At the current run-rate, application revenue has passed compute revenue.

The Compute Ledger tracks the supply-side money: hyperscalers, neoclouds, hosted model APIs and dedicated AI workload compute. It was the layer earning revenue first; at the current run-rate the applications it serves now earn more than it does, which is the order a working system settles into.

Q2 2026 annualised compute revenue vs current customer-paid run-rate

Compute : Revenue

~0.6×

~$79B of gross AI compute revenue, annualised from Q2 2026, against $138B of customer-paid AI revenue (run-rate, 31/08/2026).

Reference point, 2025 full year: 1.8×. $43B of compute revenue against the $24B collected by applications. Collected revenue and a run-rate are not the same measure; each is shown with what it is.

Annualised compute revenue

~$79B

Gross AI compute revenue, annualised from the most recent reported quarter: up +112% on the same quarter a year earlier, and above 2025’s $43B full-year total.

Buyer and supplier concentration

Concentrated

Most of it flows to or from a handful of frontier labs (led by OpenAI and Anthropic), billed by a small supplier set led by Microsoft, Google and AWS.

Forward commitments

Committed

Multi-year compute already contracted by the largest labs converts to revenue through 2030, not at signing.

Why this matters — compute is an input to application revenue, not a rival to it: every dollar spent on compute has to be recovered through the applications running on top of it, so a working system ends with compute the smaller of the two. What the ratio does not yet say is whether the gap is wide enough, or widening fast enough, to cover the capital behind it.

Who is earning the compute revenue

Modelled estimate

Frontier model labs paid an estimated ~$34B to compute providers in 2025, or ~79% of all AI compute revenue.

2025 lookback · post-Copilot · ecosystem-wide compute revenue · cite-able to 10-Qs

AI Ledger · Hepburn Advisory

Provider detail

Click any segment of the bar above for its filing detail.

Site rebuilt 02/07/2026 · that is the rebuild date, not a refresh of the figures below · provider-by-provider totals through 2026Q2; the three-segment breakdown in the panel below is still 2026Q1, which is the last quarter with a published decomposition · gross-disclosed, consistent with each issuer's revenue recognition.

What actually earns — Microsoft, Google and AWS take the large majority of every compute dollar today, much of it billed to the same frontier labs they also invest in. The neoclouds (CoreWeave, Nebius, Lambda and Crusoe) are growing fast but remain a small slice.

Quarterly Trajectory

Q1 2025 to Q2 2026, one line per major seller, on the sum-of-quarterlies basis. Quarters are anchored to each provider’s own disclosed run-rate divided by four. Microsoft $9.25B from $37B, AWS $3.75B from $15B in Q1, rising in Q2 on Andy Jassy’s stated floor of more than $25B. Two lines need reading with care. Microsoft is flat from Q1 to Q2 because it withdrew its AI run-rate disclosure at FY26 Q4 — the phrase appears nowhere in the 8-K or 10-K, so the line is held at its last disclosed level. Flat here is the absence of a measurement, not a measurement of no growth. CoreWeave, Nebius, Lambda and Crusoe are also held at Q1: they report after this cycle.

ai-index.hepburnadvisory.com.au

i All figures for the big three (Microsoft, Amazon and Google) plus Oracle are reported on a gross basis (consistent with each company's revenue-recognition policy). Quarterly numbers are pre-Copilot for trajectory continuity; the post-Copilot total is what flows to the headline. Pure-play neoclouds (CoreWeave, Nebius, Lambda, Crusoe) have no Copilot exposure and no token-API pass-through.

What to watch

Compute gets paid first; the open question is whether that holds. Three levers decide it. Each is shown against where it stands now.

Detail & references

Frontier lab computeper-provider attribution

How much frontier model labs (OpenAI, Anthropic, Mistral, others) paid external compute providers in 2025, and which provider each lab pays.

Forward commitmentsmulti-year contracted compute

Multi-year contractual commitments reported gross, separately from realised compute revenue. Per the Cross-Ledger Reconciliation rule, these dollars are not rolled into the trajectory chart above. They are forward obligations that flow through over years.

Anthropic and OpenAI have committed roughly $718B of multi-year compute to the largest hyperscalers, reported gross and separately from realised compute revenue. Detail loads below.

i Why this is separate from the trajectory chart. Google's remaining performance obligations (contracted revenue it has not yet delivered) jumped in the quarter. That is the order book, not Q1 2026 revenue. The Anthropic-Google $200B and Anthropic+OpenAI $718B are forward commitments to the same three hyperscalers. They convert into compute revenue over a five- to eight-year window, not at signing. Mixing them into a quarterly trajectory would double-count: the contracted dollars already drive the realised quarterly bars over time. Sourced from issuer disclosures of remaining performance obligations and announced multi-year compute commitments.

Convergence signalsthresholds, observed values & readings

The signals behind "What to watch", each with a defined threshold and observed value, plus the growth and concentration readings of the same data.

How compute fits the five-layer reconciliationlayer-by-layer bridge

The Compute : Revenue ratio at the top of this page is the single statement of compute against the customer-paid baseline. The old per-page "Layer-Stack Multiplier" is now the site-wide reconciliation spine, so the picks-and-shovels-versus-apps comparison is shown once. See the full five-layer reconciliation spine on the Overview.

Methodology caveat

Projections are directional arithmetic, not forecasts. Benchmarks are dated and sourced. Provider revenue defaults to the sum of the reported quarters, gross of pass-through; every assumption is editable. Where a number appears on a Ledger page, that page is the canonical source. Full method on the Methodology page.