The H100 SXM week of October 5, 2026 closed at 00:00 UTC on October 12 with a median of $3.49 per GPU-hour over 33 public on-demand rate cards at 31 providers - the same $3.49 as the week of September 28, which was taken over 31 rate cards at 29 providers. Our base-100 index moved in the same week from 102.6 to 104.2: 104.2 ÷ 102.6 = 1.016. Both figures are on /price-index, and each quote behind them is in the week’s receipt with its source page and read time.
Why one figure stood still and the other rose
The two numbers answer different questions. The dollar median is the middle rate card of whoever was in the basket that week. The base-100 index is chained: it moves only with quotes we read in both weeks, so a provider entering or leaving does not move it. In the week of October 5, three quotes read in both weeks rose and none fell, so the index went up. At the same time three cheaper quotes entered and one left, and the middle of the list stayed on the same figure.
The quote in the middle was JarvisLabs at $3.49. In the week of September 28 two quotes stood at exactly $3.49, JarvisLabs and RunPod Secure Cloud, with 15 below and 14 above. In the week of October 5 one quote stood at $3.49, with 16 below and 16 above.
The three rates that rose
RunPod’s Secure Cloud H100 SXM went from $3.49 to $3.99, first read at the new rate at 16:41 UTC on October 8. Verda (DataCrunch) moved its on-demand 1H100.80S.30V instance from $3.737 to $4.007 in six steps of about one percent between October 5 and October 11. GPU.ai’s H100 SXM row, an advertised “from” rate, was $2.95 at the last read of the week of September 28 and $3.88 at the last read of the week of October 5. GPU.ai reprices its catalogue within the hour, so its weekly figure is the rate on the page at that last read, not an average of the week. We did not change the modules that read these three pages during the week; the moves are on the pages.
Two of the three crossed the middle. RunPod Secure Cloud left the $3.49 mark for $3.99, and GPU.ai went from below $3.49 to above it. On their own these would have lifted the median.
Who entered and who left
Three providers entered, each below the median: TensorPool at $1.99 for the GPU alone, Bitdeer AI at $2.36 and VESSL AI at $2.98. All three are new to us. We first read Bitdeer on October 10 and TensorPool and VESSL on October 11, so none of these rates is a change in their prices. Edgevana’s quote, $3.07 in the week of September 28, left. That exit is ours, not Edgevana’s: on October 7 we stopped reading Edgevana’s fleet index, which had reported whole machines and reserved servers as single GPUs, and began reading its card pages instead. No H100 SXM figure from those pages was read in the 48 hours before the week closed, so the quote was kept out.
The count below the median went 15 - 1 - 1 + 3 = 16: GPU.ai crossed upward, Edgevana left, three entered. The count above went 14 + 2 = 16 with GPU.ai and RunPod Secure Cloud. With 16 on each side, JarvisLabs at $3.49 stayed in the middle. Split by operator class, the specialist-cloud median was $3.20 over 29 quotes at 27 providers and the four hyperscalers’ median was $9.90. Nothing here is forecast.
Check it yourself
Open /price-index and its receipt for the week, 2026-W41, which lists all 33 included quotes and the seven read and kept out, each with its rule. Every rate named above links to the provider’s own pricing page, and the read time beside it is the scrape at 23:40 UTC on October 11, the last of the week. The week of September 28 has its own receipt, 2026-W40, for the figures it is compared with.