Indices/AI & Compute

LIVE

KH100 — The H100 Index

What does an H100-hour actually rent for?

There is no single H100 price. The same GPU-hour clears from ~$1.8 on marketplaces to $12+ at big clouds — a 2–7× spread that any one-source number hides. KH100 prints the stratified composite of the three tiers the market actually clears at, the dispersion between them, and the committed-tenor curve — daily, from a self-archived book of ~600 offers, with Ornn's transacted index printed beside as the cross-check.

See it live The three prices

KH100 · Jul 2026 = 100

Big-cloud OD $/hr

–

Neocloud list $/hr

–

Marketplace $/hr

–

Tier dispersion

–

The problem

Every published "H100 price" is one seller tier. Big clouds quote $4–12, neoclouds ~$3, marketplaces clear under $2.6 — same silicon. A single number isn't a price, it's a sampling choice.

KOPS KH100

A fixed-weight geometric composite of all three tiers — 40% big-cloud · 35% neocloud · 25% marketplace — printed with the tiers, the dispersion ratio, and the transacted cross-check. The spread IS the market structure, so we publish it.

Valuable for

Builders pricing a training run (the dispersion is your negotiation headroom), lenders underwriting GPU fleets, and anyone tracking when compute finally commoditizes — the tiers converging is that moment.

Live index

One number, honestly assembled.

The composite since July 2026, with the GetDeploying weekly spine (tier-B-only, labeled) extending context back to January 2025. Base day 2026-07-03 = 100, constants frozen.

KH100 · stratified H100 rental compositedaily · offer book archived 13:05 UTC

Tier dispersion · max ÷ min of the three tier mediansthe commoditization dial

The three prices

Same silicon, three markets.

Today's tier medians, with the transacted cross-check and the inside spread of the big-cloud book.

Tier A · Big-cloud on-demand gpuhunt median — 8 clouds, – offers–

Tier B · Neocloud list GetDeploying provider median — ~40 providers–

Tier C · Marketplace clearing geomean of Vast.ai & Akash medians–

Cross-check · Ornn transacted ICE-futures reference index — never averaged in–

Inside spread big-cloud book p10 → p90 (within one tier)–

the gap between these rows is the neocloud business model, measured daily

Reading it. Marketplace (–) is what you can actually pay this afternoon; big-cloud (–) is what procurement quotes. Dispersion at –× means the exact same GPU-hour costs that multiple more depending on the door you walk through. When this ratio grinds toward 1×, compute has commoditized — the scoreboard for that is CIB.

Term structure

The committed-tenor curve.

Neocloud provider-median $/hr by commitment — spot to 36 months, from the archived GetDeploying weekly ladder. Almost nobody publishes this curve; it's where scarcity expectations live.

a downward-sloping curve prices in H100 depreciation; an inverted one prices in scarcity

Index methodology

Stratified, frozen, cross-checked.

The tiers

A — big-cloud on-demand median via gpuhunt (AWS/GCP/OCI/Lambda + 4 more, ~600 H100 offers/day). B — GetDeploying neocloud list median (~40 providers). C — marketplace clearing: geomean of Vast.ai asks (full explicit-query pool, verification-tiered) and Akash on-chain leases. The full per-offer book is archived daily — the index is reconstructable from raw offers.

The composite

Fixed-weight geometric mean of tier relatives, base day 2026-07-03 = 100 ($4.19 / $2.8725 / $1.9384 — frozen v0, never restated; changes version). Pre-period context is a tier-B-only weekly spine back to Jan 2025, rebased to join at 100 and labeled as such.

Governance

Per-leg carry-last-good ≤ 14 days, else the print suspends. Ornn transacted and CCIR reference rates are cross-checks, never inputs. Measurement upgrades (Vast full-pool query, 2026-08-30) are documented splices, not silent restatements. Full spec in the KOPS methodology annex.

KH100 = 100 × (A/4.19)0.40 × (B/2.8725)0.35 × (C/1.9384)0.25

three markets, one honest number — with the disagreement between them printed beside it.

KH100 — The H100 Index

One number, honestly assembled.

Same silicon, three markets.

The committed-tenor curve.

Stratified, frozen, cross-checked.

The tiers

The composite

Governance

The rest of the compute stack.

The H100's honest price tag.