Compute

GPU capacity, run by the people who own the racks.

Reserve a cluster and drive it yourself, or hand us the model and consume it as an endpoint. Same fleet, same engineers, same operating standard behind both.

Accelerators
Blackwell and Hopper generation
Fabric
InfiniBand · high-speed Ethernet
Access
Bare metal · Kubernetes · VMs
Commercials
Per GPU-hour or per MW-month

Four ways to consume the fleet.

Reserved clusters
Training

Isolated multi-node clusters held for you across the contract term. Non-shared GPUs on a dedicated fabric, with parallel filesystem storage sized to the run and a fixed node count you can plan a training schedule around.

  • Slurm or Kubernetes scheduling
  • Dedicated InfiniBand islands
  • Checkpoint-grade shared storage
On-demand capacity
Experimentation

Smaller node counts for evaluation, ablations and burst work, drawn from the same pools as the reserved fleet so behaviour in a pilot matches behaviour in production.

  • Hourly and monthly terms
  • Same images as reserved nodes
  • Path to reservation without migration
Managed inference
Production

Dedicated endpoints for open-weight and custom models, deployed and tuned by our engineers. Capacity is held for your peak rather than borrowed from a shared pool, and the endpoint stays in the region you chose.

  • OpenAI-compatible API surface
  • Latency and throughput targets in contract
  • Autoscaling inside reserved headroom
Training services
Adaptation

Managed fine-tuning and post-training workflows to adapt open models to your data, then promote the result straight onto your own endpoints without the weights leaving our infrastructure.

  • Supervised and preference tuning
  • Evaluation harness per iteration
  • Weights stay in region

What sits under a cluster.

The platform is deliberately thin. Compute, network, storage and observability, run to one standard across every region we operate.

Non-virtualised GPUs

Direct access to the accelerator, no hypervisor tax between your job and the silicon.

Storage that keeps up

Parallel filesystem for training sets and checkpoints, object storage for everything cold.

Private networking

Private interconnect and peering into the region, with tenant isolation at the fabric level.

Fleet observability

Node health, fabric counters and job telemetry exposed to you, not just to us.

Failure handling

Spare nodes held in region, drain-and-replace during long runs, engineers on call in your timezone.

Residency and control

Data stays in the country you picked, under South American jurisdiction, with access logs you can audit.

Tell us the workload and the window.

We come back with a cluster specification, a region and a delivery date.

Reserve capacity → See the footprint