Delivered Compute Report — #1

2026-09-10 · Compute Assay · spec v0.1

Three exchanges are launching cash-settled GPU price futures. Every one of them tells you what compute costs. None of them tells you what you'd actually receive. This report covers the second question: what capacity really exists, how it's wired, and whether it can do the job the price implies.

Surveyed: 16 providers, 70 offerings (public rate cards and docs only — everything cited).

Headline prices ($/GPU-hr, on-demand)

The same H100-SXM sells for $1.99 and for $12.29 this week — a 6.2× gap. A price ratio that size between sellers of "the same" good is not a spread; it is several different goods sharing a name. The table below is why.

The deliverability gap

We grade every offering on whether it can actually run the workload the price implies (spec §4: scale-up domain × scale-out fabric × contiguity):

28 of 70 surveyed offerings are verifiably training-class. 9% publish nothing about interconnect — capacity you cannot underwrite, price, or write a delivery contract against. That opacity is the market's missing half.

Cheapest verifiable training-class capacity (Grade A/B)

Basis watch: the financial layer

• Silicon Data H100 Rental Index (SDH100RT): live, published daily — $2.53 per GPU-hour (Neo-Cloud reading displayed on product page) (source)
• Silicon Data H100 Hyperscaler On-Demand Index: live, separate hyperscaler-tier reading — $7.48 (April 20, 2026 print); traded $7.40-$7.52 Mar 1-Apr 20, 2026; roughly 3x the Neocloud tier (source)
• CME Group Compute futures (with Silicon Data): confirmed announced for October 5, 2026 listing on NYMEX, pending regulatory review; not yet trading as of research date — Two contracts: Silicon Data H100 Rental Index Futures and Silicon Data B200 Rental Index Futures; each represents a month's worth of rent for the respective GPU. CME's compute-futures product page (cmegroup.com/markets/energy/power/compute-futures.html) repeatedly timed out, so full contract specs (size, tick, codes, settlement mechanics) are unconfirmed beyond the press release. (source)
• ICE GPU compute futures on Ornn Compute Price Index (OCPI): announced May 19, 2026; subject to regulatory approval; no listing date given in the release; no confirmed launch found as of 2026-09-10 — Suite of USD-denominated, cash-settled futures; OCPI series may cover H100, H200, B200, RTX 5090, with more GPU types to follow; OCPI is transaction-based and distributed on Bloomberg Terminal (source)
• Architect (AX) GPU/DRAM perpetual futures: announced January 21, 2026, pending regulatory approval; launch described as 'imminent' in the release; live-trading status as of Sept 2026 not confirmed by a fetched primary source — Perpetual futures on daily GPU rental prices and DRAM prices, referencing Ornn indexes; tradable on margin with fiat USD or USD stablecoin collateral on the AX exchange (source)
• ComputeConnect - Architect x Compute Desk exchange-for-physical (EFP) network: announced July 8, 2026; exchange described as forthcoming and pending regulatory review; this is the main physical-delivery compute contract effort found — EFP mechanism converting compute futures positions into real GPU capacity, giving physical settlement to centrally cleared compute futures on the American Innovation Exchange (source)
• SF Compute (San Francisco Compute) - spot/forward GPU cluster marketplace: live marketplace with buy/resell of cluster time (a deliverable-compute market, though not an exchange-listed futures product) — Displayed average $1.95/gpu/hr vs $3.00/gpu/hr reserved; contracts from one node for one hour to thousands of nodes for multiple years (source)

All of the above settle in cash. None deliver a GPU. The gap between the index print and what you can actually take delivery of — the basis — is what this report exists to measure.

What sellers actually publish about their fabric

Verified claims only — each fetched from the operator's own pages this week:

• CoreWeave — CoreWeave docs state its clusters use NVIDIA Quantum-2 (400G NDR) and Quantum-X800 (800G XDR) InfiniBand in a non-blocking fat-tree architecture. (source)
• CoreWeave — B200 instances get a 400G NDR non-blocking Quantum-2 fabric with eight ConnectX-7 HCAs per node (i.e., one 400G rail per GPU). (source)
• CoreWeave — CoreWeave's networking product page claims 300k+ GPU megaclusters built from interconnected 100,000-GPU superclusters with one-to-one non-blocking SHARP-enabled InfiniBand. (source)
• CoreWeave — GB200 NVL72 instances are deployed as full 18-node racks (72-GPU NVLink domain) with a shared nvlink.domain label, connected by 400 Gb/s NDR Quantum-2 InfiniBand. (source)
• Lambda (1-Click Clusters) — 1-Click Clusters span 16 to 512 H100/B200 GPUs on a Quantum-2 400 Gb/s non-blocking, rail-optimized InfiniBand fabric at up to 3200 Gb/s per node. (source)
• Crusoe — Crusoe docs define IB networks as machines on one physical fabric per location, using NVIDIA Mellanox InfiniBand; the instance table lists 3200 Gbps total IB bandwidth for h100-80gb-sxm-ib.8x and h200-141gb-sxm-ib.8x (1600 Gbps for A100), with P-Key-style partitions (max five per network) for isolation. (source)
• Crusoe — Crusoe publicly describes its IB fabric as rail-optimized via its NVIDIA GTC 2024 session; the rail-optimized claim does not appear in Crusoe's own docs pages I fetched. (source)
• Nebius — Nebius docs state each GPU has a dedicated 400 Gbps NIC, totaling 3.2 Tbps per 8-GPU node, on fabrics with limited per-fabric GPU capacity; the marketing page calls it a non-blocking Quantum-2 InfiniBand fabric. (source)
• Oracle OCI — OCI shape docs list per-node RDMA bandwidth: BM.GPU.H200.8 and BM.GPU.B200.8 at 8 x 400 Gbps RDMA; BM.GPU.H100.8 at 8 x 2 x 200 Gbps RDMA; BM.GPU.A100-v2.8 at 16 x 100 Gbps RDMA. (source)
• Oracle OCI — OCI cluster networks provide a bare-metal RDMA fabric with single-digit-microsecond latency between physically co-located nodes. (source)
• Azure (ND-series) — ND H100 v5 scales to thousands of GPUs with 3.2 Tbps per VM, one dedicated 400 Gbps Quantum-2 CX7 InfiniBand link per GPU, and NVLink 4.0 as the intra-VM (8-GPU) domain. (source)
• AWS (P5/P6 UltraClusters) — EC2 UltraClusters are AZ-co-located instances on a petabit-scale nonblocking EFA fabric; AWS does not use InfiniBand for this tier. (source)
• AWS (P5/P6 UltraClusters) — P5-family instances deliver up to 3,200 Gbps via second-generation EFA with GPUDirect RDMA, scale to 20,000 H100/H200 GPUs in an UltraCluster, and have a 900 GB/s NVSwitch intra-instance (8-GPU) NVLink domain. (source)
• Together AI (GPU Clusters) — Together's Instant GPU Clusters (8-64 GPUs self-serve) run on the NVIDIA Cloud Partner reference architecture with non-blocking Quantum-2 InfiniBand and NVLink. (source)
• Together AI (GPU Clusters) — Together's reserved GPU Clusters page claims scale from 8 to 4,000+ GPUs over InfiniBand (up to 256 GPUs on-demand; 512 to 1,000+ for GB200 NVL72), without stating fabric generation or topology on this page. (source)

This week's read

The number everyone will trade against next month is $2.53 — Silicon Data's H100 rental print, the index underneath the CME contracts listing October 5. Here is what that number is hiding: Silicon Data's own hyperscaler-tier H100 index sits near $7.48, and in this week's survey the same chip clears at $1.97 on the open marketplace, $3.20–$3.99 at neoclouds with a stated InfiniBand fabric, and $6.88–$12.29 at hyperscalers. That is not one commodity with a wide spread; it is at least three different goods sharing a ticker. The difference between them is not brand — it is whether anyone will state, in writing, what the network looks like.

Three details from this week's data that make the point:

1. Training-class capacity does not trade on the open market. Our live sweep of the largest GPU marketplace found zero 8-GPU H100-SXM hosts — the largest contiguous H100-SXM listing was 4 GPUs, B200 the same. Anything that can actually run a distributed training job is locked in reserved contracts, which is precisely the capacity a cash-settled future cannot reach.

2. Hyperscaler list prices are administrative, not market-clearing. Oracle prices H100 and H200 identically at $10.00/GPU-hr. Azure's H200 lists below its H100 ($10.60 vs $12.29). Google publishes no on-demand price for H200 or B200 at all — those chips are reservation-only. A hedger settling against a hyperscaler-tier reading is hedging a posted price, not a market.

3. 40% of surveyed capacity can't verifiably do the job its price implies. Of 70 offerings across 16 providers, only 10 meet the bar for large-scale training on published specs (Grade A), and 6 publish nothing about interconnect at all. The sellers who publish the most fabric detail — Azure, Oracle, Nebius, CoreWeave — are also the ones whose capacity you could actually write a delivery contract against. That correlation is the entire case for a deliverable spec.

One more thing worth watching: the physical layer is starting to move. Architect and Compute Desk announced ComputeConnect in July — an exchange-for-physical mechanism for converting compute futures into real capacity — and SF Compute is running a live spot/forward cluster market. An EFP network without a standardized definition of the deliverable is a settlement dispute waiting to happen. Somebody has to write the spec. That's what this report is.

Method

Prices and specs collected 2026-09-10 from public rate cards and official docs only; every row carries its source. Fields a provider does not publish are recorded as unknown, never estimated. Deliverability grades follow the Deliverable Compute Registry spec v0.1. Corrections welcome — the registry improves by being wrong in public.