GPUs

On-demand accelerated compute for training, fine-tuning, and inference. Provision GPU clusters, watch live utilization, and meter spend from the console or the API. GPUs bill by the hour to the same per-org cloud-usage ledger as the rest of your fleet.

How it works

Hanzo Cloud GPUs are real accelerators resold and metered by the hour. Capacity is sourced from cloud GPU providers — DigitalOcean GPU Droplets are the primary pool (H100, A100, L40S), with Paperspace and AWS as secondary pools for burst and specialized instances — and reconciled by the operator alongside your machines. You get the accelerator; Hanzo handles provisioning, scheduling, telemetry, and a single unified bill.

• One key, one bill — GPU-hours meter to your org's cloud-usage ledger through commerce billing, priced through api.hanzo.ai/v1/gpu.
• No idle lock-in — hourly metering starts when a node comes online and stops when you tear the cluster down.
• Honest telemetry — utilization, memory, temperature, and power come straight from the machine; fields the provider does not report render as —, never a fabricated value.

Accelerators

Accelerators are grouped into clusters (a node pool of one GPU size) and scheduled into pools. Node sizes map to real accelerator counts — a gpu-h100x8-640gb node genuinely holds eight H100s:

Pricing

GPUs are metered per GPU-hour. Representative on-demand rates:

Live pricing is served from GET /v1/gpu and mirrored under Compute → GPUs → Pricing in the console. Sustained workloads qualify for reserved discounts (10% for a 1-month commitment, 20% for 3 months) — see compute pricing for the full rate card, regions, and reserved terms.

Provision a cluster

Create a GPU cluster from Compute → GPUs → Clusters in the console, or from the CLI / API.

# Launch an 8× H100 training cluster
hanzo vm launch \
  --size gpu-h100x8-640gb \
  --region sfo3 \
  --name train-sfo3

# Or straight against the control plane
curl -X POST https://api.hanzo.ai/v1/visor/machines \
  -H "Authorization: Bearer hk-..." \
  -H "Content-Type: application/json" \
  -d '{"size":"gpu-h100x1-80gb","region":"sfo3","name":"infer-0"}'

Each cluster's nodes become your machines; the operator reconciles them into the running fleet and starts metering when they report online.

Inventory & telemetry

List every GPU with live telemetry — model, cluster, region, utilization, memory, temperature, and power — through the control plane:

curl https://api.hanzo.ai/v1/gpu \
  -H "Authorization: Bearer hk-..."

{
  "gpus": [
    {
      "id": "gpu-0",
      "model": "H100",
      "cluster": "train-sfo3",
      "region": "sfo3",
      "status": "online",
      "utilization": 87,
      "memoryUtil": 74,
      "temperature": 61,
      "power": 640
    }
  ]
}

Fields the platform does not report render as — in the console rather than a fabricated value. Alerts and scheduling pools are available at /v1/gpu/alerts and /v1/gpu/pools.

Billing

GPU-hours are metered into your org's cloud-usage ledger and settled through commerce billing. Billing is prepaid: top up credits at console.hanzo.ai and spend draws down as clusters run. When the balance is exhausted, new provisioning is refused with 402 Payment Required ("Add credits at console.hanzo.ai") — running clusters are never silently overdrawn.

Track spend under GPUs → Pricing in the console, or pull usage from the billing API. See API Keys for auth and the API Reference for the full surface.

Related

• Machines — the cluster nodes your GPUs run on
• Pipelines — build and deploy onto GPU clusters
• Billing — credits, invoices, and usage metering
• Edge — on-device inference when a cloud GPU is overkill