Methodology · version 0.9 · 23 Oct 2026

How every figure is made

Each number we publish is computed once a day from stored, dated inputs and kept in one table, so the site, the share images, the free post and the briefs show the same value with the same date. When a method or a figure changes, the change is logged on the corrections page.

Compute price board: on-demand $/GPU-hour by provider class

INPUTS

Rate cards and marketplace listings collected every morning and stored with their capture date (in today's samples: AWS, Microsoft Azure, CoreWeave, Crusoe, DigitalOcean, Google Cloud, Jarvislabs, Lambda, Nebius, Oracle Cloud Infrastructure, RunPod, Together AI, Vast.ai, Verda). GPU families: H100, H200, B200, B300, A100, GB200, MI300X; variants (SXM, PCIe, NVL, HGX) count in their family and are recorded with each listing, as is the region.

CLASSES

Neocloud: specialist GPU clouds' on-demand prices. Public rate-card average (list): a census of the 4 hyperscalers (AWS, Microsoft Azure, Google Cloud, Oracle Cloud Infrastructure), list prices before discounts. Marketplace and spot: marketplace listings, plus spot and interruptible prices from any provider. All on-demand providers: every on-demand price of the SKU, whatever the class.

NORMALISATION

Every price is per GPU-hour: an instance price is divided by its GPU count (an 8-GPU instance by eight). On-demand terms only, except the spot and interruptible prices in the marketplace class; reserved and committed prices, and marketplace medians, are left out.

EXCLUSIONS

Fractional GPUs (MIG slices); listings the source flags as unavailable; listings not observed in the last 7 days, and stale listings not re-observed for 30 days or more; and any price more than 50% below its SKU-class median on 2 consecutive days, which is held for manual review before it can count again.

FORMULA

One price per provider and cell (the provider's lowest qualifying listing), then the 20% trimmed mean: 20% of the providers cut from each end, the rest averaged. Every cell shows its sample count and as-of date.

PUBLICATION RULE

A cell with fewer than 5 providers is not published and reads “n/a (sample too small)”. The hyperscaler census is published when all 4 hyperscalers report. The hyperscaler premium is the hyperscaler cell over the neocloud cell, less one, shown only when both are published.

HOW WE COMPARE

Other GPU price benchmarks are built differently, often from transactions or marketplace offers, and can sit well below list prices. We link to them rather than republish them: Silicon Data · Ornn. Internally, each morning our neocloud H100 is compared with the outside benchmarks whose terms allow it (today the GetDeploying GPU price dataset, CC BY 4.0); a gap of more than 30% is flagged for review and logged on the corrections page. Benchmark values are not shown.

CHANGES

The week-on-week change compares the all-provider cell with its print seven or more days earlier. It is shown only when both prints have the same providers: when a provider joins or leaves, the average moves without any price moving, so the change reads n/a.

WEEKLY HISTORY

The weekly price history on the dashboards is the GetDeploying GPU price dataset (getdeploying.com/dataset/gpu-prices), CC BY 4.0 (CC BY 4.0): weekly median of provider median advertised on-demand prices, $ per gpu-hour. It is published as its owners publish it and is not part of the class cells above.

CADENCE AND VERSION

Daily, every day of the week, after the morning capture. Version 0.9 from 23 Oct 2026; it replaces the Public Rate Card Composite (version 1.1, a median of each class's list prices).

Hyperscaler capex

DEFINITION

One definition everywhere: cash capex (purchases of property and equipment, cash-flow statement), summed for the 4 largest hyperscalers. Assets acquired under finance leases are not added.

INPUTS

SEC EDGAR XBRL facts from each company's 10-Q and 10-K: Microsoft, “Additions to property and equipment (cash; excludes assets acquired under finance leases)”; Alphabet, “Purchases of property and equipment (cash)”; Amazon, “Purchases of property and equipment (cash capex)”; Meta Platforms, “Purchases of property and equipment (cash; excludes principal payments on finance leases)”.

FORMULA

The latest calendar quarter in which all 4 companies have reported, and its change on the same quarter a year earlier. Cash-flow statements report year-to-date totals; a single quarter is derived as the difference of consecutive totals and marked as derived. Fiscal quarters map to the calendar quarter in which they end.

CADENCE AND VERSION

Checked daily; the figure moves when a new filing arrives. Version 0.9.

Equity basket

DEFINITION

Compute and datacentre basket: every covered company in the coverage universe (
47Compute Terminal basket statistics from exchange closesAs of 30 Oct 2026Source
names on the latest session), thematic ETFs and names awaiting their first trading day excluded. Narrower groups, such as hyperscalers or neoclouds, are labelled sub-baskets.

FORMULA

The equal-weighted average of the names' returns over the day and year to date, and breadth: how many names rose and fell. A name with a stale close is left out of that session.

INPUTS, CADENCE AND VERSION

Exchange closes for each listing on its own trading calendar, after every US session. Version 0.9.

Compute Terminal Index (CTI): Compute Cost and AI-Factory Capacity

CTI COMPUTE COST INDEX

  • For H100, H200, B200 and B300: the 20% trimmed mean of neocloud on-demand list prices on our price board, each day's change measured over the providers present on both days, so a provider joining or leaving does not move the index. With fewer than 5 providers on both days, the last value is held for up to 5 business days.
  • Composite weights 40% H100, 20% H200, 25% B200 and 15% B300, rescaled over the GPUs with enough providers and chain-linked when a GPU joins. 5 Jan 2026 = 100.
  • History before our own rate cards is back-calculated from the GetDeploying dataset's weekly on-demand provider median as published, including weeks when its sample of offers changed, and shown as such.
  • A daily move beyond 5% is reviewed before it is published.

CTI AI-FACTORY CAPACITY INDEX

  • Operational capacity: the IT power energised at the AI datacentres in the Epoch AI Frontier Data Centers register, each Monday.
  • Pipeline: capacity not yet energised, weighted by stage (0.25 announced, 0.5 approved or financed, 0.8 under construction).
  • 5 Jan 2026 = 100. A site moving by more than 200 MW in a week is reviewed before publication; Epoch's revisions are reconciled weekly.
  • Our own campus register adds campuses of 500 MW or more that Epoch does not track. They are proposed automatically from company filings and news reports, and each counts only after an analyst has checked it against its source.
  • The neocloud contracted-to-active ratio reads the active and contracted power companies state in their SEC filings, with the sentence each figure came from.

REPRODUCIBILITY AND PUBLICATION

Both indices are computed only from stored inputs and can be recomputed from them to four decimals. Values before the first publication on 11 Nov 2026 are back-calculated and marked so; a published value is never rewritten, and any restatement is logged on the corrections page. GetDeploying GPU price dataset, CC BY 4.0. Contains data from Epoch AI, CC BY 4.0.

Deals count and closes

DEALS AND LISTINGS

Counted over the last 30 days from the analyst-approved deals register. USD values only where the value is in USD or the source reported a USD conversion; each deal counted once; talks, raise targets and market capitalisations excluded.

SHARE PRICES

NVIDIA and the covered names: exchange closes of the latest session, with the date shown.

NEWS

Counted over the same last 30 days to the day's run from the analyst-approved news register: one summary and one source link per item; single-source items are flagged.

The instruments

An instrument shows a headline reading only once it has a computed value with its date; until then its card stays hidden. Full methods are on each instrument's page.

01 · Neocloud & AI-Infra Scorecard

INPUTS

Filings, transcripts, contract ledger, financing tape, exchange prices

METHOD

Eight factors, each with published sub-metrics and weights held in configuration: (1) power — energised and contracted MW, time-to-power; (2) contracts — take-or-pay share, tenor, prepayment, counterparty grade, $m per MW-year; (3) balance sheet — leverage, liquidity, contract cover of debt; (4) execution — delivery against promised MW and dates; (5) silicon access — allocation, generation mix, NVIDIA and AMD relationships; (6) diversification — customer concentration; (7) geography — power cost, grid queue and policy exposure by site; (8) cost of capital — marginal funding cost from the Financing Tape.

Operators and REITs use an operator pillar set: leased versus powered capacity, pre-leasing, development yield, cost per MW, power secured and leverage.

A factor with missing inputs is reported as n/a and excluded, with the gap shown; ranks and the change on the month are recomputed on every run.

EXCLUSIONS AND CAVEATS

Scores describe disclosed operating and balance-sheet positions, not expected returns.

CADENCE, STATUS AND VERSION

Daily. Not yet publishing a headline reading. Version 0.9. Full method →

02 · AI-Infra Credit Stress Index

INPUTS

Financing tape of loans, bonds, converts and joint ventures; bond prices; backlog disclosures

METHOD

Funding-cost ladder: the latest all-in cost by who stands behind the cash flows (hyperscaler-anchored, investment-grade lab or enterprise, unrated lab, merchant) and by instrument, with the spread between rungs.

Debt-maturity wall: principal by maturity year and issuer, net of scheduled amortisation, rebuilt bottom-up from each issuer's latest 10-Q debt note.

Contract cover: contracted backlog divided by debt principal, by issuer.

Index level, components and three-month range are published together.

EXCLUSIONS AND CAVEATS

No aggregate maturity wall is published until every issuer's note has been rebuilt bottom-up.

CADENCE, STATUS AND VERSION

Weekly. Not yet publishing a headline reading. Version 0.9. Full method →

03 · Compute Price & Basis Heatmap

INPUTS

Licensed rental indices, public rate cards, disclosed contracts, lease comps

METHOD

Lease layer from the Lease Comps register; contract layer from the Contract Ledger ($m per MW-year); on-demand layer from our price board (trimmed means by provider class, methodology 0.9).

Conversion between layers uses configured GPUs per critical-IT MW: H100 ~700, B200/B300 ~520, GB300 ~550, Rubin ~330.

The term-discount curve compares 6-month, 1-year and 3-year commitments with on-demand, from public disclosures only.

EXCLUSIONS AND CAVEATS

Licensed index values appear only in rendered publications until their licences allow redistribution.

CADENCE, STATUS AND VERSION

Daily. Publishing: H100 on-demand, all providers (trimmed mean), $/GPU-hour,
$5.14Compute Terminal price board (public rate cards and marketplaces)As of 1 Nov 2026Source
as of 1 Nov 2026. Version 0.9. Full method →

04 · Catalyst Probability Matrix

INPUTS

Catalyst calendar, analyst assessments

METHOD

Every catalyst row carries a source link; probability and impact edits carry a dated note in the revision log.

EXCLUSIONS AND CAVEATS

Probabilities are analyst judgements, published as such.

CADENCE, STATUS AND VERSION

Weekly. Publishing: events in the next 30 days,
2Compute Terminal Catalyst CalendarAs of 1 Nov 2026Source
as of 1 Nov 2026. Version 0.9. Full method →

05 · Options Positioning

INPUTS

Exchange options snapshots

METHOD

Snapshots for the thirty options names are taken after the close and stored with their capture time: implied and realised volatility and option volume, near-term open interest (three nearest expiries) and IV rank.

The implied move is the at-the-money straddle over the close; the daily options read states the strongest readings as facts.

EXCLUSIONS AND CAVEATS

A snapshot taken outside the overnight window is not used for the session.

CADENCE, STATUS AND VERSION

Daily. Not yet publishing a headline reading. Version 0.9. Full method →

06 · GPU Economic Life Tracker

INPUTS

Stored daily rental history, company depreciation policies

METHOD

Implied life and residual value come from stored rental history, compared with each company's stated useful life.

EXCLUSIONS AND CAVEATS

Short price histories widen the range, which is shown.

CADENCE, STATUS AND VERSION

Monthly. Not yet publishing a headline reading. Version 0.9. Full method →

07 · Capex & Circularity Map

INPUTS

Guidance, lease commitments, supplier disclosures

METHOD

Capex guidance is stored as ranges, with each company's fiscal-year basis stated; supplier-financed demand (for example capacity buy-back commitments) is shown separately.

EXCLUSIONS AND CAVEATS

Companies define capex differently; each definition is stated beside its figure.

CADENCE, STATUS AND VERSION

Monthly. Not yet publishing a headline reading. Version 0.9. Full method →

08 · Global AI Datacentre Pipeline Tracker

INPUTS

Permits, filings, press releases, open site datasets

METHOD

Statuses: announced, approved, financed, under construction, energised; each with a configured probability.

Broker totals with different scopes are never added together.

EXCLUSIONS AND CAVEATS

Evidence grades are shown per row; single-source rows are flagged.

CADENCE, STATUS AND VERSION

Weekly. Not yet publishing a headline reading. Version 0.9. Full method →

09 · Regional Power Readiness Index

INPUTS

ISO queues, utility filings, lead-time series, regulation map

METHOD

Markets: Northern Virginia, Texas, US Midwest, US Southwest, the Nordics, Netherlands/Germany, UK, France/Spain, the Gulf, India, Johor/Singapore, Japan/Korea and Australia.

Sub-scores: queue depth, firm power and wholesale price, equipment lead times, land and permitting, policy (moratoria, large-load tariffs). Weights live in configuration.

EXCLUSIONS AND CAVEATS

Markets with missing sub-scores are flagged rather than filled.

CADENCE, STATUS AND VERSION

Monthly. Not yet publishing a headline reading. Version 0.9. Full method →

10 · Compute Markets Monitor

INPUTS

Index levels, exchange notices, Federal Register and CFTC filings, venue data

METHOD

Each venue milestone is a dated event with a status: announced, pending regulatory review, approved, listed, trading or withdrawn.

Contract specifications are stored as data (CME: 730 GPU-hours per contract, $0.01 per GPU-hour tick, 36 monthly listings, cash-settled; ICE: up to 24 monthly contracts, $0.001 tick).

No venue is described as launched or listed until it confirms first trade.

EXCLUSIONS AND CAVEATS

Prices, volume and open interest are stored only once a venue trades.

CADENCE, STATUS AND VERSION

Weekly. Publishing: venues trading compute futures,
0Compute Terminal venue-status tableAs of 1 Nov 2026Source
as of 1 Nov 2026. Version 0.9. Full method →

11 · Build Cost & Lead-Time Index

INPUTS

Broker and engineering cost studies, OEM and utility disclosures, contract-award register, labour statistics

METHOD

Cost studies are stored with their scope (market, design class, what is included) and never blended across scopes.

EXCLUSIONS AND CAVEATS

Cost studies differ in scope; each figure carries its definition.

CADENCE, STATUS AND VERSION

Monthly. Not yet publishing a headline reading. Version 0.9. Full method →

Independence and data contributors

Our founder is also a senior executive at one of the leading global AI infrastructure companies. Compute Terminal is editorially independent of that company: it has no role in index methodology, data selection or publication. That company contributes no data to the price board: its prices are not among the rate cards we collect. Impersonal, general research for professional market participants, including professional and sophisticated investors where receipt of such research is lawful. Not investment, procurement, engineering, legal or tax advice, and not a recommendation to buy or sell any security.

Versions

Version 0.9, 23 Oct 2026: price board by provider class (trimmed means, sample counts, provider minimum, outlier review), one capex definition, one basket definition, instruments shown only with a reading. Every change to a method or a published figure is on the corrections page.

Get the daily AI compute prices free in your inbox every morning, and the CTI Compute Cost Index from its launch on 11 Nov 2026.

By continuing you agree to the Terms of Subscription and the Terms of Use.