inference hardware.
| Workload | Cluster task | Execution destination |
|---|---|---|
| Vehicle autonomy | Train vision, planning and control models from selected fleet data. | Vehicle inference computer after validation and compilation. |
| Robotaxi | Develop and iterate autonomy models and supporting world representations. | Robotaxi vehicles and service infrastructure. |
| Optimus | Train perception, policy and embodied-intelligence models. | Robot inference hardware. |
07
Current state and adjacent silicon
| System | State at verification | Boundary |
|---|---|---|
| Cortex 1 | Production; more than 100,000 H100-equivalent capacity. | Operational training cluster. |
| Cortex 2 | Online, running workloads; more than 130,000 H100 equivalents in early ramp. | Additional onsite training capacity. |
| Dojo 3 | Custom-silicon development continuing. | A separate future effort to reduce training cost; no current production specification published. |
| AI5 / AI6 | Inference chips on separate development timelines. | Not Cortex training accelerators. |
08
Terms
- H100e
- Capacity normalised to NVIDIA H100-equivalent GPUs.
- Training
- Optimising model parameters from data.
- Inference
- Running a trained model to produce an output.
- Slurm
- Batch scheduler and resource manager for HPC clusters.
- Backfill
- Running smaller jobs while reserved resources await a larger job.
- Checkpoint
- Saved training state used for recovery or later continuation.
- Fabric
- Network carrying traffic between training workers.
- Early ramp
- Capacity is online but still moving towards its intended operating level.
09
Limits of the evidence
- H100-equivalent capacity is not an audited physical GPU count or sustained-performance measure.
- Tesla has not published Cortex 2’s accelerator mix, node configuration, precision basis or conversion method.
- Network topology, link bandwidth, storage throughput, power demand and cooling design remain undisclosed.
- The combined figure above adds two published capacity values; it does not prove simultaneous usable capacity.
- Recruitment pages document operational tools and practices but are not complete architecture specifications.
- Workload descriptions do not prove that every named programme runs exclusively or continuously on Cortex.
- Price, return-on-investment claims and unofficial site estimates are excluded.
10
Sources
- Tesla Q1 2026 UpdateCortex 1 and 2 capacity, status and training-capacity chart.
- Tesla Q4 2025 UpdateCortex 2 construction and H1 2026 expansion plan.
- Tesla Q4 2024 UpdateCompletion of the roughly 50,000-H100 original Cortex deployment.
- Tesla 2025 Form 10-KInfrastructure purpose and Cortex 2 construction context.
- AI workload scheduling roleSlurm, scheduling, observability and cluster-operating practices.
- Tesla AI and RoboticsRelationship between vision-based AI, autonomy and robotics.