On July 8, Architect Financial Technologies and Compute Desk announced ComputeConnect — the US financial industry’s first compute exchange-for-physical network, linking CFTC-regulated compute derivatives to physical GPU capacity delivery. It is the least glamorous announcement in the compute-markets race this year, and probably the most important one.

The pieces: Architect’s forthcoming American Innovation Exchange (AX) — a US designated contract market for futures and options on compute and AI supply-chain commodities, pending regulatory review — will list futures referencing Compute Desk’s rental-price indexes on NVIDIA H100, H200, B200, and B300. ComputeConnect sits beside the exchange and lets a futures position convert into actual capacity: an open protocol where capacity providers respond to delivery requests, published basis tables for different SKUs, memory configurations, and locations, and futures legs booked back to AX. The physical leg clears through Compute Desk’s Compute Clear platform, which guarantees the operational integrity of delivery.

The announcement’s framing is unusually candid about why this matters. The AI build-out is the largest capex cycle in history — the five largest hyperscalers alone are on track to invest nearly $700 billion in 2026 — and compute, the single largest input cost, has lacked any standardized hedge. The incumbent risk tool, the multi-year offtake agreement, is a blunt one: buyers over-commit on tenor, providers give up the margin that flexible terms would earn, and lenders won’t finance capacity without a large creditworthy name on the contract. Futures fix the standardization problem; the EFP fixes the credibility problem. A cash-settled contract with no physical channel is a bet on an index. A contract that converts into racks is forward procurement — and hedged capacity becomes collateral a lender can underwrite.

What every mature commodity already knows

EFP mechanics are the connective tissue between paper and physical markets everywhere they exist. In WTI crude, the exchange delivery mechanism at Cushing is the constitutional backstop, but the overwhelming majority of physical conversion happens via EFP — two parties agree a physical trade at a negotiated basis and exchange the futures legs, keeping the paper price welded to pipeline reality. In gold, the COMEX-to-loco-London EFP is quoted continuously as a two-way basis and is the standing bridge between the futures market and the OTC bullion market. Henry Hub gas, LME metals — same architecture, different nouns. The futures price stays honest not because delivery happens often, but because it credibly can, at a basis everyone can see.

And the strategic point deserves stating plainly: Architect is first to market with the physical layer. CME × Silicon Data and ICE × Ornn filed cash-settled contracts in May; Kalshi’s event ladders trade, but settle in dollars against an index print. Nobody else has announced a delivery mechanism, a clearing layer for physical capacity, or published basis tables. In every prior commodity, the venue that first connected paper to physical set the delivery spec, the basis conventions, and the location differentials that every later contract had to reference. That first-mover advantage compounds — it is much harder to displace a delivery network than an index.

The explorer below works through the compute translation: the five-step EFP lifecycle as ComputeConnect describes it, a basis-table simulator in the SKU × memory × location shape the network says it will publish, the commodity analog map, and a convergence simulator showing what a physical channel does to the futures basis — and what its absence leaves unresolved.

Silicon Data vs Compute Desk — the contract nuance that decides who hedges where

The two US index providers behind the leading futures designs look similar from a distance — both publish GPU rental-price benchmarks — but the contracts built on them diverge on almost every dimension that matters to a hedger:

The nuance worth underlining: assessment-based versus transaction-referenced indexing is the WTI-versus-Brent argument of the compute complex. An assessed benchmark is smoother and harder to manipulate through a single trade, but embeds methodology judgment; a transaction-referenced index is closer to realized economics but inherits whatever the private deal flow looks like in thin months. Layer the settlement difference on top and the two contracts serve different users: Silicon Data’s design suits financial participants managing index exposure, while Compute Desk’s four-SKU, EFP-backed design targets the operator who ultimately needs the GPUs. If both succeed, the spread between them — assessed vs transacted, cash vs deliverable — becomes its own market, exactly as WTI-Brent did.

What the precedents predict

Convergence will be friction-bounded, not exact. The EFP disciplines the basis to within the all-in cost of taking delivery: workload migration, qualification of the provider, credit terms, and the basis-table spread itself. Expect the compute equivalent of location differentials — a Tier-2 Midwest discount, an EU premium — to emerge as the first persistent, quotable structure. That is where the tradeable market microstructure will live, exactly as Midland-Cushing and the Midwest aluminum premium became markets in their own right.

The basis table is the real product. Publishing standard differentials for SKUs, memory configurations, and locations converts a bilateral, opaque procurement negotiation into a visible curve. Even participants who never touch the futures will mark their offtakes against it — the same way physical crude deals price off posted differentials to the screen.

Delivery quality is the tail risk. In commodities, the delivery spec fight (which grades, which warehouses, whose assay) is where contracts succeed or die. For compute the analogous questions: what interconnect and topology counts as deliverable H100 capacity, how SLAs are enforced mid-delivery, and whether Compute Clear’s operational guarantee holds when a provider fails during a delivery month. The nickel market’s 2022 lesson applies — physical-delivery credibility is only as strong as the clearing layer behind it.

The venue race just gained a second axis. CME × Silicon Data and ICE × Ornn compete on index methodology; Kalshi competes on event-ladder implied forwards; Architect now competes on physicalization — and ICE’s answer, an energy-normalized contract with NATIVX, is the subject of the companion piece. Benchmark battles in young commodities are usually won by whichever contract hedgers can actually use. The EFP is Architect’s argument that its contract is that one.

Sources: Architect / Compute Desk announcement (PR Newswire, July 8, 2026); Architect AX and Compute Desk company materials; exchange rulebooks for WTI, COMEX gold, Henry Hub, and LME delivery mechanics. All Architect products pending regulatory review. Basis values in the explorer are illustrative. Not investment advice.