Market benchmarks / compute economics / market structure

The price of compute, tracked.

GPU rental markets, market-implied forward curves, chip economics, token prices, and the gigawatt buildout — one independent desk for the machines that make intelligence.

Market tape
H100 ON-DEMAND FLOOR$1.99/hrdecreased·B200 ON-DEMAND MEDIAN$6.92/hrincreased·H200 ON-DEMAND MEDIAN$4.39/hrincreased·B300 ON-DEMANDfrom $7.39/hr·A100from $1.35/hrincreased·MI355X ON-DEMANDno primary listing·NVDA DC REV$75.2Bincreased·BIG-4 CAPEX ’26E$710Bincreased·GPT-4-CLASS TOKENS$0.18/1Mdecreased·FRONTIER PIPELINE~32 GW·

01Forward curves

What the market thinks compute will cost

Instrument-specific near anchors, monthly-average tenors through Jul ’27, and the shape of implied supply pressure ahead.

As of JUL 31 2026 · 20:06 UTCHow curves are derived

● Market-implied term structure

Implied forward curve

First point is a weekly exact-time anchor (RTX 5090: front-month monthly average); remaining points are monthly-average contracts — shown in $/GPU·hr or % vs anchor

Kalshi event ladders referenced to Ornn. Raw implied median = the 50% bid/ask-mid crossing; monthly back points use two [1,2,1] smoothing passes from the third month onward. Hollow points are thin or interpolated.

How to read this curve

The shape is the story. H200 trades backwardated; A100 sits in contango; RTX 5090 eases lower out the strip; B200 and H100 are flat, priced to hold through Jul ’27.

Quality-passed

Solid points are implied medians whose underlying contract ladders pass the book-quality gates.

Inferred

Hollow points use thin or interpolated ladders. Every forward value is market-implied, smoothed where disclosed, and non-executable.

Curve time machine

Replay the archived market-implied term structure.

Computed signals

The market in four numbers

Derived from the current datasets on every build — never hand-entered.

B200 on-demand specialist median
$6.92/GPU·hr
+20% y/yBlackwell demand outran supply
NVIDIA data-center revenue · FY27 Q1
$75.2B
+92% y/yNext-quarter guide $91B total
Tracked AI capex · 2026E
$780B
+66% vs ’25MSFT · GOOG · META · AMZN · ORCL
GPT-4-class token floor
$0.18/1M
−99.5% since Mar ’23DeepSeek V4 Flash · blended 3:1

Today’s read · market snapshot JUL 31 2026 · 20:06 UTC

Compute is becoming a commodity — the fourth factor of production — and for the first time it trades like one.

RTX 5090 repriced +7% since Jul 17. Its curve is easing: the strip drifts lower into the back months. H100 is the anchor of the board: ~$2.68 through Jul ’27 on a flat curve.

01
RTX 5090 sets the tone.

The monthly-average anchor is $0.60; the Jul ’27 tenor implies $0.55.

02
H100 rental prices are still easing.

The marketplace floor moved −3.9% over the latest three-quarter window to $1.73/GPU·hr.

03
Intelligence keeps repricing faster than hardware.

The GPT-4-class blended floor is now about 210× cheaper than launch at $0.18 per million tokens.

04
Listed compute futures are slated for Oct 5, 2026.

CME Group is set to list Silicon Data H100 and B200 rental-index futures on NYMEX, pending CFTC review. Venue monitor

02GPU cloud

Renting an H100, 2023 → today

US$ per GPU-hour. From $12 at peak scarcity to a stable ~$4; primary on-demand now starts at $1.99, with separately labeled marketplace access near $1.73.

As of 2026-07-06All GPUs & providers

H100 SXM · provider and marketplace price

Quarterly midpoints by provider class

Price-history chart awaiting interactive loadStatic chart frame · exact values in table view

Provider list prices, launch announcements, and archived snapshots; legacy quarterly midpoints ±15%. Current 2026-07-06 quote-table aggregates are provisional unless an evidence record pins an exact archived Git object.

03The macro picture

Compute demand is compounding

The frontier training-compute trend runs at 4.4× a year; tracked 2026E capex is +66% versus 2025.

Training compute of notable models

Total FLOP, log scale — estimates marked

Training-compute chart awaiting interactive loadStatic chart frame · exact values in table view

Epoch AI estimates and lab disclosures; estimated points carry wide error bars.

Big-tech capital expenditure

US$B per calendar year (2026 = guidance)

Capital-expenditure chart awaiting interactive loadStatic chart frame · exact values in table view

Company earnings releases and SEC filings; Microsoft includes finance leases; Oracle fiscal year mapped.

04Inference

Intelligence, repriced monthly

GPT-4-level capability became roughly 210× cheaper since launch. Then 2026 flagships reintroduced a frontier premium.

Cheapest GPT-4-class · blended

$0.18/1M

−99.5% since Mar ’23 · DeepSeek V4 Flash

OpenAI flagship · GPT-5.5

$11.25/1M

+227% vs GPT-5 · Flagship list price rose in Apr 2026

Batch & cached reality

−50 / −90%

Batch APIs ~50% off list; cached input ~90% off — production pays below list

05Quick answers

The methodology, without the footnotes

The three definitions readers need before comparing the curves, provider classes, or token-price series.

How are the GPU forward curves derived?

From Kalshi event-contract ladders on GPU rental prices referenced to Ornn's hourly index. The first point is a future exact-time weekly implied median for B200, H200, H100, and A100; RTX 5090 uses its front-month monthly-average implied median. Each raw implied level is the ladder's 50% crossing. The monthly strip receives two [1,2,1] smoothing passes from the third monthly tenor onward. Thin and interpolated points are labeled; none are current spot or executable quotes.

What does the blended token price mean?

Blended price = (3 × input + output) ÷ 4 per million tokens, reflecting a typical 3:1 read/write mix at standard list rates.

What are the provider classes?

Under primary on-demand policy primary-on-demand-v1, hyperscaler is the undiscounted big-cloud provider-list floor and specialist is the AI-native provider-list median. Marketplace access is the separately labeled open-market floor on community hardware.

06Go deeper

Explore the full desk