Longtime AI skeptic Gary Marcus calls the slow deflation of generative AI equities the 'Generative AI Fizzle™,' a framing that hinges on a capex to revenue mismatch: roughly $700 billion a year in locked in Big Tech AI spending on one side of the ledger, and leaked OpenAI…
The phrase "Generative AI Fizzle™" is Gary Marcus's, and it is deliberately not the same word as "crash." Marcus, a longtime skeptic of generative AI, coined it on his Substack to describe a slow-deflation scenario in which the gap between AI infrastructure spending and AI revenue widens quietly until equity multiples compress under their own weight. The metaphor is the easy part. The mechanism is more specific.
That mechanism is a unit-of-account constraint. On one side of the ledger, roughly $700 billion a year in Big Tech AI capex commitments this year, locked in by Microsoft, Amazon, Google, and Meta and projected by Goldman Sachs to exceed Japan's GDP over the multi-year horizon through 2030, a framing Business Insider has used to translate that run-rate into a comparable scale. On the other side, leaked OpenAI financial documents reported by Ars Technica and Fortune show approximately $13 billion in revenue against approximately $21 billion in losses. The capex side is committed. The revenue side is not.
OpenAI is the cleanest case study because it is the purest exposure. Its losses are not hidden in a broader services business; they are the business. The leaked figures, which have not been confirmed by an audited filing, frame a company whose compute costs scale with usage while its pricing model still relies on subscriptions and per-seat enterprise contracts. A loss-to-revenue ratio above 1.5x, sustained, is not a margin problem to be optimized away. It is a structural mismatch between cost of goods sold and the willingness of customers to pay.
The hyperscalers look different, and that is the part of the fizzle story that separates pure-play exposure from vertically integrated exposure. Microsoft, Amazon, Google, and Meta can absorb hundreds of billions in AI capex because each one monetizes AI as an additive layer on top of cloud, advertising, search, and productivity revenue that already exists. The $700 billion figure is real money, but it is being financed out of operating cash flow from businesses that do not require generative AI to function. OpenAI does not have that option. Its investors do.
Marcus is careful to disclaim stock-pick status. He notes, in the same post, that markets can remain irrational longer than any one short can remain solvent. That caveat matters because it limits what the fizzle framing is actually claiming. It is not a prediction that AI equities fall next quarter. It is a claim about the trajectory implied by the capex-to-revenue ratio when extended across the next several capital cycles: each cycle will demand another trillion-dollar commitment before the last one is paid for.
The structural question then becomes what would falsify the framing. The cleanest test is straightforward: can OpenAI, or any equivalent pure-play generative AI company, demonstrate sustained quarter-on-quarter revenue growth at a rate that meaningfully narrows the loss-to-revenue ratio within 18 months? If yes, the capex commitment is a normal infrastructure bet that compounds. If no, the next funding round is not a step up but a step sideways, and equity dilution or recapitalization becomes the more likely path than the next capability leap.
The next signal to watch is the capex guidance in Microsoft's, Amazon's, Google's, and Meta's upcoming earnings calls, which arrives ahead of OpenAI's next disclosed financial period. If those numbers hold, the fizzle is a slow process. If they pull back, the mechanism tightens faster than Marcus's timeline.