TCI‑CRI‑H100$3.49TCI‑CRI‑H100‑7D— GAPTCI‑CRI‑H100‑MKT— GAPTCI‑CRI‑COMPUTE107.4TCI‑CRI‑H100$3.49TCI‑CRI‑H100‑7D— GAPTCI‑CRI‑H100‑MKT— GAPTCI‑CRI‑COMPUTE107.4TCI‑CRI‑H100$3.49TCI‑CRI‑H100‑7D— GAPTCI‑CRI‑H100‑MKT— GAPTCI‑CRI‑COMPUTE107.4
AS OF 2026-09-12 11:08:37 UTC
The Compute Indices
EU–US basis

What a European buyer carries when the hedge is priced in the US.

CME Group plans to list compute futures on 5 October 2026, pending regulatory review, that settle in cash on Silicon Data's H100 and B200 rental indices. A European buyer who hedges with them is exposed to the difference between what the EU/EEA population of sellers charges and what the reference population charges. TCI‑CRI‑H100‑BASIS‑US is that difference for one H100 SXM GPU-hour: the EU/EEA headline minus a US series priced with exactly the same rules.

First print 1 Oct 2026 Methodology Notice 2026-N3 latest.json

The two legs and the spread

Each leg publishes only with five qualifying sellers; the spread publishes only when both legs do.

TCI‑CRI‑H100EU/EEA
3.49
6 providers · 12 Sep 2026
TCI‑CRI‑H100‑USUnited States
——
Not yet computed
TCI‑CRI‑H100‑BASIS‑USEU minus US
——
Not yet computed

The month-to-date average starts with the first print on 1 Oct 2026. Each CME contract covers a month of rent (730 GPU-hours), so the month is the unit a hedger's basis is measured in.

Last 60 sessions

The basis publishes from 1 Oct 2026 under methodology v0.6.0 (notice 2026-N3). There is no value before that date, and none is back-filled.

What the spread measures, and what it does not

Both legs use the same unit definition, node floor, weighted median over offers, trim, tier weights, concentration cap and publication gate. Only the region differs, which is what makes the spread a regional basis rather than a comparison of two methods.

It is not the basis to the Silicon Data index the CME contracts settle on. That index's methodology is not public, and a spread against it would mix a regional difference with a methodological one that nobody outside can measure. What is published here is the regional part, with the method held constant.

The US leg draws on five candidate sellers (vast.ai, RunPod, Lambda, DigitalOcean, Voltage Park) against a gate of five. On any day one of them has no qualifying offer, the US leg gaps and so does the spread, with the reason in the audit set.

RunPod and DigitalOcean charge one price in every region, so on a day both qualify in both legs the same two prices sit on each side of the spread. That pulls the basis toward zero, and it is correct: a buyer can rent from either at the same price on either side of the Atlantic. A basis driven by the other sellers is the one that reflects regional pricing.

Sellers admitted to the H100 class, v0.6.0
SellerSegment CollectorH100 SXM rows, last 30 days
digitaloceanneoclouddigitaloceanEU/EEA + US
genesis_cloudneocloudstatic_yamlnot yet seen
hetznerneocloudstatic_yamlnot yet seen
lambdalabsneocloudgpuhuntEU/EEA + US
leasewebneocloudstatic_yamlnot yet seen
nebiusneocloudgpuhuntEU/EEA
runpodmarketplacerunpodEU/EEA + US
scalewayneocloudscalewayEU/EEA
seewebneocloudstatic_yamlEU/EEA
vast.aimarketplacevast_aiEU/EEA + US
verdaneocloudgpuhuntEU/EEA
voltageparkneocloudvoltageparkUS

Sellers in both legs

Shown on each day the basis publishes: the sellers that qualify in both legs, and what each charged on each side. There is no basis print this session.

Where a region-flat price is placed

  • vast.ai: every offer carries its own location.
  • Lambda: catalogue rows carry a region (europe-central-1 is Germany).
  • DigitalOcean: one price, recorded once for each region its availability page lists for the H100 (Amsterdam, New York, Toronto).
  • RunPod: one price everywhere. A US row is recorded only on a day RunPod reports stock of that GPU type in a US datacentre.
  • Voltage Park: sells only in the United States, by its own statement.