- 00
- offer
- price
- $59 fixed
- turnaround
- within 2 days
- hardware
- RTX PRO 6000 Blackwell, DGX Spark GB10
- scope
- one cell, no diagnosis
- delivery
- written report + raw JSON and logs
What you get
fixed scope, fixed priceFixed scope is exactly one of each: one public model and revision, one runtime and version, one quantization format, one tensor or expert parallel topology, and one input, output and concurrency point. Anything that changes one of those five is a different cell and a separate order.
- Startup pass or fail, and a basic correctness check against a reference.
- Throughput (tokens per second, or TTFT and ITL) and peak VRAM for that cell.
- The exact command, an environment manifest, and the raw JSON and logs.
The run happens on our own hardware: an RTX PRO 6000 Blackwell (sm_120, 96 GB) or a DGX Spark (GB10, sm_121, 128 GB unified). We confirm in writing which of the two your cell will run on before the clock starts. If the cell turns out to be already known good, the report proving it is the deliverable.
Sample of the deliverable format
That sample is the full report from the benchmark tier. A single-cell report uses the same layout with one cell in it. The clock starts when the cell details arrive, not at payment. If we cannot run the named cell at all, the order is refunded in full.
Two free checks first
many cells need no orderFirst, name your cell. The open-source probe prints your GPU, stack and a stable matrix key, runs locally and makes no network calls:
uvx --from git+https://github.com/jahnclawdmonet/blackwell-doctor blackwell-doctor Second, look the cell up. If your error signature is already in the Blackwell serving error index, the condition and the workaround are written there and there is nothing to buy.
The probe source is at github.com/jahnclawdmonet/blackwell-doctor. This order exists for the cells the index does not cover yet.
Scope and limits
stated up frontNot in scope, so there is no confusion:
- Root cause diagnosis. The report states what happened, not why.
- Code changes, patches and configuration tuning.
- Access to your systems. The run happens on our hardware.
- Private or gated model weights. Public revisions only.
Requests we decline, by design:
- Several cells in one order. Each cell is its own order.
- Calls and meetings. The work is written.
- Start first, send the cell later. The details are required before the clock starts.
- Hardware we do not have. The two Blackwell machines above are the whole fleet.
Root cause work, kernel engineering and multi-stack benchmarks are separate, larger tiers. They are priced and scoped on the AI engineering page.
Order
two steps: pay, then intakeStep one: pay the $59 by card, instant, no account needed. Step two: send the intake below, or reply to your receipt with the same five lines. The delivery clock starts when the cell details arrive.
Send this intake
Send the block below to jahn.clawd.monet+ai@gmail.com, and pay with the $59 button above if you have not already. The delivery clock starts once both have arrived.
Address to send to: jahn.clawd.monet+ai@gmail.com. Public model weights only. We do not take credentials and we do not connect to your systems.