Per-GPU memory and allocation-hour price. Each point is a listed configuration.
Compute priceboard
Ordered by allocation cost for your selected hours.
| + | GPU / allocation | Provider | GB / GPU | $/allocation-h | Your period | Basis |
|---|
When does owning the machine make sense?
Model the whole system, including idle power. Compare the same work when you have measured rates.
Choose the machine you would rent
Energy + maintenance + amortization
Upfront purchase / operating savings
Keep the calculation with the decision.
Your inputs stay editable. A source-bound snapshot preserves what you compared.
What if you stopped paying premium prices for routine work?
Separate reusable procedures, smaller-model work and the judgment that remains. Count every retry.
Requests off premium
Premium GPU time saved
Less GPU time is not automatically a released allocation. Residency, isolation and scheduling must establish reusable capacity.
Bring your measurements. Keep the evidence.
The Hot Aisle runner measures your workload on a GPU allocation and publishes a checksummed record. Connect it here to keep its checksummed, recomputable records beside your price, ownership and reuse decisions. Until a runner is connected, a synthetic sample shows what an observed run looks like.
Observed runs
Published qualification records
Connect the Hot Aisle runner
Read-only. Everything runs on your machine; nothing is uploaded.
Download the runner kit (a separate download from this desk's offline kit), start it with node runner/bin/workload.cjs serve --port 8787, then start this desk's local helper and open the private link it prints. Runs and records the runner publishes appear above.
For developers
A working stdio MCP server, not a prompt template. Search the priceboard, calculate ownership, model qualified-route proposals and inspect permitted results with the same engine.
node compute/scripts/connect.cjs --mcp \ --results /path/to/approved-results
Node 22+. Exact local paths are chosen once during setup. Your normal orchestration remains the execution authority. Optional read-only inputs: an Aperture aperture-support/1 hardware receipt (--aperture) and one explicitly configured model endpoint (--endpoint).
Saved decisions
Explicitly saved on this browser, up to 12. Export a decision to keep it independently.
Offline import and verification
Open the full source-bound benchmark comparator ↗. The retained v2 workbench supports matched-side comparison and per-request quality gates.
Existing vLLM benchmark files (JSON, JSONL, console summaries), published runner records and saved decisions are accepted locally. Import is the fallback when a connection is unavailable.