CME and ICE have announced compute futures but nothing trades yet. Until it does, the observable market is spot rental indexes and term-rental rates. This page tracks both — and the forward prices they imply.
On May 12, 2026, CME Group announced plans to launch cash-settled compute futures on Silicon Data's daily GPU rental indexes, pending regulatory review. A week later, ICE announced GPU compute futures settled on Ornn's transaction-based Compute Price Index (OCPI). Both are expected to list in late 2026. Until contracts trade there are no market futures prices — what exists is the underlying cash market: daily spot rental indexes, and multi-year reserved (term-rental) rates from which forward prices can be inferred. This page tracks the publicly published values of both.
Sources: CME Group press release, May 12, 2026 · ICE press release, May 19, 2026
The two settlement benchmarks disagree — by design. Silicon Data's indexes are survey-based aggregations of standardized rental observations; Ornn's OCPI is built from executed transactions on its exchange. The gap between them is exactly the basis risk a hedger settling against one while renting at the other would carry. Same discipline as the rest of this site: the two lanes are shown side by side, never blended.
Silicon Data values are each index page's published current reading (neo-cloud tickers where listed); Ornn values are pulled from the public OCPI API. Note the H200 divergence — $3.10 vs $4.78 on the same day — a live illustration of index basis risk: a hedger settling against one benchmark while renting at the other carries that gap. Silicon Data's own pages also display different H100/A100 figures in their cross-index widgets than in their per-index readings; where they conflict, this table uses the per-index reading and links the exact page.
Compute cannot be stored, so futures cannot be priced off today's spot by cash-and-carry. The economically meaningful forward information lives in term-rental contracts: multi-year reserved rates are fixed strips of future compute, and the difference between adjacent strips implies a forward price. The only fully public, like-for-like term structure comes from hyperscaler list prices. Below: AWS on-demand vs reserved rates, converted to $/GPU-hour (instance price ÷ 8 GPUs).
All rates: AWS us-east-1, Linux, list prices as of Aug 9, 2026, via instances.vantage.sh (open-source mirror of AWS pricing).
A term rate is an average over its window, so the marginal forward for a later window is the difference of adjacent strips: F(yr 2–3) = (36 × Π₃₆ − 12 × Π₁₂) ÷ 24, following Bandi & Su (2026), Eq. 3, no discounting. Values below are computed in your browser from the table above, so they always reflect the latest published rates.
| GPU | Window | Implied avg forward, $ / GPU-hr | Read |
|---|---|---|---|
| Computing… |
Two caveats, per the research: reserved contracts bundle price insurance with capacity locking, so these synthetic forwards are upper bounds on where financial futures should price. And hyperscaler list prices sit well above neo-cloud indexes — the level is not comparable to the spot table above; the shape of the curve (deep term discounts, i.e. backwardation) is the signal.
Bandi & Su (2026) built the first synthetic compute futures return panel from Silicon Data's licensed term-rental curves (data through April 2026). Two published findings matter most for anyone pricing the upcoming contracts. First, average hold-to-maturity returns on long synthetic futures were positive across nearly every GPU generation and maturity — consistent with a positive risk premium paid by compute providers hedging revenue, meaning futures should price below expected spot. Second, on a per-unit-of-compute basis newer GPUs rent cheaper: the market prices roughly two H100-equivalents of B200 compute at a discount to one H100.
Annualized average hold-to-maturity returns on long synthetic compute futures, by GPU and maturity. Data: Bandi & Su (2026), Table 3, Panel B; sample Mar 2025 – Apr 2026. Long-maturity estimates rest on few observations; the negative B200 3-month reading is attributed by the authors to the run-off of the physical-access wedge, not a negative premium.
Rental price per H100-equivalent unit of compute (peak dense 8-bit throughput conversion via Epoch AI's GPU dataset). Data: Bandi & Su (2026), Table 2, end of sample (Apr 2026): A100 $4.54, H100 $2.51, B200 $2.02 per H100-equivalent-hour.
The full papers behind this page — compute asset pricing, forward-curve models, and the electricity-market template — are on the Research page. What is deliberately not shown here: historical index series from Silicon Data or Ornn, which are licensed products — only their published current values, linked at the source.