Gary Marcus Trademarked a Rebranded Failed Prediction

AI Datum Point

Jun 26, 2026

Gary Marcus has been predicting the collapse of generative AI since August 2023. The prediction has not been right. It has, however, been consistent: every negative data point confirming the thesis, every positive data point treated as noise or anomaly, and the goalposts moved at intervals when the collapse failed to materialise.

Last week Marcus filed Accenture's Q3 2026 earnings as confirmation. The stock fell 18%. Marcus did not mention that adjusted EPS beat consensus by nine cents, that revenue came in at $18.7 billion up 6% year on year, or that management attributed the guidance cut to the Middle East conflict, elongated EMEA decision cycles, and two large managed services deals slipping to FY27. RBC Capital, cutting its price target after the print, noted that "underlying AI demand remains healthy" and cited 100 new advanced AI projects initiated in Q3. The revenue miss against consensus was $80 million: less than 0.4% of quarterly revenue. Marcus read the stock chart. He did not read the earnings report.

Wednesday he coined a new term. The "Generative AI Fizzle™" replaces the collapse thesis with a slow decline narrative. The trademark symbol is the tell. A concept that required trademarking is a concept being protected from scrutiny.

Marcus argues that LLMs have become commodities, price wars have made profits elusive, and the moat he warned about never existed. On the commoditisation point he is not wrong. The MMLU benchmark gap between frontier closed models and open-source alternatives has narrowed from 17.5 percentage points to 0.3 points in under two years. DeepSeek trained a frontier-competitive model for $5.9 million. The moat argument was always more fragile than the valuations implied.

But "profits are elusive" is not the same as "profits do not exist." The capital flowing through AI markets tells a more granular story than Marcus is telling. Cursor, the AI coding tool founded by four MIT graduates in 2022, hit an estimated $4 billion in annualised revenue in May 2026, per Sacra. SpaceX signed a definitive agreement to acquire it for $60 billion on June 16. Sierra, the enterprise AI agent platform founded in 2023, crossed $200 million ARR in May 2026, seven quarters after launch. Sierra's own blog addressed the apparent contradiction directly: the difference between the 95% of AI pilots that fail and the companies generating real returns is the difference between "AI tourism" and tools focused on specific jobs to be done. These are not GPT wrapper businesses. They are purpose-built tools solving specific problems that people pay for at scale.

There is also a documented problem with how AI revenue is being measured at the early stage. A May 2026 TechCrunch investigation found that VCs and founders are using inflated ARR figures to crown AI startups, with boards knowingly counting pilot programme revenue as ARR before contracts converted to paying customers. This is not the Cursor or Sierra story. Both companies have disclosed funding rounds, acquisition deals, and independently estimated revenue at a scale that pilot inflation does not explain. It is the story of the layer below them, where the "profits are elusive" narrative lives. If that layer's figures are inflated, Marcus's argument is understating the problem rather than overstating it. Either way, the fizzle thesis is not the analysis. It is the reframing.

The structural condition Marcus keeps circling without landing on is this: the LLM layer has commoditised and the moat has largely gone. That is true and the evidence confirms it. But the investment is not primarily going to the LLM layer anymore. Mega-rounds above $100 million are concentrating in AI companies with proven product-market fit and enterprise revenue traction. Capital has already sorted. Physical AI: robotics, autonomous systems, real-world applications. Crunchbase confirmed in June that robotics and physical AI startups had already exceeded every prior full-year funding record before July, with $18.8 billion raised year-to-date on its narrower measure and $55.8 billion on Dealroom's broader physical AI definition. Either way, a record before the year is half done. The money that understood the commoditisation thesis moved before Marcus finished writing about it.

The fizzle may come. The structural pressures Marcus identifies are real. But the evidence for where the money actually went, and what it is producing, does not support the "nothing is working" reading that the Fizzle™ implies. The ones that were always going to work are working: infrastructure, physical AI, purpose-built enterprise tools. The ones built on the assumption that an LLM API wrapper was a business are not.

That is not a fizzle. That is a market doing what markets do.