The price of AI is collapsing, while the cost of building it is not. Equity investors have spent the summer trying to work out who gets caught in between.
Consider the past 10 days: A free model called Ox Alpha appeared online, performing near the frontier of what AI can do, and nobody will even say who built it. OpenAI cut prices on its flagship model for the third time in about a month. Meanwhile, the token-price gauge this column flagged as a warning signal back in July has kept sliding since. Intelligence, as a product, is deflating in real time.
On its own, that last point isn’t the drastic warning sign it’s sometimes made out to be. The Silicon Data LLM Token Expenditure Index blends prices with usage, and its maker calls it a proxy for what buyers are willing to pay, rather than a simple price tag.
More fundamentally, falling prices are how new technologies conquer the world, and each chip generation cuts the cost of producing a token far faster than sellers reduce what they charge for one. The trouble starts only when the arithmetic flips, when what AI sells for falls faster than the expense of building it.
The semiconductor market is entering a period of structural undersupply, from foundry to memory, suggesting compute pricing will stay elevated. Some of Nvidia Corp.’s biggest customers have been told that the prices of servers containing its AI chips are going up more than 15% in many systems to be shipped early next year, Bloomberg News has reported. The company gave a bullish sales outlook last night, expecting to grow revenue by approximately 70% in the fiscal year of 2028.
Samsung Electronics Co. has raised prices for some advanced contract chipmaking by up to 15% for new orders and is locking as much as 70% of its memory capacity into multiyear contracts with data-center buyers. SK Group’s chairman has warned that the memory crunch will worsen in 2027, even as SK Hynix Inc. explores a joint-venture model to fund new plants. The product is deflating while materials are rationed and repriced upward. That’s the arithmetic flipping.
The bulls have a good riposte to this: volume. Cheaper intelligence means more gets consumed. The share of U.S. businesses paying for AI is approaching 60% and spending has more than tripled at every level of the distribution. Three hyperscalers recorded combined cloud revenue of roughly $106 billion last quarter, up more than 40% in a year. Ramp estimates the heaviest spenders’ outlay at $7,400 per employee monthly, which might sound stretched, while the median firm’s is at about $12, which is anything but. On this reading, the price collapse is the business model, not the threat to it.
Which brings us to the question readers pushed hardest after last month’s piece: where, exactly, is the return? A study by Milos Maricic, founder of the AI advisory firm Maximand, went looking for it in a sector where the productivity story should be easy to prove: white-collar work built on text and numbers. Across 919 earnings calls of the 60 largest US-listed financial firms over three years, four in five teleconferences mentioned AI and more than half discussed its cost. As for the number that highlighted a realized dollar return from using it: that was just one firm — twice, for about $19 million combined. Benefits touted at investor days and on television rarely survive contact with an earnings call.
“Three years into the AI buildout, the firms buying the technology still cannot put a dollar figure on the payoff,” Maricic said.
Credit markets are showing increased nervousness around financing AI infrastructure. Broadcom Inc. is in talks to raise more than $60 billion in debt to fund AI chips, Bloomberg News reported last week, part of a widening rush to borrow against tomorrow’s compute demand. Credit has started charging for its doubts. Insuring Broadcom’s debt costs about 80 basis points more than it did in January, and the premium is accelerating.
“If you assign any probability that companies ultimately cannot finance everything they intend to build, the answer is either more equity or less capex. Neither deserves a higher multiple,” wrote Rich Privorotsky, Goldman Sachs Group Inc.’s head of European one-delta trading. “Nobody has cut an EPS number or said they plan to spend $1 less… the equity market is simply pricing a wider range of outcomes.”
That last line sums up the summer. Estimates keep rising and multiples keep shrinking, which is what a market does when it stops debating whether the machine works and switches focus to who gets paid for it. Nvidia’s earnings provided some relief with the caveat that the company warned about a potential narrowing of margins as memory costs rise.
The bull case focused on volume survives all of this. Adoption keeps climbing, and the return numbers may yet show up on earnings calls the way e-commerce profits once did — late but enormous. Cheap intelligence would then be the mechanism that pays for everything, including the debt.
It could just as easily go wrong. A product deflating at this speed, built from inputs priced for scarcity, funded by ever costlier leverage and sold to customers unable to yet measure what it earns them requires a lot of moving parts to align.
In July, this was a story about whether AI’s pricing power holds. The fresh evidence says it is wearing thin from both ends, with buyers pushing back on price while suppliers reprice the inputs. The market’s response has been neither panic nor conviction, just a growing premium to compensate for uncertainty. That’s not a verdict, but an admission that one is coming.
Barnert writes for Bloomberg.