Methodology Change Log
The record of changes to what the published series measure. Pending changes are announced with an estimated impact and a comment window; the realized impact is appended at the effective fixing. Changes apply forward. A restatement is rare, is recorded here with the version every republished row carries, and follows the How We Publish. Comments: research@ccir.io.
v2.4.1 effective: realized decomposition
The five sources announced 2026-10-01 entered the panels at the 2026-10-08 fixing. The figures compare that fixing's own rows with and without them.
The last column gives the price changes between the estimate and 2026-10-08 that explain each gap. No comments were received.
| On-demand cell | Estimated | Realized | Why they differ |
|---|---|---|---|
| Neocloud H100 | 0.0% | 0.0% | |
| Neocloud H200 | +3.6% | +8.0% | Larger rise: a candidate increased. |
| Neocloud B200 | −2.8% | −8.9% | Larger drop: incumbents increased, the candidates held. |
| Neocloud B300 | +2.6% | +3.2% | |
| Neocloud A100 | +4.2% | +6.7% | Larger rise: an incumbent near the middle of the distribution increased. |
| Marketplace H100 | −4.3% | −4.0% | |
| Marketplace H200 | −7.0% | 0.0% | No drop: incumbents and the candidate increased. |
| Marketplace B200 | −6.0% | −10.1% | Larger drop: incumbents increased, the candidate held. |
| Marketplace B300 | +3.9% | +7.9% | Larger rise: an incumbent stopped posting and the candidate increased. |
| Marketplace A100 | −4.2% | −0.8% | Smaller drop: incumbents and the candidate increased. |
On /term, the batch moved every B300 tenor by +2.4% and every B200 tenor by −1.5%.
| Ladder cell | Without | With | Change |
|---|---|---|---|
| B300 on-demand | $8.19 | $8.38 | +2.4% |
| B300 1-year | $6.45 | $6.61 | +2.4% |
| B300 2-year | $6.34 | $6.50 | +2.4% |
| B200 on-demand | $6.69 | $6.59 | −1.5% |
| B200 3-month | $6.59 | $6.49 | −1.5% |
| B200 6-month | $6.42 | $6.33 | −1.5% |
| B200 1-year | $6.42 | $6.33 | −1.5% |
v2.4.1 — five sources join the panels
Five operators that post public prices join the rate and term panels at the 2026-10-08 fixing. Current members add new chips, and price floors keep placeholder prices out. Estimated effects are listed with this entry; realized figures publish at the effective fixing. Comments through 2026-10-07: research@ccir.io.
Effective 2026-10-08 as announced. Realized figures: entry 261009-1.
Contracted rates now weight recent deals
Each contracted cell uses publicly filed contracted rates signed in the last 12 months. A deal's weight falls over those 12 months. The model is GPU weighted, so larger deals carry more weight.
Lower on the page, /term shows these contracted rates against the tenor rates providers post.
Hardware values are now market based
Hardware values now start from the rates at which each chip is being contracted today. They are checked against public fleet sales and retail GPU-only sales, which appear beside each value.
Values are per GPU, all-in: the GPU with its share of the server and networking. GB200 and GB300 NVL72 racks are valued on the same basis.
Contract terms and rates come from recent executed deals, on the same weighting as /term. When a chip's recent deals thin out, its value moves toward re-leasing on the spot market.
v2.4.0 — one construction across the record
Each daily print now reports the prior complete day in Central time and is dated that day. Distinct products a provider posts on one chip are distinct listings; the sizes of one product are one listing, priced at the eight-GPU node where posted. A listing prices the region of its posted location, whichever provider posts it, so hyperscaler capacity in Europe prices the European cells. Every marketplace enters on the same terms. The headline statistic is the operator-equal median at every panel depth; a cell no longer switches to the listing mean when it reaches ten sources.
v2.3.3 effective — realized decomposition
The four additions announced 2026-09-03 took effect at the 2026-09-11 fixing. Realized with/without on that fixing's own rows, blended H100 on-demand rate: $3.58 with (n = 24) vs $3.61 without (n = 23), −0.7% composition. Day-over-day −0.8% is −0.7% composition and −0.04% market on the incumbent panel; use $3.61 for the market leg across the boundary. Blended H200: $4.31 with (n = 14) vs $4.40 without (n = 13), −2.1% composition, market −0.03%.
EU H200: $3.49 with (n = 7) vs $3.74 without (n = 6), −6.7% composition, as estimated. APAC H100: $3.73 with (n = 7) vs $3.77 without (n = 5), −1.1%, as estimated. The market leg on both cells was zero on the day. Two APAC one-month cells reached a first provisional print on two sources each.
The EU H100 cell crossed the ten-source line. A cell of ten or more sources publishes the listing mean; below ten it publishes the operator-equal median. With the addition the cell prints $3.02 as a mean of ten. Without it the cell prints $2.99 as a median of nine, so the published rule shows +1.0%. Like for like the added card sits below the cell: mean against mean −1.1%, median against median −2.3%, the estimated figure. No published series crossed the material threshold. No comments were received by the close.
v2.3.3 — four sources proposed for the on-demand rate panel
Four operators that post public rate cards are proposed for the on-demand rate panel, effective with the 2026-09-11 fixing. Two are European operators. Their rows enter the EU region cells and the blended headline pools. Two are Asia-Pacific operators. Their rows enter their own region cells only. All four have run as capture-only sources since 2026-07-15. Non-USD cards convert at the ECB reference rate at collection, the treatment already applied to the panel's European members.
Estimated effect, from a with/without shadow over the last seven fixings. The blended H100 on-demand rate moves about −0.8% on one added source. The EU H100 and H200 cells move −2.3% and −6.7% as two cards priced below the cell enter. The APAC H100 cell moves −1.0% on two added sources. Two APAC one-month cells reach a first provisional print. No published series crosses the material threshold. Realized with/without publishes at the effective fixing. Comments through 2026-09-10: research@ccir.io.
The comment window closed on 2026-09-10 with no comments received. The four sources entered the panel at the 2026-09-11 fixing as announced. The realized with/without is the entry dated 2026-09-11.
v2.3.1 — a quorum bound on the marketplace vote floor
The v2.3.0 vote floor applies source by source. In cells where most members are marketplace-class, it can remove most of the panel's votes at once and leave the headline to a small minority. From the 2026-08-19 fixing the floor carries a quorum bound. Where the floor would remove half or more of a cell's voters, the floored votes stand for that day. The full panel then prices the cell. Cells where floored sources are a minority follow v2.3.0 unchanged. The bound applies from the 2026-08-19 fixing itself. Two marketplace-tier cells first published that morning under v2.3.0 were re-published under the bound the same day, before the daily record closed. Forward-only; no prior series value is restated.
v2.3.0 — single-point marketplace medians no longer vote in the operator-equal aggregate
The headline is the median of per-source medians, one vote per source. At marketplace-class sources that vote can form over a single surviving listing when catalog churn empties the rest of the venue's cross-section — one host's asking price then moves the aggregate by a full source vote, and reverts when the listing rotates out. From the 2026-08-19 fixing, a marketplace-class source votes only on days its median formed over at least two distinct price points. The floor governs the vote alone: the source's rows still stand in every distribution field, count, and concentration measure, and a cell whose only sources are floored keeps its median unchanged. Forward-only; no series is restated.
v2.2.1 effective — realized decomposition
The additions announced 2026-07-15 took effect. Realized with/without on the effective fixing: $3.68 with (n = 14) vs $3.62 without (n = 12), +1.8% composition. Day-over-day +1.6% = +1.8% composition − 0.3% market; use $3.62 for the market leg across the boundary. The retirement contributed 0.0% by construction. No comments were received by the close.
Amended 2026-07-23: one flagged row — a price block whose hardware identity rested on page-markup inference — entered two A100 cells at this fixing contrary to the announcement's terms. Published headlines were unaffected (median-rule cells); the affected cells' count and maximum fields carried the row for one fixing. Rows carrying that flag are excluded from every cell from the 2026-07-24 fixing, forward-only. The operator's parent publishes an operational rate table that disambiguates the block's identity; the A100-80G on-demand rate enters from that surface at the venue-labeled price from 2026-07-24.
v2.2.1 — two sources proposed for the on-demand rate panel; one member retired
Two neocloud-class sources are proposed for the on-demand rate panel. One is a subsidiary of an existing member with a materially different live rate card, admitted as a separate venue (no splice). Estimated impact: the two entries largely offset (−3.0% and +5.0% alone). Realized with/without publishes at the effective fixing. Comments through 2026-07-22: research@ccir.io.
Amended 2026-07-16: one panel member is retired from the roster at the same fixing. The member has published no purchasable rates since before the panel froze and has never contributed an observation to a published cell; no series is affected.
v2.2.0 — two sources added to committed-term cells
Two sources with firm published term prices (1/3-month and 3-year tenors) join the committed-term cells. Enhanced per-region capture of hyperscaler reservation pricing (1/3/5-yr) begins as a cross-check surface with no separate panel vote. On-demand rows from term sources enter no on-demand cell.
Candidate-source observation begins
Capture-only observation of candidate sources (rate cards, committed-term surfaces, marketplace feeds) began under panel provenance rules. No candidate holds a panel vote; proposals follow the Change Management Policy.
v2.1.0 effective — realized decomposition
The addition announced 2026-07-10 took effect. Realized with/without on the effective fixing: $3.60 with (n = 12) vs $3.52 without (n = 11), +2.3% composition. Day-over-day +1.0% = +2.3% composition − 1.3% market; use $3.52 for the market leg across the boundary. No comments were received by the close.
v2.1.0 — one source added to the on-demand rate panel
One source posting guaranteed on-demand and interruptible H100 and B200 rates joins the neocloud panel. Its own capacity only: re-sold capacity of existing members is excluded, one operator one vote. A corporate affiliate remains a separate panel source while both rate cards stay live; no splice. Comments were open through 2026-07-14.
July 1–2 panel expansion — collection panel 31 → 41 sources
Eleven sources entered over 2026-07-01/02 and several cells stepped where new price levels entered. The steps are composition effects, not repricing; treat week-over-week comparisons across the boundary accordingly.
Fourth source enters the hyperscaler H100 US cell
A fourth source's on-demand H100 US rows began flowing 2026-07-02/03. The step is a composition effect — no panel source changed a posted rate on those dates.
v2.0.0 — first public methodology
Chip × operator-class × geography series grammar, operator-equal weighting, published headline-statistic rule, frozen panel roster. Every published row carries its methodology version.