TCI‑CRI‑H200H200 SXM 141GB
——never published
below the provider gate
Independent. Transparent.
Built for the compute
market.
TCI publishes reproducible reference prices for AI and HPC compute — methodology-driven, vendor-neutral, and rebuildable from public sources by anyone.
Headline index
TCI‑CRI‑H100
NVIDIA H100 SXM 80GB, on-demand, EU/EEA regions. Weighted median over offers, per GPU-hour, ex-VAT.
3.993.56/GPU‑hr
±0.00%+0.00vs 9 Oct 2026 print
Index live · 8 of 5 providersHash‑locked v0.11.02 executable inputs
Headline series — today
One row per class the index prices, for the 10 Oct 2026 session. A class below the provider gate keeps its row and states the gate it missed; the last good value is never promoted into today's slot.
Live · updated daily 11:00 UTC
| Series | Region | $/GPU‑hr | 30d Δ |
|---|---|---|---|
| TCI‑CRI‑H100 | EU/EEA | $3.99 | +22.8% |
| TCI‑CRI‑H200 | EU/EEA | — | gap · below the provider gate |
| TCI‑CRI‑B200 | EU/EEA | — | gap · below the provider gate |
| TCI‑CRI‑B300 | EU/EEA | — | gap · below the provider gate |
| TCI‑CRI‑A100 | EU/EEA | — | gap · below the provider gate |
| TCI‑CRI‑H100‑PCIE | EU/EEA | — | gap · below the provider gate |
Sub-indices
The segment and variant cuts of the headline. A series below the provider gate keeps its slot and shows a gap with the reason.
TCI‑CRI‑B300B300 SXM
——never published
below the provider gate
TCI‑CRI‑A100A100 SXM 80GB
——never published
below the provider gate
TCI‑CRI‑H100‑PCIEH100 PCIe — priced as its own class
——never published
below the provider gate
TCI‑CRI‑H100‑SOVEU/EEA-headquartered operators
——never published
below the provider gate
TCI‑CRI‑H100‑MKTMarketplace segment only
——never published
below the provider gate
TCI‑CRI‑H100‑NCNeocloud segment only
3.93
±0.00%
TCI‑CRI‑H100‑HSHyperscaler catalog segment
——never published
below the provider gate
EU, US and Global, 30 sessions
The same H100 calculation over three regions: EU/EEA, the headline; the United States; and every country. Truncated y-axis, so this is a line and never an area fill. Hover or tab through the plot for the crosshair readout; it is CSS-only and works with scripting disabled.
TCI‑CRI‑H100
USD per GPU-hour · 30 sessions to 10 Oct 2026 · 28 published, 2 gapped · today $3.99 from 8 providers
A gapped session is drawn as a hairline tick on the baseline and the line breaks across it. Nothing is interpolated: the index publishes a gap rather than a value it cannot defend. The same calculation, replayed at every hourly read.
Table view — 30 sessions
| Session | USD/GPU-hr | State |
|---|---|---|
| 11 Sep 2026 | — | gap · below the provider gate |
| 12 Sep 2026 | 3.49 | published |
| 13 Sep 2026 | 3.25 | published |
| 14 Sep 2026 | 3.25 | published |
| 15 Sep 2026 | 3.25 | published |
| 16 Sep 2026 | — | gap · no print computed for this session |
| 17 Sep 2026 | 3.49 | published |
| 18 Sep 2026 | 3.28 | published |
| 19 Sep 2026 | 3.49 | published |
| 20 Sep 2026 | 3.28 | published |
| 21 Sep 2026 | 3.49 | published |
| 22 Sep 2026 | 3.49 | published |
| 23 Sep 2026 | 3.49 | published |
| 24 Sep 2026 | 3.78 | published |
| 25 Sep 2026 | 3.76 | published |
| 26 Sep 2026 | 3.49 | published |
| 27 Sep 2026 | 3.49 | published |
| 28 Sep 2026 | 3.64 | published |
| 29 Sep 2026 | 3.56 | published |
| 30 Sep 2026 | 3.49 | published |
| 1 Oct 2026 | 3.49 | published |
| 2 Oct 2026 | 3.49 | published |
| 3 Oct 2026 | 3.49 | published |
| 4 Oct 2026 | 3.49 | published |
| 5 Oct 2026 | 3.72 | published |
| 6 Oct 2026 | 3.71 | published |
| 7 Oct 2026 | 3.73 | published |
| 8 Oct 2026 | 3.70 | published |
| 9 Oct 2026 | 3.99 | published |
| 10 Oct 2026 | 3.99 | published |
TCI‑CRI‑H100‑US
USD per GPU-hour · 10 sessions to 10 Oct 2026 · 6 published, 4 gapped · today $3.99 from 5 providers
A gapped session is drawn as a hairline tick on the baseline and the line breaks across it. Nothing is interpolated: the index publishes a gap rather than a value it cannot defend. The same calculation, replayed at every hourly read.
Table view — 30 sessions
| Session | USD/GPU-hr | State |
|---|---|---|
| 1 Oct 2026 | — | gap · below the provider gate |
| 2 Oct 2026 | — | gap · below the provider gate |
| 3 Oct 2026 | — | gap · below the provider gate |
| 4 Oct 2026 | — | gap · below the provider gate |
| 5 Oct 2026 | 3.99 | published |
| 6 Oct 2026 | 3.99 | published |
| 7 Oct 2026 | 3.99 | published |
| 8 Oct 2026 | 3.99 | published |
| 9 Oct 2026 | 3.99 | published |
| 10 Oct 2026 | 3.99 | published |
TCI‑CRI‑H100‑GLOBAL
USD per GPU-hour · 7 sessions to 10 Oct 2026 · 7 published, 0 gapped · today $3.93 from 12 providers
A gapped session is drawn as a hairline tick on the baseline and the line breaks across it. Nothing is interpolated: the index publishes a gap rather than a value it cannot defend. The same calculation, replayed at every hourly read.
Table view — 30 sessions
| Session | USD/GPU-hr | State |
|---|---|---|
| 4 Oct 2026 | 3.49 | published |
| 5 Oct 2026 | 3.72 | published |
| 6 Oct 2026 | 3.71 | published |
| 7 Oct 2026 | 3.73 | published |
| 8 Oct 2026 | 3.70 | published |
| 9 Oct 2026 | 3.99 | published |
| 10 Oct 2026 | 3.93 | published |
Constituents at the 2026-10-10 print
Every candidate the index saw, including the ones it rejected and why. Tier is ordinal: L1 executable quote, L2 neocloud list, L3 hyperscaler catalog.
| Provider | Segment | Tier | Price USD/GPU‑hr | Weight | Flags | Source |
|---|---|---|---|---|---|---|
| seeweb | neocloud | L2 | 2.1179 | 10.0% | TRIMSOV | seeweb |
| vast.ai | marketplace | L1 | 2.8970 | 20.0% | TRIM | vast_ai |
| scaleway | neocloud | L2 | 3.5778 | 10.0% | SOV | scaleway |
| verda | neocloud | L2 | 3.9275 | 10.0% | SOV | gpuhunt |
| lambdalabs | neocloud | L2 | 3.9900 | 10.0% | — | gpuhunt |
| runpod | marketplace | L1 | 3.9900 | 20.0% | — | runpod |
| digitalocean | neocloud | L2 | 4.4100 | 10.0% | TRIM | digitalocean |
| nebius | neocloud | L2 | 4.5000 | 10.0% | TRIMSOV | gpuhunt |
| aws | hyperscaler | L3 | — | 0.0% | OOP | page |
| azure | hyperscaler | L3 | — | 0.0% | OOP | page |
| coreweave | neocloud | L2 | — | 0.0% | NOT IN PAN | page |
| gcp | hyperscaler | L3 | — | 0.0% | OOP | page |
| oci | hyperscaler | L3 | — | 0.0% | OOP | page |
| Weighted median over offers | 3.9900 | 100.0% | Trim k=1 each end · 25% concentration cap | |||
What the sellers declare
The same GPU-hour from a hyperscaler and from a marketplace is the same unit and not the same product. This is the part of that difference the sellers state themselves, as they state it, in their own units. Where a seller publishes nothing, the cell says so.
| Provider | Interconnect | GPUs per node | GPU memory | vCPUs | System memory | Local storage | Contract form |
|---|---|---|---|---|---|---|---|
| seeweb | NVLink (read from the product name, not published) | 8 (read from the product name, not published) | not published | not published | not published | not published | published rate card |
| vast.ai | NVLink, 478 GB/s as measured by the host | varies across 2 offers | 80 GB | not published | not published | not published | bookable at this price now |
| scaleway | NVLink (read from the product name, not published) | varies across 2 offers | not published | not published | not published | not published | published rate card |
| verda | NVLink (read from the product name, not published) | varies across 9 offers (read from the product name, not published) | 80 GB | not published | not published | not published | published rate card |
| lambdalabs | NVLink (read from the product name, not published) | varies across 54 offers (read from the product name, not published) | 80 GB | not published | not published | not published | published rate card |
| runpod | NVLink (read from the product name, not published) | 4 (read from the product name, not published) | 80 GB | not published | not published | not published | bookable at this price now |
| digitalocean | not published | varies across 6 offers (read from the product name, not published) | not published | not published | not published | not published | published rate card |
| nebius | NVLink (read from the product name, not published) | 8 (read from the product name, not published) | 80 GB | not published | not published | not published | published rate card |
Of 56 cells across the constituents of this print, 16 carry something the seller publishes as a field (29%). A value marked ‡ is read from the product name and is this index's reading rather than the seller's statement. Nothing here is inferred from a neighbouring offer, and nothing here enters the calculation.
Data quality and sources
The publication gate, the executable share, and the collectors behind them. A gap is credible; a fabricated print is fatal.
Data quality
Everything that decides whether this session publishes at all
Publication gate
8 / 5
Qualifying providers against the minimum. Below it the value is
null and the session is flagged, never estimated.
Executable share
40.0%
Share of index weight from executable marketplace quotes;
the rest is published list price.
Session coverage
28 / 30
Sessions in the last 30 days that cleared the gate.
The remainder are published as gaps.
FX vintage
T−1
ECB reference rate 1.1206 dated
2026-10-09. The ECB publishes after the cut-off, so the EUR leg is
T−1 by construction.
Concentration cap
SLACK
The 25% per-provider cap did not bind on this print; it is
mathematically inert at n=4 and binds only from n=5.
Trim
k=1
Count-based: the k highest and k lowest offers are clamped to the
k-th order statistic. Percentile winsorising is inert at this panel size.
Model mix
Composite weight by GPU class, effective 2026-10-05 · same basket as the ledger below
- H10039.6%
- A10020.0%
- H20017.4%
- B20012.1%
- B3005.9%
- H100P5.0%
Composite basket, effective 2026-10-05
Window 2026-09-07 to 2026-10-04 · 24 collection days · recomputed by a fixed published formula, stored append-only
- 0139.6%
H100 class share
Share of total observed qualifying capacity over the review window, floored at 5% and capped at 75%.
- 0220.0%
A100 class share
Share of total observed qualifying capacity over the review window, floored at 5% and capped at 75%.
- 0317.4%
H200 class share
Share of total observed qualifying capacity over the review window, floored at 5% and capped at 75%.
- 0412.1%
B200 class share
Share of total observed qualifying capacity over the review window, floored at 5% and capped at 75%.
- 055.9%
B300 class share
Share of total observed qualifying capacity over the review window, floored at 5% and capped at 75%.
- 065.0%
H100P class share
Share of total observed qualifying capacity over the review window, floored at 5% and capped at 75%.
Sources
Every collector the index reads. Overlay data (power prices) is published beside the index and never enters the calculation.
static_yamlManually verified public pricing pages
config/providers/*.yaml, refreshed manually with a last_verified date
SiteHow the number is made
Three commitments, each one checkable against the repository rather than taken on trust.
Reproducible
Every parameter is a visible config value, and the five files that can change a print are hash-locked. Every print can be rebuilt from public sources by anyone with the repo.
Vendor-neutral
A weighted median over offers from marketplaces, neoclouds and sovereign operators, gated at 5 providers, with a 25% concentration cap — no single source sets the price.
Open methodology
Design follows the IOSCO Principles for Financial Benchmarks as voluntary best practice — published, versioned, and change-controlled, with the full constituent set behind every print.
A family of indices
One index is published today. The rest are stated as what they are — in development or planned — and carry no numbers until they do.
TCI‑CRI
Compute Reference Index — GPU-hour pricing across 6 hardware classes, published daily with its full constituent set.
TCI‑SRI
Storage Reference Index — reference pricing for high-throughput storage, the other metered cost on a training run. No data is collected for this yet.
Latest research
Notes on what the index measures and what it cannot.