Enterprise compute
A control plane built for multi-node training, not a Docker wrapper.
Every launch is health-gated: a machine only counts as ready after it passes a real health check, and a job that never becomes ready is refunded as account credit.
$ curl -X POST https://www.kilawattcloud.dev/api/public/v1/agent/exec \$ -H "Authorization: Bearer $KILAWATT_API_KEY" \$ -d '{"gpu_type":"nvidia-h100","card_count":8,"duration_seconds":600}'Multi-Supplier Job Orchestration
Each job is routed to a supplier by live cost and availability, launched through the supplier's API, health-checked, and torn down at the end of the paid window.
- › Live cost-based routing
- › Health-gated launches
- › Failover across suppliers
Provider-Native Interconnect
Jobs land on the interconnect the serving supplier offers: NVLink within a node, and RDMA-capable fabric where the supplier supports it. We do not guarantee a specific fabric.
- › NVLink intra-node
- › Provider RDMA where available
- › Interconnect set by the supplier
Control Plane & API
Lightweight orchestration for provider selection, live cost quoting, pre-flight profitability checks and automated teardown at the end of the paid window.
- › Live cost-based quoting
- › Pre-flight margin guardrail
- › Automated teardown & reaping
Auditability & Fault Recovery
Every provider attempt, winning and failing, is written to an append-only ledger with the raw response body, enforced by the database rather than application discipline.
- › Append-only attempt ledger
- › Multi-provider failover routing
- › Raw provider responses retained
Cluster architectures
Example cluster shapes we can source from suppliers
Planning references for enterprise builds, not guaranteed on-demand inventory. Large datacenter and B200 clusters are sourced and quoted with our team. Availability, interconnect and cooling depend on the supplier that serves the job.
| Example shape | Scale | Workload |
|---|---|---|
| GB200 NVL72 Rack | 72 GPUs · 1 rack | Frontier-scale foundation model training |
| B200 SXM Pod | 16 nodes · 128 GPUs | Multi-node pre-training |
| H200 Inference Fleet | 8 nodes · 64 GPUs | High-volume LLM serving at low TTFT |
| H100 Training Cell | 32 nodes · 256 GPUs | Distributed pre-training & fine-tuning |
Platform safeguards
Enforced in code, not promised on paper
Usage metering interval
Every 15 min
Automatic instance shutdown at zero balance
Job pre-authorization
100% of estimated cost
402 before any node is contacted
Rate limiting
30 req/min per key
Deterministic 429 responses
Concurrency
3 jobs per key
429 until a slot frees
Payment security
PCI-DSS via Stripe
Card data never touches our servers