What changed, and why

30 Oct 2026

The H100 neocloud on-demand average for 24–29 Oct 2026 now counts Crusoe's GPU instance price instead of the price of its dedicated inference endpoints.

Why: From 16 to 22 Oct the rate-card reader took the prices in Crusoe's Self-Serve Deployments table, which are for dedicated inference endpoints, as its GPU instance prices. The board then kept the 22 Oct prices while they were within its window. The corrected values use the instance prices on the pages the reader stored. From 28 Oct the averages also lack Nebius: its on-demand prices were not read from 21 Oct and cannot be recovered. Both readers have been corrected.

30 Oct 2026

The H200 neocloud on-demand average for 24–29 Oct 2026 now counts Crusoe's GPU instance price instead of the price of its dedicated inference endpoints.

Why: From 16 to 22 Oct the rate-card reader took the prices in Crusoe's Self-Serve Deployments table, which are for dedicated inference endpoints, as its GPU instance prices. The board then kept the 22 Oct prices while they were within its window. The corrected values use the instance prices on the pages the reader stored. From 28 Oct the averages also lack Nebius: its on-demand prices were not read from 21 Oct and cannot be recovered. Both readers have been corrected.

30 Oct 2026

The B200 neocloud on-demand average for 24–29 Oct 2026 no longer counts Crusoe, which publishes no on-demand B200 instance price. On 28 and 29 Oct the average has too few providers and is withdrawn.

Why: From 16 to 22 Oct the rate-card reader took the prices in Crusoe's Self-Serve Deployments table, which are for dedicated inference endpoints, as its GPU instance prices. The board then kept the 22 Oct prices while they were within its window. The corrected values use the instance prices on the pages the reader stored. From 28 Oct the averages also lack Nebius: its on-demand prices were not read from 21 Oct and cannot be recovered. Both readers have been corrected.

30 Oct 2026

The H100 average across all on-demand providers for 24–29 Oct 2026 now counts Crusoe's GPU instance price instead of the price of its dedicated inference endpoints.

Why: From 16 to 22 Oct the rate-card reader took the prices in Crusoe's Self-Serve Deployments table, which are for dedicated inference endpoints, as its GPU instance prices. The board then kept the 22 Oct prices while they were within its window. The corrected values use the instance prices on the pages the reader stored. From 28 Oct the averages also lack Nebius: its on-demand prices were not read from 21 Oct and cannot be recovered. Both readers have been corrected.

30 Oct 2026

The H200 average across all on-demand providers for 24–29 Oct 2026 now counts Crusoe's GPU instance price instead of the price of its dedicated inference endpoints.

Why: From 16 to 22 Oct the rate-card reader took the prices in Crusoe's Self-Serve Deployments table, which are for dedicated inference endpoints, as its GPU instance prices. The board then kept the 22 Oct prices while they were within its window. The corrected values use the instance prices on the pages the reader stored. From 28 Oct the averages also lack Nebius: its on-demand prices were not read from 21 Oct and cannot be recovered. Both readers have been corrected.

30 Oct 2026

The B200 average across all on-demand providers for 24–29 Oct 2026 no longer counts Crusoe, which publishes no on-demand B200 instance price.

Why: From 16 to 22 Oct the rate-card reader took the prices in Crusoe's Self-Serve Deployments table, which are for dedicated inference endpoints, as its GPU instance prices. The board then kept the 22 Oct prices while they were within its window. The corrected values use the instance prices on the pages the reader stored. From 28 Oct the averages also lack Nebius: its on-demand prices were not read from 21 Oct and cannot be recovered. Both readers have been corrected.

30 Oct 2026

The B300 average across all on-demand providers for 24–29 Oct 2026 now counts Crusoe's GPU instance price instead of the price of its dedicated inference endpoints.

Why: From 16 to 22 Oct the rate-card reader took the prices in Crusoe's Self-Serve Deployments table, which are for dedicated inference endpoints, as its GPU instance prices. The board then kept the 22 Oct prices while they were within its window. The corrected values use the instance prices on the pages the reader stored. From 28 Oct the averages also lack Nebius: its on-demand prices were not read from 21 Oct and cannot be recovered. Both readers have been corrected.

30 Oct 2026

The H100 hyperscaler premium over the neocloud average for 28 and 29 Oct 2026 is recomputed from the corrected neocloud average.

Why: The neocloud average it is measured against counted Crusoe's inference-endpoint prices; see the H100 neocloud correction of the same date.

30 Oct 2026

The d/d change posted with the H100 neocloud price in the 30 Oct 2026 card on X is withdrawn: it compared averages over different providers, so it was never a like-for-like price move.

Why: It compared the 29 Oct average, over 7 providers, with the 30 Oct average, over 6. Crusoe was in the 29 Oct sample; on 30 Oct it was left out (not observed in the last 7 days). A provider leaving the sample moves the average with no price moving. The 29 Oct average was also restated on 30 Oct (see the H100 neocloud correction of that date), and the change against the restated average is not like-for-like either.

23 Oct 2026

The home page's H100 neocloud price is now the 20% trimmed mean of at least five neocloud providers' on-demand prices, shown with its sample count and date. Until five providers report, it reads "n/a (sample too small)".

Why: The old figure was the median of three providers' list prices, in practice one provider's rate card, and its label did not say so.

23 Oct 2026

The B300 marketplace price no longer counts fractional GPUs.

Why: The old value came from a single listing for a fraction of one B300 (a MIG slice), priced as if it were a whole GPU. Fractional GPUs are now excluded, and a price more than 50% below its class median on two consecutive days is held for review.

23 Oct 2026

The price board now publishes, for each GPU and provider class, the 20% trimmed mean, the number of providers and the as-of date (methodology 0.9). Spot and interruptible prices count in the marketplace class; the hyperscaler cell is a census of the four hyperscalers' list prices.

Why: Medians of a few list prices disagreed from page to page and with transaction benchmarks. One method, with its sample shown, removes the disagreement.

23 Oct 2026

Hyperscaler capex has one definition wherever it appears: cash capex, the purchases of property and equipment in each company's cash-flow statement, summed for the four largest hyperscalers.

Why: A brief quoted a higher total that included assets acquired under finance leases, so two different numbers were published for the same quarter.

23 Oct 2026

The equity basket is one definition everywhere: every covered company, thematic ETFs excluded, with its count shown beside it.

Why: A card quoted a different count from the site.

23 Oct 2026

Instrument cards without a computed reading are hidden until they have one, with its date.

Why: Several cards showed a placeholder instead of a value.

23 Oct 2026

The Pro annual price is read from one setting wherever it appears: the pricing page, the home page, emails and checkout.

Why: Two pages had shown different Pro annual prices.

23 Oct 2026

Dates on the home page come from the data: a day's post is announced as upcoming only until it goes out.

Why: The home page announced a post for a day that had already passed.

21 Oct 2026

The header and footer name the publisher, Apex Management Int LLC.

Why: They had named a former trading name.