Changelog

Company coverage splits into three surfaces. The Companies index gains a band switch that separates the AI-native application layer from the twenty-nine established software companies selling AI, which were already in the index but had no control that could surface them; each band now carries its own headline figure on its own stated basis, and only the AI-native band feeds the index total. Two new pages join it: Model providers, a profile per frontier lab covering run-rate and collected revenue, operating loss, funding chain and compute commitments, with a roster of the labs that publish nothing; and Compute providers, which reports gross disclosed revenue and revenue net of the apps carve-out as two separate figures rather than one. Both new pages show how their totals reconcile to the Revenue Ledger. Four model providers that carried a revenue figure with no dated source behind it were researched; three now carry one, and two of those figures turned out to be five months stale.

The Revenue Ledger headline moves onto the buyer-side gross basis for 2026: $120B of customer-paid revenue, replacing the $97B run-rate reading that had led the page since 10/08/2026. The baseline tile and the forecast band move to the same figure, so every revenue headline on the page reads one number. The narrower run-rate stays on the page as a labelled reference. Because the basis changed, the growth multiple shown beside it resets from 7.7× to 1.7×. The earlier figure measured two years of growth from a smaller, differently-counted base.

Capital moves to a 2026 frame: $1.2T of cumulative capex committed through 2026, against $151.5B of cumulative customer revenue. The capital-to-revenue ratio is rebuilt like-for-like: $766B of 2023–25 capex against $31.5B collected over those same years, or $24 of capex for every $1 of revenue. It previously compared a capital stock that stopped at December 2025 against a revenue rate read in August 2026, which is neither cumulative nor current and flattered the build-out because the numerator had stopped counting. The published cumulative capex figure moves from $745B to $766B.

Depreciation moves off a flat four-year assumption and onto the schedules the four largest buyers disclose: Amazon 5.0 years (shortened from 6.0 in February 2025, citing the pace of AI), Alphabet 6.0, Microsoft 6.0, Meta 5.5, capex-weighted 5.6 years. The flat 4.0 had overstated annual depreciation by about 40%, while the page's own footnote already said two of the four had moved to six years. The 2030 cost of owning the infrastructure falls from $747B to $535B.

The forward revenue line in the capex-and-depreciation chart is rebased. It had carried a hand-typed series putting 2026 revenue at $64B, against the $120B the site now publishes. Corrected alongside the depreciation schedule, revenue and the annual cost of ownership now cross around 2026 to 2028 rather than never. This reverses the page's previous reading, which put convergence at about 2030.

Compute's headline moves to a current run-rate basis: $79B, annualised from the most recent reported quarter. The share chart, the four-quarter trend and the trajectory line all now run to Q2 2026; they had been two quarters stale against the underlying disclosures, and the share chart was also carrying a live understatement because corrections made on 31/07/2026 had reached the quarterly series but not the per-component block the chart read. AWS moves from 23.5% to 31.5% of compute revenue in a single quarter. The Copilot scope-out moves from a blanket ratio to a per-provider deduction. On this basis application revenue passes compute revenue: the compute-to-applications multiple crosses below 1× to about 0.8×, against 1.8× before.

The forward-looking band on the Revenue Ledger is replaced. It now reports investor money burned for every $1 of customer revenue: $1.50 in 2025, $0.42 in 2026, $0.26 in 2027. The pool of losses itself rises over that period, from $35B to $55B; the ratio falls because revenue grows faster than the losses do.

The reconciliation strip on the homepage is rebuilt on 2026. Between 2025 and 2026 every layer grows (capital, usage, compute and power all rise) while every multiple against revenue falls, because revenue grows faster than any of them. A strip encoding only the multiples would have shown five shrinking bars while every layer was in fact growing. Each row now carries its 2026 figure, the growth on 2025, and the 2025 reading it replaces. The five homepage cards and the page's search description move onto the same basis.

The customer-paid revenue baseline that every multiple on the site is measured against moves from $17.36B to $24B for 2025, onto a gross-of-channel-margin basis. This entry is a backfill. The change went live with the site redesign on 28/06/2026, in a release whose scope explicitly set the Changelog aside, which is how it went unrecorded.

A dedicated Companies index is added for the AI application layer: 58 companies with revenue observation chains (every figure dated, tiered and linked to its source), funding histories, pricing models, regional presence backed by cited evidence, and a locked category taxonomy replacing the previous free-text labels. Revenue run-rate figures are classified on a four-rung ladder (disclosed, reported, estimate, not published) and estimates never enter headline totals. Growth is computed only where two credible observations sit at least ninety days apart; nothing is annualised from a single point. The page publishes its own data health, including the median age of its observations.

The homepage is refreshed to reflect the five-ledger architecture. The masthead reads "The AI economy, layer by layer." with a tagline that names every ledger in causal order ("Capital in. Compute spent. Power drawn. Tokens out. Revenue back. Five ledgers, one system."), replacing the prior three-ledger framing.

A new hero visual — The AI “$” Stack — replaces the previous tile strip and the four-step narrative-flow loop. Five horizontal pill bars descend in size by 2025 figure: Capex $330B → Usage $164B (notional at OpenRouter median output rate) → Compute $43B → Power $25B → Apps Revenue $17B. Each bar links through to its ledger; Power carries an explicit Tier 3 pill until its own data hardens.

The Apps Revenue figure published on the homepage is restated to the 2025 cohort sum-of-quarterlies basis ($17B), aligning with the Revenue Ledger and the Compute Ledger's Layer Stack. The hook one-liner is now Compute-anchored: $2.50 of compute spend stands behind every $1 of customer-paid AI Apps Revenue. Several editorial sentences (homepage card commentary, hero hook sub-line, sub-headers on the Compute Ledger page) were trimmed at the same time so ledger surfaces stay methodology-first.

The Bear / Base / Bull scenario toggle is retired from the homepage.

The Compute Ledger goes live as the fourth Ledger surface, answering "what hyperscalers and neoclouds earned from AI compute in 2025." Built around a three-segment model: frontier-lab compute — what model labs (OpenAI, Anthropic, Mistral and others) paid external compute providers — at ~79% of the dollar; AI workload compute — the rest of business AI demand running on hyperscaler and neocloud infrastructure; and hosted model APIs — Bedrock, Vertex, Foundry and the equivalents.

The methodology iterated through four refinements before ship: (1) bucket decomposition with a Copilot scope-out and a pass-through correction to avoid double-counting flow that hyperscalers recognise from labs they themselves fund; (2) segment sizing correction following Andy Jassy's Q1 2026 commentary, taking AMZN AI compute from a top-down $18B to a bottom-up $10B and GCP from $18B to $7B; (3) a quarterly trajectory back-cast that uses Q4 2025 ≈ Q1 2026 run-rate where issuers haven't yet published; (4) a switch to the sum-of-quarterlies basis for the 2025 calendar so the year doesn't drop on Q4 → Q1 methodology changes.

Published 2025 total: $43.07B. The page also introduces the Layer Stack — a four-row proportional bar showing Apps Revenue ↔ Compute ↔ Silicon ↔ Power at canonical 2025 dollars.

The Revenue Ledger's published headline figure is now the 2025 cohort sum-of-quarterlies total ($17.36B). This is the same cohort that flows through to the homepage Apps Revenue card, The AI “$” Stack hero, and the Compute Ledger's Layer Stack, so the figure quoted on every public surface lines up.

The cumulative-since-2023 framing ($22B) that previously appeared in the homepage hero and meta tags is retired; the Revenue Sankey continues to publish per-provider quarterly flows on the same cohort scope (rebalanced 3 May, see entry below). The change is a framing alignment. Provider-level data and the Sankey methodology are unchanged.

The Follow the Dollar Sankey now derives every published value from our underlying entity database, applying the same documented methodology to every provider rather than hand-curating each one. Previously, per-provider customer revenue and investor subsidy were calculated separately and the chart was rebalanced manually whenever an input changed.

The methodological shift: investor subsidy now functions as a balancing figure. When customer revenue is revised (for example, when stronger data arrives for a provider's reported revenue) the chart holds total provider value steady (anchored to documented inference costs and operating losses) and adjusts the investor subsidy to keep the Sankey balanced. Future revenue revisions will visibly shift the customer-versus-investor split for each provider without changing the provider's overall flow size.

Specific changes published in this update:

OpenAI: Customer revenue revised $7.65B → $9.31B (+22%). Provider total holds at $13.65B; investor subsidy correspondingly adjusts $5.20B → $4.34B. The methodology change in action.

Google (Gemini): Customer revenue $2.00B → $2.25B; provider total $2.50B → $2.93B (+17%). Lower data depth than OpenAI. No published operating loss is available yet, so methodology-derived estimates apply throughout. Worth flagging in case more granular Google AI revenue data lands.

Anthropic: Provider total $7.71B → $7.30B (−5.3%). Small methodology-driven adjustment.

Five smaller providers (DeepSeek, Mistral, xAI, Minimax, Moonshot) now appear collapsed into a single "Other Model Providers" node at $1.33B combined. Each is preserved individually in the provenance trail; the aggregation rule keeps the chart readable as more providers enter the registry.

For full methodology and per-provider source classifications, see methodology.

An entity directory now lists all 103 tracked entities with their current snapshot, best-source confidence, and a "qualifies for detail page" flag based on a data threshold (≥3 fields, ≥medium confidence, ≥1 provenance entry). 57 of 103 qualify at the default threshold, and a detail page is published for each entity that crosses the threshold.

The Ledger replaces its single token-centric dashboard with three ledgers (Capital, Revenue and Usage), each answering a different question about the same underlying economy. Numbers now carry explicit provenance tiers (1A–3C) and can be viewed under Bear, Base or Bull assumptions. A fourth ledger, tracking physical power and grid constraints, follows later.

CapEx revised to $745B following Platformonomics cross-check. Previously $770B. The $25B delta came from double-counting non-AI networking spend in hyperscaler totals.

Provenance tier system (1A through 4) applied to all Sankey nodes. Every flow and stock in the Capital Ledger now carries an explicit data-quality tag.

Revenue Sankey fully rebalanced from consensus sources. All flows verified against entity-level filings. Total system revenue: $26.86B quarterly.

Entity-level revenue figures refreshed from earnings calls and analyst commentary across 32 source records. Higher-tier sources (10-K filings, IR disclosures) take precedence over commentary and podcast estimates.