Anthropic is reportedly heading for an October IPO that could value it at two trillion dollars, the largest in history. OpenAI has filed too. And into that euphoria this week arrived a chorus of people asking, from very different directions, whether the numbers underneath actually add up.
Start with the demand, because on the surface it looks unstoppable. Business Insider reported that after OpenAI cut the price of its Luna model by 80 percent, the effective cost of using it fell roughly tenfold while consumption jumped about fourteenfold. Revenue rose even as the price collapsed. The writer reluctantly reached for the phrase everyone in Silicon Valley loves: Jevons Paradox, the 19th-century observation that making a resource cheaper can grow its total use rather than shrink it. Cheaper AI simply gets pointed at more jobs.
The trouble is that soaring usage is not the same as customers making money. Buried on page 35 of a 69-page OpenAI report on enterprise adoption, Fortune found a small table showing no statistically significant correlation between how much employees use AI and their company’s revenue per head. The most profitable firms were the earliest adopters, but spending more tokens did not track with earning more. Perhaps, Fortune noted drily, the richest companies simply have the most to spend. The same report showed executives, the people signing the cheques, are the lightest users of all.
If the buyers cannot yet prove the return, the sellers face a harder squeeze still. Steve Eisman, the investor made famous by The Big Short, told CNBC that the “Achilles heel” of the entire AI trade is that roughly 70 percent of AI revenue at Microsoft, Amazon, Google and Oracle flows from just two customers: OpenAI and Anthropic. A price war with cheaper Chinese models, he warned, would leave everyone exposed. He is holding judgement until Anthropic goes public and real quarterly numbers replace anecdote.
Security writer Bruce Schneier and co-author Nathan Sanders went further in The Guardian, arguing the labs may never be sustainably profitable at all: frontier models are expensive to train, depreciate within months, and behave enough like commodities that open-source and Chinese rivals give away for free what the leaders sell. Their provocative conclusion is that if the market rejects OpenAI and Anthropic, the United States should nationalise them and run them as public labs, too valuable to the public to let die even if they are worthless as equities.
Zoom out and the pattern is familiar. Writing in TechRadar, the head of an AI firm noted there are more than 70,000 AI companies today and pointed to history: of some 50,000 internet startups founded around the dot-com boom, fewer than ten became enduring giants. The middleware layer, he argues, tends to get absorbed by the platforms while durable value accrues to a handful of application builders on top.
The nervous energy is visible inside the leaders themselves. OpenAI this week lost its second senior executive in three days, with chief revenue officer Denise Dresser departing after eight months, just as the company tries to show public markets a stable, growing enterprise business. Usage has never been higher. Whether that translates into profit remains, for now, the industry’s most expensive unanswered question.