The AI industry now grows on everything except intelligence

OpenAI's breakout growth product this year is an ad network: ChatGPT's ads business went from a $100 million annual run rate — the year's total, if the pace held — in March to $1 billion by August ^[1]. Amazon Web Services just committed a billion dollars to an army of human consultants, engineers it embeds inside customers' offices ^[2]. And Meta, the company that could not get anyone to pay for its chatbot, is renting out raw GPU hours — a move analysts read as an admission it could not monetize AI directly ^[3][4]. Three of the most powerful companies in the field, and none of them is betting its growth on intelligence anymore. Look at the buyers' ledger and the reason shows up fast. Apollo's chief economist measured the entire S&P 500 and found AI has delivered zero financial payoff to corporate America outside tech — the Magnificent Seven's margin and revenue-per-worker gains since 2022 have not extended to a single other industry ^[5]. MIT's State of AI in Business report reached the same wall from inside: 95% of enterprise generative AI pilots show no measurable result, out of an estimated $30 to $40 billion spent ^[6]. PNC's chief executive did the buyer's arithmetic out loud.

Any impact that AI can have on the productivity of a bank, that productivity can be taken away by the cost of tokens. — Bill Demchak

Then the trap, in one line. Token prices — what a lab charges for each sliver of text a model reads or writes — halved in five weeks this fall, and corporate AI spending still stalled ^[7]. Cheaper intelligence did not unlock more buying. It devalued the thing its buyers couldn't turn into profit. Even the buildout's own arithmetic is underwater — Morgan Stanley reckons a fully loaded data center rents for $25 billion a year against $23 billion of output ^[8]. The labs know the ledger, and their own words carry it. AWS announced its billion-dollar consulting corps with an explanation.

The AI skills gap affects every customer and every partner, so it is a market problem, and Anthropic is committed to helping close it. — Anthropic

That is a remediation product — the seller sending people in to finish what the software could not, a fix aimed at the customers' failure to convert spending into results ^[2]. OpenAI's stated law of the buildout is blunter.

Intelligence has to keep getting cheaper. — OpenAI

Cheap is the diagnosis, not the pitch. And the human who gets dropped into a customer to make the model pay — the forward-deployed engineer — has gone from under 200 job postings in 2023 to 7,300 in the first nine months of this year, at a median salary of $195,456 ^[2]. The industry's fix for its software is people. When intelligence becomes a deflating commodity, the growth stops coming from the product and starts coming from what surrounds it: the hand-holding, the attention, the raw compute. Where value does appear, it is moved, not made. Klarna's $60 million in savings came from replacing 853 customer-service jobs with a single agent ^[9]. Meta's Muse agent, which cancels subscriptions and negotiates bills, knocked shares of Booking.com, Airbnb and Planet Fitness on launch day — the consumer's win is subtracted from someone else's revenue ^[10]. Which is why the incumbents are walling agents off their data ^[11]. There is a counter-signal, and it deserves its clause: application vendors like Klaviyo are growing fast on agent software, targeting 30%. But the 30% is the seller's growth — money landing in the P&L of whoever sells the agents, one layer above the buyers whose payoff Apollo could not find ^[12][5]. Apollo's economist, Torsten Slok, has put the whole buildout's question in one sentence.

The question looking ahead is whether, even as models become commoditized, mainstream corporate America can turn increasingly cheap intelligence into higher revenue growth and higher profit margins. — Torsten Slok

The money has already voted, in the ad slots, the consulting corps and the rented chips, on whether cheap intelligence can ever show up in a non-tech company's margin. The one number that has not come in is the one that matters.