Generative AI just had its worst stretch in a long time, and the warning is coming from one of the field’s loudest skeptics. According to Marcus on AI, the past few weeks have stripped the shine off the whole sector: OpenAI is now leaning toward delaying its IPO until next year, AI-linked stocks are sliding, and Washington has quietly started rationing access to frontier models. Marcus on AI calls it the month generative AI lost its mojo, and the data backs the headline.
The OpenAI IPO wobble
The New York Times reports OpenAI is leaning toward holding off its public offering until 2027. The reason, per people involved in the deliberations, is a lack of confidence that retail investors will buy the story Sam Altman wants to sell.
One rumor making the rounds: Altman pitched a trillion-dollar valuation, and his own advisers told him the math wouldn’t fly with everyday buyers. The choice on the table was wait until 2027 for the $1T number, or accept a lower valuation now. Marcus on AI has compared OpenAI to WeWork for a while, and a stalled IPO fits that script.
The selloff is broad
This isn’t one company having a bad week. The damage runs across the AI supply chain. Over the last month, per Marcus on AI:
- Nvidia: down over 8%
- Oracle: down about 22%
- Microsoft: down 10%
- SoftBank: down about 12%
- CoreWeave: down a bit over 4%
- Cerebras: down 32%
SpaceX, increasingly pitched as an AI play, is getting hit too. Its $25 billion bond sale is showing paper losses north of $300 million, with traders saying fast-money investors may be heading for the exits.
What stands out here is the spread. When chipmakers, cloud providers, and the marquee model labs all bleed at once, that’s a sentiment shift, not a single bad earnings print.
Washington tightens the valve
The policy story is just as striking. Marcus on AI flags reporting that the Trump administration asked OpenAI to stagger the release of GPT-5.6 over security concerns, with the government approving access customer by customer. Andrew Curran reported the same staggered rollout and suggested it could become the norm for all frontier models from every lab.
Marcus on AI’s read is blunt: U.S. federal AI policy has swung from implausibly libertarian to draconian and opaque in a matter of weeks. If the White House decides ad hoc who gets frontier intelligence, that’s a real problem for any company building on top of these models.
Reality checks on the hype
The technical claims are taking hits too. Cardiologist and medical AI bull Eric Topol said his team stress-tested frontier models for multimodal medical reasoning and found them “not ready,” citing faulty reasoning, bad shortcuts, and hallucinations. Nobel laureate Jennifer Doudna, who co-invented CRISPR, said breakthrough innovation is still a human domain, noting she isn’t seeing chatbots come up with brand-new ideas.
Meanwhile, Chinese models are gaining fast. Marcus on AI points to OpenRouter data showing the share of tokens used for U.S. models has collapsed. His long-running thesis is that LLMs are becoming a commodity, and a price war plus capable open-source competitors is exactly what commoditization looks like.
The bull case, and why it’s contested
Not everyone is bearish. Economist Noah Smith argues the bubble talk is fading, pointing to rising revenue and what may be Anthropic’s first profitable quarter. Marcus on AI’s counter is that revenue going up doesn’t prove profitability is sustainable, especially with token costs, price wars, and open-source pressure all working against margins.
That’s the real debate. Revenue growth is genuine. Whether it survives contact with the cost structure is the open question.
Why it matters
For anyone building on or investing around these models, the signal is to plan for turbulence. A delayed OpenAI IPO chills the funding climate. Per-customer government approval slows your access to the newest models. And commoditization means today’s frontier edge may not be worth the premium tomorrow.
One rough month doesn’t end the AI era. But it does puncture the idea that the line only goes up. For the full breakdown and Marcus’s policy prescription, the original analysis at Marcus on AI is worth a read.