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 series — today
One row per class the index prices, for the 7 Sep 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 | $3.25 | ±0.00% |
| TCI‑CRI‑H200 | EU/EEA | — | gap · below the provider gate |
| TCI‑CRI‑B200 | EU/EEA | — | not computed |
| 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 |
Headline index
TCI‑CRI‑H100
NVIDIA H100 SXM 80GB, on-demand, EU/EEA regions. Weighted median over offers, per GPU-hour, ex-VAT.
3.252.80/GPU‑hr
±0.00%+0.00vs 6 Sep 2026 print
Index live · 5 of 5 providersHash‑locked v0.3.0-dev4 executable inputs
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
——never published
below the provider gate
TCI‑CRI‑H100‑HSHyperscaler catalog segment
——never published
below the provider gate
Headline, 30 sessions
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 7 Sep 2026 · 16 published, 14 gapped
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.
Table view — 30 sessions
| Session | USD/GPU-hr | State |
|---|---|---|
| 9 Aug 2026 | — | gap · below the provider gate · revised |
| 10 Aug 2026 | 3.25 | published |
| 11 Aug 2026 | — | gap · below the provider gate · revised |
| 12 Aug 2026 | — | gap · below the provider gate · revised |
| 13 Aug 2026 | — | gap · below the provider gate · revised |
| 14 Aug 2026 | — | gap · below the provider gate · revised |
| 15 Aug 2026 | — | gap · below the provider gate · revised |
| 16 Aug 2026 | 3.29 | published |
| 17 Aug 2026 | 3.29 | published |
| 18 Aug 2026 | — | gap · below the provider gate |
| 19 Aug 2026 | — | gap · below the provider gate |
| 20 Aug 2026 | 3.29 | published |
| 21 Aug 2026 | — | gap · below the provider gate |
| 22 Aug 2026 | 3.29 | published |
| 23 Aug 2026 | 3.29 | published |
| 24 Aug 2026 | 3.29 | published |
| 25 Aug 2026 | 3.29 | published |
| 26 Aug 2026 | — | gap · below the provider gate |
| 27 Aug 2026 | 3.29 | published |
| 28 Aug 2026 | 3.25 | published |
| 29 Aug 2026 | 3.29 | published |
| 30 Aug 2026 | 3.25 | published |
| 31 Aug 2026 | 3.29 | published |
| 1 Sep 2026 | — | gap · below the provider gate |
| 2 Sep 2026 | — | gap · below the provider gate |
| 3 Sep 2026 | — | gap · below the provider gate |
| 4 Sep 2026 | — | gap · below the provider gate |
| 5 Sep 2026 | 3.25 | published |
| 6 Sep 2026 | 3.25 | published |
| 7 Sep 2026 | 3.25 | published |
Constituents at the 2026-09-07 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 | 2.1600 | 16.7% | TRIMSOV | static_yaml |
| vast.ai | marketplace | L1 | 2.2669 | 25.0% | CAP | vast_ai |
| datacrunch | neocloud | L2 | 3.2500 | 16.7% | SOV | static_yaml |
| runpod | marketplace | L1 | 3.4900 | 25.0% | CAP | runpod |
| nebius | neocloud | L2 | 3.8500 | 16.7% | TRIMSOV | static_yaml |
| aws | hyperscaler | L3 | — | 0.0% | OOP | page |
| azure | hyperscaler | L3 | — | 0.0% | OOP | page |
| gcp | hyperscaler | L3 | — | 0.0% | OOP | page |
| Weighted median over offers | 3.2500 | 100.0% | Trim k=1 each end · 25% concentration cap | |||
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
5 / 5
Qualifying providers against the minimum. Below it the value is
null and the session is flagged, never estimated.
Executable share
50.0%
Share of index weight from executable marketplace quotes;
the rest is published list price.
Session coverage
16 / 30
Sessions in the last 30 days that cleared the gate.
The remainder are published as gaps.
FX vintage
T−3
ECB reference rate 1.1622 dated
2026-09-04. The ECB publishes after the cut-off, so the EUR leg is
T−1 by construction.
Concentration cap
BOUND
The 25% per-provider cap bound 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 2026-09-07 · same basket as the ledger below
- H10068.0%
- A10016.3%
- H100P7.9%
- H2007.9%
Composite basket, effective 2026-09-07
Window 2026-08-10 to 2026-09-06 · 28 collection days · recomputed by a fixed published formula, stored append-only
- 0168.0%
H100 class share
Share of total observed qualifying capacity over the review window, floored at 5% and capped at 75%.
- 0216.3%
A100 class share
Share of total observed qualifying capacity over the review window, floored at 5% and capped at 75%.
- 037.9%
H100P class share
Share of total observed qualifying capacity over the review window, floored at 5% and capped at 75%.
- 047.9%
H200 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‑ERI
Energy Reference Index — the power cost behind the compute cost. Day-ahead power prices are already collected as an overlay; they do not enter any index value.
TCI‑SRI
Storage Reference Index — reference pricing for high-throughput storage. No data is collected for this yet.
Latest research
Notes on what the index measures and what it cannot.