The GPUs that train Forge policies.
This is internal infrastructure. Tenon does not sell compute, hosting or GPU access; the product is the trained policy and the edge runtime on Jetson AGX Thor. This page documents the training node today, the target for the next six months and what it costs.
Topology
| Today | 8×H100 80 GB, burst regime |
|---|---|
| Target, in six months | 12×H100 80 GB, reserved capacity |
| Parallel environments | 4,096 in Isaac Lab |
| Per generation | 90 GPU-hours, 11 hours of wall-clock |
Orchestration
Slurm schedules each generation on the node. Ray distributes the 4,096 simulation environments and the rollout workers inside that allocation.
Storage
Each generation writes 1.8 TB of synthetic frames. The seed, the manifest revision, the scene hash and the container digest are archived with it, so the generation can be rebuilt later. The procedure is on reproducibility.
Containers
Training runs in NGC containers referenced by digest, never by tag, starting from nvcr.io/nvidia/pytorch:26.08. A tag can be moved to a new image; a digest cannot, so every published policy points at the exact image it was trained in.
Consumption
| Today | Target | |
|---|---|---|
| GPUs | 8×H100 | 12×H100 |
| GPU-hours per month | 1,980 | 6,000 |
| Generations per month | 22 | 40 |
| Compute spend per month | US$ 4,400 | US$ 13,200 |
What we are asking for
- Reserved H100 capacity instead of burst.
- Expansion from 8 to 12 H100 GPUs, ramped over six months.
- A US$ 150k allocation, which covers eleven months at the target rate of US$ 13,200 a month.
The manifest preview that runs on H100 before a cell is committed is described on inference.