ICE’s second compute-futures initiative is a different animal from its first. ICE and NATIVX announced GPU compute futures on NATIVX’s COIL Index — which tracks the price of tokenized, energy-normalized compute and connectivity, “normalized to one stable unit and auditable at every step.” USD-denominated, cash-settled, launch expected later this year pending regulatory processes, with sub-indices for training (COIL-T), inference (COIL-I), graphics (COIL-G), and connectivity (COIL-CO).

The design argument, in ICE’s own framing: power is a dominant input cost of large-scale compute, fluctuating electricity prices directly move AI workload economics, and normalizing compute prices to a consistent energy unit “strips out the noise introduced by regional power cost disparities.” The venue argument is just as deliberate — the contracts will trade alongside ICE’s natural gas and power futures, so compute operators can hedge GPU exposure in the same venue where they already hedge energy, and energy traders gain a listed window into the fastest-growing source of electricity demand in the world. This is the first contract design that explicitly fuses the compute and power risk stacks — and it intensifies a benchmark race that now runs five venues deep: CME × Silicon Data, ICE × Ornn, Kalshi, Architect’s AX (with the ComputeConnect physical channel), and now ICE × NATIVX.

The question normalization begs

“Strip out regional power cost” is an index-design decision with a trading consequence: whatever the benchmark treats as noise becomes basis someone else can own. This site already carries the dataset to quantify exactly what gets stripped — 374,550 hourly settlements across ERCOT’s four hubs and PJM’s three, July 2023 through July 2026, built for the hourly power futures piece. The dashboard below reuses it to put numbers under three questions the COIL design raises:

What is the energy content of a GPU-hour? An H100 at nameplate 700W and a PUE of 1.3 consumes ~0.91 kWh per GPU-hour. At ERCOT North’s summer real-time shape, that swings from roughly two cents overnight to nearly twenty at the evening peak — a 10:1 intraday range in the input cost the normalized benchmark deliberately averages away. What is the arb against the normalized view? A flat-normalized benchmark implicitly prices compute at 24×7 average power; an operator who can time-shift training jobs runs at the cheapest-8-hours cost, and the spread between those two numbers — hub by hub, up to double-digit percentages of the energy bill — is a recurring edge the index cannot see, hedgeable now that Nodal’s hourly futures and ICE’s own TB4 list the shape directly. And what is the compute heat rate? Rental divided by energy consumed puts compute’s implied value in $/MWh — thousands of dollars against power in the tens — which is why the interesting trade is not the level but the spread’s volatility, with the compute leg (COIL), the fuel leg (ICE power and gas), and the shape leg (hourly futures) all finally quoting on listed venues.

What to watch

The normalization methodology itself. COIL is described as tokenized, energy-normalized, and “built around maximizing constituent capacity” — the precise formula (which energy unit, whose power price, how connectivity weights in) is the entire ballgame for basis traders, and it is not yet published. When the contract specs file, the first thing to compute is the residual between COIL and the rental-price indexes (Compute Desk, Ornn, Silicon Data) — that residual is the embedded energy position.

The cross-venue benchmark race. Five venues, four index philosophies, no incumbent. Commodity history says one benchmark wins and the others become basis markets against it. The COIL bet is that energy normalization makes the cleanest global comparison unit; the Architect bet is that physical delivery makes the most usable hedge. Both can be right — WTI and Brent split exactly this way.

The integrated-venue effect. Listing compute next to gas and power in one clearing house is a margin story: cross-margining a long-compute / short-power book at a single CCP materially changes the capital cost of the compute spark spread. If ICE grants meaningful offsets, the spread trade gets structurally cheaper at ICE than anywhere else — venue choice becomes part of the trade construction, as it already is in the energy complex.

Sources: ICE / NATIVX announcement (Business Wire, July 2026); NATIVX and Synova Global company materials; hourly power data from ERCOT MIS and PJM Data Miner 2 (this site’s dataset, Jul 2023 – Jul 2026). GPU TDP figures are nameplate; rental marks are illustrative. Contract specifications pending regulatory processes and may change. Not investment advice.