Cloud GPU Rentals

hwLedger integrates with major GPU rental platforms to query pricing, availability, and automatically spin up instances.

Supported providers

Provider	Pricing API	Launch API	Shutdown API	Spot	Interruptibility
Vast.ai	Yes	Yes	Yes	Yes	5-30% undercut, can interrupt
RunPod	Yes	Yes	Yes	Yes	Stable (24h min)
Lambda Labs	Yes (manual)	Yes	Yes	No	Dedicated, 10h monthly
Modal Labs	Yes	Deployment	Auto	No	Via timeout

Pricing cache

hwLedger caches prices locally to avoid repeated API calls:

File: ~/.cache/hwledger/pricing-cache.json

{
  "providers": {
    "vast": {
      "last_updated": "2026-04-18T21:45:00Z",
      "instances": [
        {
          "id": "123456",
          "gpu": "RTX4090",
          "vram": 24,
          "$/hour": 0.18,
          "availability": "available",
          "location": "US-West"
        }
      ]
    },
    "runpod": {
      "last_updated": "2026-04-18T21:43:00Z",
      "instances": [...]
    }
  }
}

Refresh: cache invalidated after 1 hour or manual hwledger cloud refresh-prices.

Configuration

File: ~/.config/hwledger/cloud.toml

[providers.vast]
api_key = "vast-api-key-here"
min_vram_gb = 16
max_$/hour = 0.50
filter = { location = "US", min_bandwidth_gpbs = 1.0 }

[providers.runpod]
api_key = "runpod-api-key"
min_vram_gb = 20
max_$/hour = 0.40

[providers.lambda]
api_key = "lambda-api-key"
min_vram_gb = 24

[budget]
total_$/month = 100
alert_at_$/hour = 0.60

Launch workflow

User command:

hwledger fleet launch --model llama-70b-instruct \
  --context 32000 \
  --provider vast \
  --timeout 2h

Server process:

1. Query prices: GET /api/pricing from Vast → cache
2. Filter by model: find GPUs with >= 48GB VRAM (70B model requirement)
3. Sort by cost: pick cheapest available
4. Create instance: POST /api/instances with image ID (Ubuntu + hwledger-agent preinstalled)
5. Wait for boot: poll instance status until "running"
6. SSH health check: ssh ubuntu@instance.ip nvidia-smi
7. Register with fleet server: agent heartbeats in
8. Schedule job: planner sends model + inference request
9. Teardown: on timeout or user command, DELETE /api/instances/{id}

Cost estimation

Before launching, planner estimates cost:

Model: llama-70b-instruct (70B params, FP16)
Context: 32K tokens
Cost per inference: (model_size + kv_cache) / bandwidth = cost/s
Vast.ai RTX4090: $0.18/hr = $0.00005/sec

Prefill (32K tokens): ~10 sec × $0.00005 = $0.0005
Decode (1K tokens): ~3 sec × $0.00005 = $0.00015
Total per request: $0.00065 (+ overhead)

Monthly (1000 requests): $0.65 (+ idle cost)

User confirms before launch.

Spot instance risks

Vast.ai spot instances save 30% but can be interrupted:

[budget]
prefer_spot = true  # Use spot if available
timeout_on_interrupt = 60  # Seconds to migrate job before forced kill

hwLedger automatically:

• Saves job state to ledger (persist-in-flight)
• Migrates to new instance
• Resumes from checkpoint

Regional selection

Placement scores consider region:

Score = (availability / instance_cost) × (latency_ms / 100)

Lower score = better. Server picks top 3 instances, user selects region interactively.

Related

• Fleet Agent: Deployment
• Placement Algorithm (coming soon)
• Cost Estimation