The cluster tier.

What an H100 costs at the clouds that sell them by the rack — Lambda, Nebius, Crusoe, CoreWeave. This is the tier the futures market trades, and it is a different kind of number from the one on our marketplace index.

Posted rates, not transacted prices. Every figure here is an asking price a provider advertises on its own public page. Our marketplace index is what machines actually rent for. Quoted and cleared are not the same thing, the gap between them is the interesting part, and the two series are never merged.

reading the tape…

Methodology, in full.

Where a provider does not state its unit, we do not guess. CoreWeave publishes an hourly figure without saying whether it covers one GPU or a whole node — and it demonstrably mixes both. Those rows are recorded raw and carry no per-GPU price. A row marked raw only is us declining to assert something we cannot support.

On-demand only in the headline. Preemptible and spot figures are collected and shown, but the low and median are computed over on-demand quotes so we are not comparing a spot price against a committed one.

Honesty box.

This series is brand new — it began 25 Aug 2026 and this sentence will still be here when it is embarrassing. A handful of snapshots is not a price history, and nothing here should be treated as an index yet. It is a recorder that has just been switched on, published early because rows not collected today cannot be collected later.

Coverage is four providers. Voltage Park is missing on purpose: it merged into Lightning AI and no longer publishes a rate table. We would rather name the gap than quietly have one.

Raw: /api/clusters — every quote, its source URL, and the basis we recorded it on. The marketplace index and its tape live at /spot/ and /data/.