$10K Sovereign AI Lab

Layers under testMODELS · COMPUTE · DATA

The test

Build a complete local AI environment inside a fixed $10,000 budget, then run real production work through it for a sustained period.

Budget rules

The ceiling includes the compute node, memory, internal storage, any dedicated networking or backup hardware, a UPS, a wall-power meter, tax, and shipping. Loaned, donated, prototype, or unavailable hardware does not count. Prices are captured on the day the bill of materials is locked.

An existing display, keyboard, and ordinary internet connection are excluded as general office equipment. Any service needed specifically to make the AI stack work must be priced, even if its introductory tier is free.

Workload

Before the hardware is ordered, we will publish a versioned workload pack with:

1. document extraction and summarisation over a fixed local corpus;
2. retrieval-backed question answering with citations to that corpus;
3. a bounded software task with executable acceptance tests;
4. batch transcription and structured metadata extraction; and
5. one tool-using workflow that requires human approval before it changes state.

The pack will include prompts, input hashes, expected output schemas, scoring rules, and the information that must remain on the local network. Private production material may be used as an additional case study, but never as the only evidence for a published result.

What gets measured

• Total cost, including everything that is easy to forget: power, storage, networking, time.

• Which models the machine can hold, and at what quality and context length.

• Throughput and latency on the actual workloads, not on synthetic benchmarks.

• What breaks, how often, and what it takes to keep it running.

Method

• Lock the bill of materials, software versions, model revisions, quantisations, context limits, and workload before the measured run.
• Record cold start and warm performance separately. Report time to first token, generation throughput, end-to-end task time, peak memory, and wall power rather than one headline number.
• Run each scored workload at least three times and publish every run, including failures and operator interventions.
• Complete a seven-day mixed-workload soak after initial setup. Log restarts, thermal throttling, unavailable services, manual recovery, and maintenance time.
• Re-run the acceptance suite from a clean machine using the published setup instructions before calling the environment reproducible.

Publication packet

The result will include the full bill of materials, receipts with personal information removed, architecture diagram, install procedure, configuration, workload pack, raw measurements, scoring code, and a change log. Conclusions will distinguish what the machine can hold, what it can run interactively, and what it can complete reliably.

Status

Protocol v0.1 · planned. The budget rules and measurement contract are now public. Hardware selection, the final workload pack, and the run have not been completed. No performance result is implied by this protocol.