What the $200k buys.

At a planning band of $3 to $4 per H200-hour, $200k buys approximately 50,000 to 67,000 H200-hours. Depending on the model size, 10 to 15 benchmarks across two or three checkpoints or model variants will be a meaningful jump in safety testing.

Task	Compute	Notes
Initial benchmarks	$20k–$30k	Ten to fifteen benchmarks across initial checkpoints and scaffolds
Replication and elicitation	$30k–$50k	Repeats, alternate scaffolds, capability elicitation and confirmation runs
Agentic and human-directed red-teaming	$40k–$70k	Long-horizon environments and adaptive branches from observed failures
Interpretability and diagnosis	$25k–$50k	Multi-week H200-node blocks, activation capture and targeted analysis
Failure analysis and retesting	$20k–$40k	Reproducing findings and repeating affected evaluations after a voluntary model change
Storage, egress and variance reserve	$10k–$20k	Large intermediate artifacts, failed runs, provider variance and workload variance
Total range	$145k–$260k	Campaign-dependent

This type of compute arrangement is identical to the existing workstream of many labs. As a partner we help reproduce and diagnose findings, retest, and run safety fine-tuning before launch where a lab wants help fixing what we found.

A clean model may finish well below $200k and a difficult model may run well above it. Unused compute rolls into the next launch, so this number is an approximation rather than a hard cap. The moment an evaluation discovers something important is the wrong moment to run out of compute; we need to stay flexible in the spend while making the cluster fair to use for all partners involved.

A year

We cannot predict how many open-weight models will launch next year. Last year is the best guide: 42 releases with 1,000 or more Hugging Face likes. At the $200k estimate per launch, a year like that costs $8.4M in compute. At the table's range, $6.1M to $10.9M.

Counts from the ledger, last 365 days. The per-launch estimate is the table above.

Back to funders