← pland.world
Compute Capital Markets · Field Notes

Turning GPU hours into a tradable commodity

By PlanD/ Tabula/ Snapshot 2026-07-05/ Origin: a thread by @0xfishylosopher (Jay Yu)

Compute behaves less like corn or gold and more like electricity — temporal, non-fungible, priced by the hour. A thread argued that before compute futures could exist, someone had to build a credible price index. This is a field study of that market as it actually stands in mid-2026: three index providers racing to become the benchmark, live GPU-rental prices, a working spot marketplace, and a running collection of the questions that came up along the way.

Live Silicon Data readings · captured 2026-07-05

H100 · Neocloud
$2.71/hr
▼ 2.9% · 30d
SDH100RT · the working benchmark
H100 · Hyperscaler
$7.24/hr
▼ 3.1% · 30d
~2.7× the neocloud price
B200 · Neocloud
$5.73/hr
▲ 5.3% · 30d
new gen rising as H100 falls
A100 · Neocloud
$1.64/hr
▲ 1.9% · 30d
older-gen floor, quietly firming
01

The race to become the benchmark

three index providers · racing for venues
The coarsest layer is the chip model (H100 vs H200 vs B200). But the thread's real point: a tradable SKU is far finer than "chip model" — and it's these axes that make compute non-fungible and push most trades into OTC.
Chip model
H100 / H200 / B200 / B300 / RTX 5090
Form & interconnect
SXM vs PCIe; NVLink/NVSwitch or not; fabric InfiniBand vs Ethernet — "H100 SXM ×8 + InfiniBand" ≠ "H100 PCIe ×1"
Region / AZ
us-east-1 (N. Virginia) vs us-west-2 (Oregon) vs eu-west-1 (Ireland) vs ap-northeast-1 (Tokyo) vs ap-south-1 (Mumbai) — plus cheap-power frontiers (Nordics, the Gulf, West Texas, Iceland). Price tracks local power cost, grid mix and data-sovereignty rules; latency back to your data usually pins the choice.
Term
on-demand / spot / reserved (1–36 months)
Cluster size & topology
1 GPU → an 8-GPU box → a single NVL72 rack (72×GB200, one NVLink domain) → a 512- or 1,024-GPU fat-tree InfiniBand pod → a 10k+ GPU SuperPod. Non-blocking vs oversubscribed fabric, rail-optimized topology — the bigger and tighter the cluster, the higher the per-GPU-hour premium.
SLA / reliability
reliability tier, oversubscription, uptime guarantee

A cash-settled future needs one number everyone trusts at expiry. Whoever owns that number owns the market — the way S&P owns "the S&P 500" and licenses it to every exchange and ETF. Three players are racing to be that index; the two furthest along each pair with a venue to settle contracts — Silicon Data with CME, Ornn with ICE (plus Architect and Kalshi). Ornn's OCPI already prints on the Bloomberg Terminal.

DRW-linked · GPU indices · CME's benchmark
CME Group
first exchange-listed compute futures
CME futures · pending CFTC
a16z-led · $33M · OCPI index
ICE
+ Architect/AX perps · Kalshi
OCPI settling live · ICE pending
data + instruments + brokerage stack
OTC brokerage
compute pricing · hedging · financing
Earliest stage
Why the index is the choke point

No trusted, manipulation-resistant reference price → no settlement, easy manipulation, useless hedges. Liquidity can't form until the index exists and is believed. That's why all three obsess over transaction-based methodology — it's the exact lesson of LIBOR → SOFR: move off self-reported surveys onto real prints, or the whole edifice is gameable.

02

What the index actually shows

Silicon Data portal · 2026-07-05

The index deliberately abstracts away region, provider and rental term to produce one representative number per chip — the "generic basket." A specific machine (a region, an interconnect, a 3-month reserve) trades at index ± basis. Note the two-tier split: hyperscaler H100 sits near $7 (largely on-demand list price from AWS/GCP/Azure), while neocloud H100 sits near $2.70 (competitive spot across CoreWeave, Lambda, RunPod, Vast.ai…). That ~2.7× gap is a basis — and the reason neoclouds exist.

Same chip, different world: on-demand rental $/hr by tier. The H100 hyperscaler-vs-neocloud gap is the headline basis.
H100 forward curve (neocloud), traced from the portal's 36-month view. Downward slope = backwardation: the market prices longer commitments cheaper — it expects H100 to keep getting cheaper. Live on the Silicon Data portal ↗
A forward curve is one slice of a surface

Price is a function of two time axes — calendar date and contract tenor. The forward curve fixes "today" and sweeps tenor (one row). To see a single contract's history you fix the tenor and sweep the calendar (one column). Silicon Data's trial shows only today's row; the historical dates sit behind the paid tier.

The full year, straight from the feed. Pulled from Silicon Data's own live archive — one year of daily prints for every index (2025-07 → 2026-07). Two things the headline number hides: H100 neocloud actually rose ~28% off a $1.96 trough, and the newest silicon (B200) is by far the most volatile.

Every index, one year of daily prints (weekly-sampled). Newer chips (H200, B200) start partway through — that's when their index went live.
Both H100 tiers — hyperscaler (gold) and neocloud (teal) — with the basis shaded between them. As neocloud H100 rallied, the gap quietly compressed from ~$5.0 to ~$4.5.
IndexCurrent1Y Low1Y High1Y ΔVolatility
Volatility = annualized stdev of daily returns. Computed from the archived daily series. B200's 25% vol vs H100-hyper's 6% is the whole story: bleeding-edge supply is priced far more nervously than mature capacity.
03

The spot market, made concrete

Compute Exchange · 19 live pools

If the index is the abstraction, Compute Exchange (same CEO as Silicon Data, co-founded by DRW's Don Wilson — the index layer and the trading venue under one roof) is the ground truth: an auction marketplace where every listing is a fully-specified SKU — model × form-factor × memory × interconnect × NIC × DPU. It's the best picture of what "a unit of compute" really means. Two things jump out: the sheer generational spread (Ampere → Rubin), and that next-gen silicon is already sold as a forward — Rubin chips that don't exist yet are listed "Expected Jul 2026 · Submit Interest." That is the thread's "capacity forward" in the wild.

Dense FP4 throughput per GPU (PFLOPS), by marketplace pool. Blackwell → Blackwell Ultra → Rubin roughly doubles again — the generational leap that keeps the index a moving target.
HBM memory per GPU (GB). More/faster memory is why newer chips win at LLM inference regardless of raw FLOPS.

The catalog as a SKU taxonomy. A sample of the pools — note the gold card: a capacity forward on hardware that hasn't shipped. Browse Compute Exchange ↗

The thread, made real

"Reservations / capacity forwards are almost always bilateral OTC trades on particular SKUs." Here it is on-screen: Vera Rubin NVL — 288GB HBM, 22 TB/s, 35 PFLOPS FP4, NVLink 6 — offered for reservation months before the silicon exists. Whether these forwards ever standardize enough to hedge on an exchange is the open question of the whole space.

04

Will there be one "general compute price"?

a comment on the thread's vision

The thread imagined dealers hedging specific SKUs against a "generalized basket exchange (e.g. an H200 basket)." Read literally, the basket is per-chip — and that already exists (SDH100RT, OCPI-H100). In that sense the vision has shipped: CME↔Silicon Data and ICE↔Ornn are building futures on exactly those per-chip baskets. But the tempting bigger idea — a single, cross-chip general compute price — runs into three walls:

No fungible unit
An H100-hour and a B200-hour do wildly different work. "Compute" has no agreed physical measure the way a barrel of oil has energy content. You'd have to normalize by performance — which precision? which workload?
Obsoletes too fast
The mix turns over every ~18 months (Hopper→Blackwell→Rubin). A fixed basket needs constant re-weighting; its drift makes it a poor settlement anchor — you'd be hedging a moving target.
Basis explodes
The coarser the index, the larger its gap to any real SKU — the worse it hedges. Per-chip looks like the sweet spot the market has already voted for.
Comment

A single "general compute index" is likely to emerge as a macro barometer (a CPI-for-AI-compute) rather than the core hedging benchmark. And the truly universal unit may not live at the GPU layer at all — it may be $/token. Ornn's Token Price Index and Silicon Data's LLM Token Index already price the useful output directly, abstracting the hardware away entirely. Tokens are more fungible than GPU-hours — so if one number ever comes to mean "the price of compute," it might be measured in tokens, not hours.

05

Field notes — the questions

a running collection · click to expand

This project was built by asking, in order, whatever didn't make sense. Grouped by theme, here's the collection — the compute- and market-structure-specific ones worth keeping.

Field notes, not investment advice. Prices are point-in-time readings from third-party portals (Silicon Data, Compute Exchange) captured 2026-07-05 and move constantly; forward-curve and time-series charts trace on-screen shapes and are illustrative. Company details from public reporting. Direction-2 (the index layer) is the focus here; the "will the market form" and dealer/basis strands remain open.