The State of the AI Economy (Exponential View, June 2026)

"The State of the AI Economy" is a sixty-six-page report published on June 25, 2026 by Azeem Azhar, William Gildea, Hannah Petrovic, Nathan Warren and Marija Gavrilov of Exponential View. It is marked "Independently produced by Exponential View. Built from public disclosures and Exponential View's own models; all conclusions are our own," under the copyright of Epiiplus1 Ltd (Source: intelligence.exponentialview.co).

The report's stated purpose is to close a visibility gap on the demand side of the AI economy. Its framing is that the supply side is well documented because most semiconductor companies and hyperscalers are public, while "the demand side, what customers are actually paying for and if the revenues are real, has been obscure," since the largest labs are private and public companies bury AI revenue inside segment totals. Without that, the authors argue, "it is impossible to judge the health of the AI economy that underpins $22.7 trillion of stock market valuation and has driven US GDP growth in the past six quarters." Alex Imas, Shanu Mathew, Patrick Rutherford, Jaime Sevilla and Amy Sutter are thanked as reviewers of an early draft.

Method

The revenue model is bottom-up and line-level across more than 1,000 firms, in four stages. Each revenue line is sourced to primary filings, audited accounts, transcripts and credible reporting, plus cloud attribution where a private firm's revenue surfaces in a public firm's accounts (the report names OpenAI via Azure and Anthropic via Bedrock); softer signals include executive comments, proxy and sample metrics, and unverified estimates and leaks. Each line is then confidence-scored before entering any model, with filed figures graded highest, above other primary sources, corroborated third-party estimates and single-sourced claims, and "all derived numbers inherit the lowest grading from input sources." Full per-company GenAI financial models are built and triangulated against independent proxies — chip-maker revenue, build cost, segment mix, industry research, traffic and capacity. Finally revenue is deduplicated: counted at every layer but never summed across them, and attributed by value-add, so that "$100 app spend that sends $60 to a model provider, which spends $30 on inference hosting, is counted as $100, not $190" — split $40 to apps, $30 to model labs, $30 to hosting.

Scope is global excluding China. Counted: app, foundation-model and infrastructure-hosting revenue. Excluded: chips and hardware (treated as a cost to the compute layer rather than a customer payment), AI advertising uplift (named as primarily Alphabet and Meta), AI features inside legacy software, and capex and financing. Capex is the AI-attributable portion of seven stack-builders' infrastructure spend, hyperscalers and neoclouds, including cash property, plant and equipment and leases, counted once on the balance sheet of the owning entity. Tokens are every token processed, input and output, across all major providers and surfaces, attributed once to the model actually run so that a model served through a foundry is not double-counted; the token series, unlike the revenue series, is global including China.

Demand

The headline figures are $110 billion of banked trailing-twelve-month revenue, at a $175 billion annualized run rate — a 1.6x gap the report attributes to the growth rate. Measured against inflation-adjusted realized-revenue trajectories time-aligned to year zero, generative AI is described as scaling "three times faster than any IT wave," with the caveat that indexing GenAI growth to past technology waves "risks understating its speed" when modeling demand projections, depreciation schedules and capex paybacks.

The acceleration is expressed two ways. In 2023 the industry needed 180 days to add $1 billion in cumulative revenue; by the report's date it needed less than two days, a 90-fold compression. Quarter-on-quarter revenue growth held across what the report labels a Chatbot Subscription Era and an Agentic Coding Era, running at 35% quarter-on-quarter — equivalent to 3.2x annually — with Claude Code's release, the Codex launch and OpenClaw going viral marked on the series.

Downstream of that demand the report traces a compute supercycle:

Rising demand also shows in the combined hyperscaler backlog of remaining performance obligations, though the report notes the measures are not like-for-like: Microsoft's figure is total RPO including M365 and Dynamics, Amazon's is total company RPO (mostly AWS), Google's is revenue backlog (mostly Cloud), and Oracle's quarter ends a month earlier.

Scale relative to the economy

Against the wider economy the report's finding is that "big is still small, and early." Global ex-China AI revenue is equivalent to 0.42% of US GDP, against the IT sector's 9.4%; to 0.8% of labor costs; and to 3.0% of corporate profits — "even a generous yardstick (corporate profits) is 32x larger than all GenAI revenues." The ratio to GDP has risen roughly 3x against Q1 2025 (0.13%) and 10x against Q1 2024 (0.04%).

At company level the report uses Uber as a worked example: a maximum cap of $1,500 per engineer per month would put Uber in the top 10% of per-employee AI spend across Ramp's customer base of about 70,000 US businesses, while its $90 million annual AI spend sits below every comparable P&L line it is charted against — $720 million (12%), $3.4 billion (2.6%), $14 billion (0.6%), $8.7 billion (1.0%) and $52 billion (0.2%) for FY2025.

The report argues measured revenue understates value, drawing on four historical cases: Nordhaus (1996) on electric lighting becoming about 99.97% cheaper between 1880 and 1920, so that an hour's wage bought roughly 40,000 times more light with approximately zero direct GDP impact; Brynjolfsson et al. (2019) on free digital goods, with search alone valued at about $17,500 per person per year, again with approximately zero direct GDP impact; Crafts (2004) on steam adding about 0.4 percentage points a year to GDP between 1850 and 1870; and Graetz and Michaels (2018) on programmable machine tools adding about 0.37 percentage points a year between 1980 and 2000. The first pair are consumer surplus that escapes measurement; the second pair producer surplus that is recorded.

Applying that frame to AI, the report puts US consumer surplus from generative AI at $3–4 billion a month, about 30% on top of revenues, using Stanford Digital Economy Lab willingness-to-accept survey responses to the question, "Would you give up access to all AI tools like ChatGPT, Gemini, Claude, or Copilot for one month starting tomorrow morning in exchange for [$US]?"

Corporate reporting is tracked through S&P 500 earnings calls from Q4 2022 to Q1 2026: mentions of AI's impact rose 3–4x since 2023 and 50–60% of claims are now quantified, but "a majority of firms have not yet reported a quantified impact from AI use." The claimed outcomes skew to efficiency — cost reduction 25%, time savings 23%, throughput increase 22%, quality improvement 18%, conversion improvement 7%, revenue gain 6%, so that roughly seven in ten claims concern cost savings or efficiency. The report offers an accounting reason rather than a technological one: on a 30% net margin, a $1 million cost saving yields a 40% net margin while $1 million of added revenue yields 36%, so the same nominal impact looks better through savings.

On firm-level dispersion it pairs Bessen, Goos, Salomons and Van den Berge (2020) on automating firms outgrowing peers in employment by 37 percentage points between 2000 and 2016 with Ramp data showing a 92-percentage-point gap in revenue growth since November 2022 between high-AI-intensity firms (top 25% of AI spenders by share of revenue) and firms with no AI usage.

Capital expenditure and payback

Hyperscaler and neocloud capex reaches $2 trillion cumulatively through 2026E on a property-plant-and-equipment-plus-leases basis, with the caveat stated on the chart that "Total CapEx ≠ AI CapEx" because announced numbers include pre-planned spending for existing cloud and SaaS businesses, Meta's metaverse work and Amazon's logistics. AI-linked capex is put at $535 billion above the pre-AI trend by 2026E. Sell-side forecasts are shown chasing the curve upward across Barclays, Citi, Goldman Sachs, JP Morgan, Morgan Stanley, New Street Research, SemiAnalysis and UBS.

The report's more distinctive finding concerns funding mix and payback. The marginal AI-infrastructure dollar is increasingly externally financed: neocloud capex is primarily debt-funded while hyperscaler capex remains primarily cash. The authors state the risk transfer plainly — "external funding moves risk outside firms, as third-parties expect repayment," whereas "paying cash keeps a bad bet inside the firm, only denting profits" — while noting that hyperscalers' market-cap weight transmits risk to the wider economy regardless.

The 2026E depreciation charge approaches $111 billion, on IT equipment depreciated over six years and buildings over fourteen. Against it:

• Q4 2025 was the first quarter in which AI revenues exceeded quarterly capex depreciation.
• Q1 2026 headroom — defined as (revenue − depreciation) ÷ revenue — reached 19% for hyperscaler and neocloud revenues and 32% across all GenAI revenues. Equivalently, depreciation absorbs about 81% of hyperscaler and neocloud GenAI revenue and 68% of total GenAI revenue before any other cost.
• Cumulatively, revenue has nearly covered cumulative depreciation but not the expected headroom, which the report calls "still ~half-covered."
• The stated next test is incremental coverage: as committed capex enters service the depreciation base rises, so "revenue growth, utilization and pricing must continue to compound or headroom will compress again."

Two pieces of evidence bear on whether the six-year schedule is right. H100 one-year rental contract prices moved from a $3.05 per GPU-hour launch-era scarcity premium down to $1.70 at "peak overbuild fears as Blackwell ships" and back to $2.40 on "surging inference demand." And gross rental yields at 50% utilization run above the 17% six-year depreciation charge for newer GPUs and continue "long beyond their six-year depreciation life" for older ones. Extending assumed chip life from six to eight years — "as old as T4s" — raises infrastructure headroom from 19% to 36%; shortening it below six years leaves revenues failing to repay capex, which the report notes would require initial H100 purchases to be obsolete already. The section is framed against Mark Zuckerberg's remark on Meta's Q3 2025 earnings call that "the kind of very worst case would be that we effectively have just prebuilt for a couple of years, in which case, of course, there would be some loss and depreciation, but we'd grow into that and use it over time."

Cost to own and operate one gigawatt

The report models the annual cost of owning and operating 1 GW of AI capacity at $7.9 billion a year, of which capital costs are $7.0 billion (89%) and OpEx $900 million (11%).

Dividing by an output of 75–85 quadrillion tokens per gigawatt per year gives an inference-provider cost of $0.10 per million tokens for an open-weight model, requiring $0.20–$0.40 to earn a 50–75% gross margin and $0.25–$0.50 of customer value for a 25% return on investment. For an illustrative Kimi-class model under closed licensing the cost is $0.42 per million tokens, of which $0.32 is a licensing fee set at about 25% of Kimi's $1.29 blended selling price, requiring $0.84–$1.68 in price and $1.05–$2.10 in customer value on the same tests. The model is labelled illustrative: 6,700 GB200 NVL72 systems, cost of ownership per Epoch AI (May 2026) annualized including cost of capital, token output from SemiAnalysis InferenceX at FP4, 8k-in/1k-out, 50 tokens per second per user and 65% utilization (April 2026), with the range reflecting a 10–25% throughput uplift from speculative decoding.

Tokens

Global inference token volumes exceed 30 quadrillion a month, growing 14x year-on-year, on a series that includes China. Two drivers are given: the shift from chat to agents, measured as rising tool use per prompt on OpenRouter and rising average tokens per task; and the joint movement of capability and price, with the Epoch Capabilities Index rising from 112 to 158 while the blended price per million tokens fell from $17 to $2 and tokens processed per output token rose from 12 to 36 as reasoning models spend more tokens thinking.

Demand appears elastic. A log-log fit of Google's average price against volume gives a trend β of −1.7 with an R² of 0.93, and across providers the magnitude of elasticity is put at 1.2–1.8, so that "every 10% price cut → 12–18% more tokens → total token spend still rises." Three executive datapoints anchor it: Sundar Pichai at Google I/O 2025 — "…we were processing 9.7 trillion tokens a month. Now, over 480 trillion — 50x more" — against a 97% price decline; Sam Altman's "Three Observations" (2025) statement that "the cost to use a given level of AI falls about 10x every 12 months, and lower prices lead to much more use," against a 90% price decline; and Tan Dai of Volcengine/ByteDance reporting Doubao's daily token usage exceeding 50 trillion, up from 4 trillion in December 2024, against a 50% price decline. The authors caveat the regression as a time series rather than a cross-section, so that "β may overstate pure price-elasticity."

Efficiency per unit of physical capacity rises on every generation, driven by smaller models and better serving, hardware gains through Hopper to Blackwell to Rubin, and a workload mix shifting from training to inference. Revenue per trillion tokens has fallen since its 2023 peak while industry-wide revenue per gigawatt of data-center capacity passed $7 billion per GW.

The report is sceptical that tokens are the right unit of value. It places generative AI alongside electricity (lamps, then kilowatt-hours), the internet (pageviews, then sessions and clicks, CPM to CPC to CPA) and mobile internet (megabytes, then active users), and concludes that for AI "the value-producing unit is still undefined." Its candidate is quality-adjusted output tokens — output tokens scaled by model capability, excluding input and reasoning tokens as costs of production rather than user-facing output. Measured from January 2025 to April 2026, total tokens grew 39x and raw output tokens 30x, while quality-adjusted output grew 33x on Epoch Capabilities Index weighting, 59x on METR task-horizon weighting and 35x on the Artificial Analysis Intelligence Index. The authors note the 33x–59x spread means "we can only draw directional conclusions."

Token-based pricing is characterized as AI's "pay-per-click moment," the third stage after free and seat-based subscription pricing, on the analogy that untracked banner advertising had a limited market that crashed with the dot-com bust while attribution through cost-per-click enabled a sustainable market after Google AdWords in 2002.

Where value is captured

Revenue is concentrated but the mix is shifting toward apps and models. On the deduplicated quarterly series, the layer split moves across three plotted quarters from 89%/8%/3% to 87%/9%/4% to 82%/11%/7%, with total quarterly revenue growing 2.95x in a year; on the chart's legend ordering the largest share is hosting, then foundation models, then apps. Chips are excluded from the split because they are capitalized as capex by the hosting layer.

Concentration is measured per layer by Herfindahl–Hirschman index, with hosting and apps defined by revenue share, foundation models by token share (because of open-weight competition) and chips by share of compute in H100-equivalents (because of hyperscaler vertical integration). The report's reading is that "pricing power follows competitive pressure, not stack position": upstream suppliers price to capture all available margin absent downstream competition; Nvidia is the largest chip provider but vertical integration by AWS and Google into custom silicon "may reduce its prominence"; and foundation-model revenue is concentrated with OpenAI and Anthropic, but open-weight models "offer low-cost competition for quality tokens."

The premium frontier labs charge is described as time-limited. Grouping models by performance at GPQA Diamond, price per capability falls quickly, so "labs earn a time-limited premium on the current frontier" and must "outrun open-weight commoditization to hold margin." The report sketches a task ladder — lightweight open-weight models for summarizing an article, leading open-weight models for drafting working code, closed-weight frontier models for multi-day agentic work. Among self-selecting OpenRouter users, the combined token share of Google, OpenAI and Anthropic fell from 72% to 33%, with the caveat that OpenRouter "is not a cross-section of the market" but shows the behaviour of model-routing users. Under this pressure, labs are described as pushing into vertical applications, with Codex for Legal and Claude for Legal named as examples.

Four scenarios are set out for what happens if layers compress: open-weight models catching up, so labs cannot charge a licensing fee and tokens get cheaper for apps and hosting; general-purpose models absorbing the integration layer, squeezing generic wrapper pricing while apps defend with proprietary data, domain-specific workflows and their own models; reduced compute needs, with smaller and more efficient models growing the addressable market and some hosting moving to edge devices; and all three combined. In each, the report's conclusion is the same: "cost reductions grow the market and result in increased consumer surplus."

Stated conclusion

The report's closing formulation is that "AI demand is more revenue-validated than any prior platform shift," and that "the investment case comes down to whether falling prices can move enough token volume to earn a return on CapEx." The open question it names at the front is "whether cheapening artificial intelligence can create enough volume and margin to service the buildout."

Provenance and limits

Fetched in full (66 of 66 pages, HTTP 200) from exponentialview.co, the publisher's own domain, on August 28, 2026 and extracted from the PDF text layer. The cover names the five authors and carries the copyright line "© Epiiplus1 Ltd 2026" and the statement "Independently produced by Exponential View." A developments digest of August 27, 2026 referenced the report in connection with an Epoch AI revenue item, dating it only as "2026"; the report's own cover dates it June 25, 2026, two months earlier.

The report is an analyst product built substantially on the publisher's own proprietary model rather than a peer-reviewed or filed document, and most series carry "Exponential View analysis" as their sole or leading source. Its figures are estimates with an internal confidence grading that is not published per line. The deck is chart-heavy, and axis values for several series sit inside raster images that the text layer does not carry, so the figures recorded above are those stated in labels, callouts and notes; the per-layer assignment of the 82%/11%/7% split follows the chart legend's ordering.

Relationships

• supports: AI Bubble Debate — supplies a demand-side revenue measure and a depreciation-coverage test.
• supports: AI Bubble vs. Buildout — Synthesis — a middle-position quantitative anchor built from customer revenue rather than supplier disclosure.
• related: AI Infrastructure Capex — the $2 trillion cumulative and $535 billion above-trend capex figures.
• related: AI Data Centers — the per-gigawatt cost model and the build-cost mix shift toward silicon.
• related: Inference Economics and Token Pricing — token cost, pricing and elasticity.
• related: Open-Weight Frontier Models — commoditization of last year's frontier as the constraint on lab margin.
• related: Azeem Azhar — lead author.
• related: Epoch AI — the Epoch Capabilities Index and cost-of-ownership figures the report builds on.
• related: METR — the task-horizon series used as one quality adjustment.