Compute room · Dataset line · Schematic

Where the evidence gets cooked.

The direction: GPUs on hardware we control, turning messy cold-chain documents into datasets and evaluations people can inspect. Below is a schematic of that pipeline in motion.

Illustrative simulation. Random motion, example rows. Not live telemetry, not our hardware inventory, not measured performance.

The compute room

32 GPUs. Zero marketing.

Simulated avg load58%

The hardware · Founder-reported

Powered by NVIDIA Blackwell.

The compute behind the dataset work is NVIDIA Blackwell-generation GPUs. These are the models and memory sizes the founder reports; unit counts and benchmarks are not published yet.

Workstation class

NVIDIA RTX PRO 6000 Blackwell

96 GB GPU memory

Room for large models and long-context work on a single card.

Performance class

NVIDIA GeForce RTX 5090 (ASUS ROG Astral)

32 GB GPU memory

Blackwell consumer flagship for fine-tuning runs and dataset generation.

Founder-reported hardware, not an audited inventory. The rack animation above stays a schematic.

The dataset line

Documents in. Checked records out.

Each record keeps its source and its verdict. A missing assumption is a result, not a failure to hide.

  1. 1

    Intake

    Leases, utility bills, equipment records

  2. 2

    Extract

    Pull the fields and assumptions out

  3. 3

    Check

    Flag gaps and inconsistent math

  4. 4

    Review

    A person confirms what holds up

  5. 5

    Publish

    Versioned, documented dataset

records.jsonlexample rows · tick 0
  • Utility bill → average load 10 kW, rate $0.12/kWhinputs present, math reproducesChecks out
  • Pro forma → “refrigeration included in NNN charges”no equipment age or service recordGap found
  • Listing sheet → “state-of-the-art cold storage”claim with no evidence attachedRejected
  • Equipment log → compressor service dated, interval stateddated, sourcedChecks out

Where this stands

Direction, not a deployed datacenter.

  • The GPU motion on this page is a schematic. A small, real, re-checked sample of the dataset output is on the Datasets page.
  • The motion above is simulated. The rows are made-up examples of what a check looks like.
  • We’ll publish real datasets, methods and evaluation results in the open repository as they hold up.

Got documents, datasets, or compute know-how?

Real cold-chain records, field knowledge and engineering help make the line worth running.

Explore the projectBring a real problem