How big is the AI bubble? We sized it: a $1-trillion-a-year hole, and growing.
The buildout needs roughly $1 to $2 trillion a year to pay off. The two largest AI companies collect about $70 billion a year; the whole market, by our estimate, is maybe $120 billion. Close to a third of even that is the industry paying itself.
- The buildout needs $1 to $2 trillion a year to pay off. The two largest AI companies collect about $70 billion a year; we estimate the whole market near $120 billion. Either way, far below the $1 to $2 trillion needed.
- The gap is widening, not closing. Over 2025 to 2026, AI revenue more than doubled, yet the dollar gap still grew about $380 billion.
- About a third of AI revenue is circular. Suppliers fund the labs that rent their compute back: roughly $50 to $60 billion is the industry paying itself.
- Even AI's own forecasts don't close it. The leaders' bullish 2030 targets reach only 15 to 20% of what's needed, and arrive unprofitable.
AI Revenue is Exploding but the Gap is Still Widening
To justify the historic capex spend on AI, the industry needs to generate far more revenue every year than is actually being collected. This gap keeps widening even as AI revenue grows fast.
Everyone agrees AI might be a bubble. The debate is also where the conversation stops: "is it a bubble" is a yes-or-no question argued mostly with adjectives, and the loudest claims (Burry, the "$600 billion question") each rest on a single back-of-envelope. We did the boring work instead, and measured the gap.
Four methods, one order of magnitude
We did not take the scary number on faith. We built the requirement four ways and added our own. They disagree on the figure, not the order of magnitude.
| Source | Revenue needed | Method |
|---|---|---|
| JPMorganfloor | ~$650B / yr | Revenue for a 10% return on the investment |
| Sequoia / Cahn | ~$1.2T / yr | NVIDIA data-center run-rate × 2 (cost) × 2 (margin) |
| WireSift (ours) | ~$1.3T / yr | Capacity build-up: ~156 GW × ~$8.5B/GW (Epoch) |
| McKinsey | ~$1.5–2T / yr | Projected AI data-center capex |
| Bainceiling | ~$2T / yr | Compute demand, by 2030 |
The single biggest swing factor is mundane: how fast the chips wear out. Shorten the assumed hardware life and the bar jumps.
| Chip Life | Required AI Revenue |
|---|---|
| 6 years | ~$1.1T |
| 5 years | ~$1.3T |
| 3 years | ~$1.9T |
Real demand is ~$120B, and it's softer than the headline
- ~$70B is collected by the two largest AI companies (OpenAI + Anthropic) combined; we estimate the whole end-customer market at ~$120B, generously counted.
- ~$185Bis the naive sum of every company’s reported AI revenue, but it double-counts: ~$28B of Microsoft’s “AI revenue” is just OpenAI renting Azure.
- In the public markets, disclosed AI dollars are dominated by picks-and-shovels suppliers (NVIDIA, Broadcom, AMD, networking) — not companies selling AI to customers. The money is in building it, not yet in using it.
A third of the revenue is the industry paying itself
The pattern repeats across every major deal: the supplier funds the customer, the customer spends it back on the supplier, who books it as AI revenue.
| Funder | Invests in | Who commits it back |
|---|---|---|
| Microsoft | OpenAI | $250B committed to Azure |
| Amazon ($8B+ in) | Anthropic | $100B+ committed to AWS |
| NVIDIA (up to $100B in) | OpenAI | compute purchases |
| Oracle | OpenAI | $300B compute deal |
We estimate about a third of current AI revenue, roughly $50 to $60 billion, is intra-industry. The clearest example: we estimate OpenAI's ~$24B/yr of Azure spending is the large majority of Microsoft's ~$28B "Azure AI" line. Bloomberg counts $800B+ of committed circular financing. Anyone who lived through 2000 will recognize the shape: Nortel and Lucent lent money to the customers who bought their gear.
Even AI's own forecasts don't close it
Grant the bulls their most optimistic numbers and demand still reaches a fraction of the requirement, and gets there unprofitably (OpenAI is not cash-flow positive until ~2030 and just added ~$111B to its burn forecast). Anthropic's 2030 figure is extrapolated from its 2027 guidance. The bull case falls short on its own arithmetic.
AI works at the task level. The economy-wide payoff hasn’t shown up yet.
- +14% support tickets resolved per hour
- +26% more code shipped by developers
- −40% time on professional writing tasks
- 80%+ of firms report no measurable productivity gain
- 95% of generative-AI pilots show no return
- US total factor productivity grew, but decelerated
General-purpose technologies take decades to show up in aggregate productivity; electricity and the computer both looked like dead weight for years. The returns may simply be early. We size the gap. We do not call the timing of when, or whether, it closes.
Every figure traces to a sourced model: the requirement triangulated across JPMorgan, Sequoia, Bain, McKinsey and our own capacity build-up; actual revenue de-duplicated layer by layer; the circular share anchored on disclosed compute commitments. Revenue figures are annualized run-rates, not full-year recognized. See the methodology for how each figure is built and sourced.
We re-measure the gap every earnings season.
Get each update, sourced and sized, before it makes the headlines.