Azeem Azhar’s Exponential View published what it claims is the first deduplicated, bottom-up measure of the generative AI economy on June 26, finding that the sector generated $110 billion in total revenue over the past twelve months and is now running at an annualized rate exceeding $175 billion.
The methodology is worth understanding before citing the headline figures. The Exponential View team built a line-item financial model for each major company in the stack, constructing P&L estimates and cash flows that they then triangulated against public statements, reported leaks, and supplier disclosures. The central design choice was to count only the dollar that reaches an end customer, not the same dollar twice as it passes through the supply chain. When an enterprise pays Anthropic and Anthropic pays Amazon, the model counts one dollar, not two. That approach is necessary because most of the revenue flows through private companies, including OpenAI, Anthropic, Cursor, and ElevenLabs, that disclose nothing, and the public hyperscalers, Amazon, Google, and Microsoft, that do not consistently break out AI segment revenues.
The $110 billion figure excludes several things that matter for context. It does not include internal AI uplift, meaning the ad revenue Meta and Google earn because recommendation systems got better. It excludes efficiency savings from internal enterprise tools. It excludes professional services and systems-integration fees. And this version of the model contains no Chinese revenue data at all. The number is, in other words, a floor rather than a ceiling of total AI economic activity.
The supply side of this market is already well-documented through public company filings. Hyperscalers, chip makers, memory suppliers, and power infrastructure companies are public and disclose their forward order books. The demand side, which is what Exponential View is measuring, has remained opaque precisely because demand is concentrated in private companies with no disclosure obligations. This asymmetry has let bullish narratives dominate, because anyone who counted only the supply side found enormous numbers with no countervailing demand signal to test them against.
The revenue-versus-investment question is the number that matters most to anyone thinking about the sector’s durability. The model separates AI-specific capital expenditure from the pre-existing infrastructure spend at the major hyperscalers, who were already running roughly $120 billion annually in CapEx before ChatGPT launched. After carving out the AI-incremental portion and depreciating compute assets over six years, the Exponential View model finds that hyperscaler revenues attributable to AI just about cover the depreciation expense. That is a thin margin. Six-year depreciation is defensible, the team argues, because current demand still exceeds available compute and operators are improving GPU fleet efficiency. Both conditions could change.
The token-price dynamic is where the analysis gets genuinely interesting. As frontier model costs fall, the standard fear is that revenue per unit collapses faster than volume can compensate. The Exponential View model finds the opposite, for now: every ten-percent price cut leads to twelve to eighteen percent more tokens consumed, so total spend still rises. That is a positive demand elasticity, and it is the core argument for why the revenue line survives deflationary model pricing. The analysis also surfaces a conceptual problem with tokens as a unit of economic measurement. Token counts do not capture model quality. A quality-adjusted output token, combining volume, user-visible outputs, and underlying model capability, gives a more honest read of what the market is actually buying.
The report’s coverage of enterprise adoption confirms a familiar pattern. Many companies have moved past pilots into early-stage deployment, but deepening and scaling are still ahead. In conversations Azhar reports having with senior executives across European and US industries spanning industrials, insurance, finance, and pharma, the consistent signal is intent to increase AI investment. A BCG survey cited in the piece found that roughly half of CEOs believe their jobs depend on getting AI right, which is less a data point than a description of the political economy inside large enterprises right now.
The macro framing the Exponential View team offers, that this market is growing roughly three times faster than the mobile or internet waves did, is the claim that most needs independent verification. Prior wave comparisons depend heavily on which starting point and deflator you choose. Operators building product and infrastructure decisions around the $175 billion run rate should treat it as a useful directional reference and note that the model’s China gap means the real global number is likely higher.
For any operator currently repricing AI contracts or negotiating consumption commitments with hyperscalers, the token-elasticity finding is the most actionable input from this report: lower prices appear to expand total spend, which means locking in high per-token rates now is a worse deal than it looks.
Published in Exponential View on June 26, 2026.
