Delivered Compute Report · 2026-10-26 · previous issue (2026-09-14)
The neoclouds repriced, and delivery became the question
We re-read every one of the registry's 133 offerings on the seller's own page this week. 26 prices changed. 18 were increases; the rest are a format change (Nebius preemptible), probable currency drift (Scaleway) and one newly published price (Google). The cheap tier of H100 capacity got smaller: the median walk-up price for an H100-SXM across the 13 providers on our headline board rose from $3.99 to $4.61 (+15%), while the top end (the hyperscalers) did not move. Six weeks ago the spread ran from cheap neoclouds to expensive clouds. Now the neoclouds are converging upward.
What moved
| Provider | Offerings | Before | Now | Change |
|---|---|---|---|---|
| Verda (DataCrunch) | B300 / B200 / H200 / H100 / A100 / L40S on-demand | $7.50 / 6.11 / 4.00 / 3.25 / 1.79 / 1.37 | $10.91 / 8.52 / 6.01 / 4.61 / 2.16 / 1.83 | +21% to +50% |
| RunPod | Secure Cloud B200 / H200 / H100 SXM / A100 SXM / RTX 4090 (H100 PCIe unchanged) | $6.79 / 4.59 / 3.49 / 1.59 / 0.74 | $7.99 / 5.29 / 3.99 / 1.79 / 0.89 | +13% to +20% |
| Nebius | HGX B200 / H200 / H100 on-demand | $7.15 / 4.50 / 3.85 | $8.50 / 5.40 / 4.50 | +17% to +20% |
| AWS | Capacity Blocks p5 / p5e / p5en / p6-b200 (GB200 UltraServer unchanged) | $5.19 / 5.97 / 6.87 / 12.36 | $5.97 / 6.87 / 7.90 / 14.21 | +15% |
| Google Cloud | a3-ultragpu-8g (H200) on-demand — first published price | — | $10.60 | new |
| Scaleway | H100-SXM / H100 PCIe / L40S / L4 (USD via aggregator; likely FX, not list changes) | $3.61 / 3.34 / 1.68 / 0.87 | $3.57 / 3.21 / 1.65 / 0.88 | −4% to +1% |
Per GPU-hour. Every one of these rows carries the price text we read on the page, in the re-survey file. RunPod's page is dated 2026-10-15. AWS's Capacity Block page says its next price update is January 2027. Nebius's preemptible tier changed format. It used to be a fixed price; it is now a dynamic spot rate with a floor ("from $0.79–0.99"). We record the floor, flagged variable-floor, and keep it out of every price band.
What is no longer published
- Verda long-term contracts: the dollar prices for its 1- and 2-year contracts are gone. The page now shows only "8% / 25% off on-demand" and "get a custom quote". We did not derive prices from the percentages, so 8 rows are delisted.
- Lambda 1× H100 PCIe: no longer on the instances page, so it is delisted.
- Google a4-highgpu-8g (B200): still has no on-demand price ("N/A"). It is now sold only through Flex-start, Calendar and Spot.
What we could not re-read (19 rows)
These keep their 2026-09-10 price and date, so the staleness shows on the row:
- Crusoe (4): the prices are on the page but not tied to rows we could read.
- Lambda 1-Click H100 (3): rendered client-side.
- AWS Savings Plans (6): rendered by script.
- Vast.ai (5): the marketplace API returned contradictory snapshots.
- Scaleway B300 (1): the aggregator contradicts itself.
A row we could not read is never marked confirmed.
Read across: the board, then and now
| H100-SXM walk-up, 13-provider board | 2026-09-10 | 2026-10-26 |
|---|---|---|
| Cheapest (Hyperstack) | $3.20 | $3.20 |
| Median | $3.99 | $4.61 |
| Most expensive (Azure) | $12.29 | $12.29 |
| Providers at or under $4.00 | 7 | 5 |
The same-chip spread is still 3.8×, but it now comes from a different place. Six weeks ago the cheap end was crowded with neoclouds a long way below the hyperscalers. Now three of them (Verda, Nebius and RunPod's Secure Cloud) have moved toward the middle within a single month. That is what a capacity crunch looks like from the outside. It also matters for the cash-settled GPU futures now coming to market: an index built on posted walk-up rates just absorbed a step-change that came from a handful of repricings, not from any change in what was delivered.
New this issue: will the capacity show up?
The more consequential question is no longer what an hour costs. It is whether committed capacity arrives on schedule. We graded the next unfinished phase of ten announced AI sites on public grid, build, hardware and track-record evidence, published a probability for each, and wrote it to a hash-chained ledger that resolves forecast by forecast.
| Site (next phase) | Target | Grade | P(on time) |
|---|---|---|---|
| IREN Childress, Horizons 2–4 → Microsoft | 2026-Q4 | DR-B | 65% |
| TeraWulf Lake Mariner, CB-4 → Fluidstack | 2026-H2 | DR-B | 62% |
| Cipher Barber Lake, first halls → Fluidstack | 2026-Q4 | DR-B | 61% |
| Microsoft Fairwater, Datacenter 2 | 2028 | DR-B | 58% |
| SpaceXAI Colossus II → Anthropic | 2026-12 | DR-C | 48% |
| Applied Digital Polaris Forge 1, Building 3 → CoreWeave | 2027 | DR-C | 47% |
| AWS New Carlisle (Rainier) phase 2 → Anthropic | 2028-Q1 | DR-C | 47% |
| Stargate Abilene, final two buildings → Oracle/OpenAI | 2026-Q4 | DR-C | 43% |
| Nebius Vineland → Microsoft | 2026-12 | DR-D | 34% |
| Nscale Monarch, Anthropic building | 2027-Q4 | DR-D | 32% |
The pattern: the three lowest-graded sites with near-term targets (Vineland, Colossus II, Abilene) all depend on on-site gas generation that is unpermitted, in litigation, or awaiting a permit, and Texas has halted TCEQ data-center permits pending an ERCOT audit. The sites with the best near-term grades (IREN, Lake Mariner, Cipher) all have grid power already energized on site. Across these ten sites, power is the binding constraint on delivery, not chips and not construction. A contract that does not say whose generation, under which permit, delivers the capacity is not a delivery contract.
The grades use stated priors (DR-1 v0.1), not fitted ones. The point of publishing now is that each forecast resolves on a date and stays in the ledger. A track record cannot be backfilled.
Also shipped
- Harness v0.1.0. The open DCB probe, with run nonces drawn from the public drand randomness beacon, so anyone can check that a result was not backdated. The founding cohort is free for the first ten fleets.
- ASSAY-TT-1, draft. A standard compute tranche and a sublease and assignment term sheet that a provider can pre-approve. Committed compute needs an exit valve, and this is ours. Redlines are welcome.
Method and corrections
- Price checks: each row was re-read with an automated fetch on 2026-10-26. Per-node prices were divided by GPUs per node, as stated in the row's note.
- Price statuses: confirmed means the same price was read today; changed means a different price was read today, and the previous one is kept in
price_previous; unavailable means the offering is no longer listed; not rechecked means the row is untouched. - Delivery evidence: each item is cited on its card, and secondary sources are labelled. We could not tie Anthropic's announced Texas Fluidstack site to a specific campus from primary sources, so it is not graded separately. Its most likely candidate, Barber Lake, is.
- Corrections: send a source and it moves a row or a grade the same week. See corrections.
Data for this issue: registry 2026-10-26 · re-survey with quotes · sites.json · forecast ledger. CC BY 4.0. Not investment advice.