TL;DR
The short version
Aggregated across a thousand-plus companies, generative AI banked about $110B over twelve months and sits at a roughly $175B annualized run rate — growing about 3x faster than any prior IT wave. Starting Q1 2026, quarterly AI revenue began exceeding the depreciation on the hardware it runs on. Demand is revenue-validated, not just hype.
Two things matter to an operator. The subsidy era funding cheap inference is ending — a wholesale model rate converts to token pricing, cloud GPU prices rose, and memory is spiking — so cost discipline returns. And the gap between firms that use AI well and those that don't is enormous: top-quartile spenders grew revenue more than 100% over three years against 15–20% for non-adopters.
Built from The AI Daily Brief's walkthrough of Exponential View's inaugural State of the AI Economy report.
Is it paying back, or is it a bubble?
Demand is revenue-validated. The report aggregates spend across a thousand-plus companies — weighting audited accounts over executive quotes and deduplicating so a dollar is counted once — and lands on roughly $110B banked over the trailing twelve months, a ~$175B annualized run rate, growing about 3x faster than any previous IT wave.
The largest buildout in tech history is paying back — for now.
Nathaniel Whittemore, The AI Daily Brief
But the hedge is load-bearing. Starting Q1 2026, quarterly AI revenue began exceeding CapEx depreciation, and GPUs are earning yields into years seven to nine — well past the six-year depreciation assumption. Independent coverage pegs Q1 2026 AI sales outside China at roughly $25B against about $21B in quarterly data-center and chip depreciation. That covers the running bill, not the cumulative one.
Still a rounding error on GDP. AI revenue is about 0.42% of US GDP against the IT sector's ~9.4%. The base is tiny; the slope is steep — AI-revenue-to-GDP has grown 3x versus a year earlier and 10x versus two years earlier. As one observer put it, not enough people are emotionally prepared for the case where it isn't a bubble.
When the market sells off, is demand falling?
A useful stress test arrived in August 2026: a sharp AI-stock selloff that turned out to be market mechanics, not a demand verdict. A Fed on hold, a risk-off macro rotation, a Korean retail-leverage margin-call cascade that hit roughly 1.2M accounts, and the forced liquidation of a 4x-levered AI hedge fund — Leopold Aschenbrenner's Situational Awareness, its book absorbed by Citadel — dragged the semiconductor names regardless of anyone's AI conviction.
Underneath the price action, lab demand kept re-accelerating. OpenAI's CFO told staff July annualized revenue topped all of Q2; AWS grew 37% year-on-year and Amazon raised capex from $200B to $220B on demand; Azure crossed $100B in annual revenue for the first time.
The operator read: treat AI-name stock weakness as a leverage-and-plumbing signal, not a demand signal. In this cycle the two came apart — the same "demand is upstream of the trade" point, now tested by an actual selloff. Read the revenue run-rate figures with care (the ~$70B Anthropic number and the fund's exact size are estimates, not disclosures), and note the systemic-risk debate around data-center debt vehicles is a separate finance argument, not a read on whether AI demand is real.
The subsidy era is ending
Cheap inference has been partly subsidized, and the props are coming off. A major wholesale model rate converts to token-based pricing next year, a large cloud raised its GPU capacity-block prices about 20%, and a memory shortage is taxing all electronics — a leading memory maker posted a record 84.9% gross margin as DRAM prices rose roughly 60% quarter-on-quarter. Blended token price still fell from about $17 to $2 per million over two years, but the tailwind behind that fall is weakening. Cost-per-outcome is about to matter again.
The distillation trap. As models become strategic assets, labs are getting defensive about being copied through their own tools. One major lab barred rival AI coding tools on its data-labeling tasks to avoid training its own frontier coding model on competitors' outputs — model distillation violates the leading labs' terms of service, and accusations of large-scale distillation are already flying. Every lab building an internal frontier model hits the same bind: the best available tool is also a contamination and legal risk.
Most of the headline growth figures here are the report's own analyst-modeled aggregates rather than audited numbers — directionally strong, but read the methodology before repeating any single figure. The pricing and margin data points are independently verified.
