The Median US Business Now Spends $188 per Employee a Year on AI. Its Margins Haven't Moved.

The AI argument has been fought in the wrong set of books for three years. The bulls point at Nvidia's revenue, at hyperscaler capex, at token volumes climbing. The bears point at valuations and at credit default swaps. Both sides are arguing about the supply side, or about what investors believe. Almost nobody argues about the customer's profit and loss statement, which is the only place a general purpose technology eventually has to pay for itself.

Torsten Slok, chief economist at Apollo Global Management, ran that version on Saturday. His conclusion, in his own words: "Despite massive investments and widespread hype around artificial intelligence, the technology has delivered zero financial payoff to corporate America outside of tech."

The note isolates what Slok calls the S&P 493, the index with Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia and Tesla taken out. Revenue per worker is climbing for those seven. It is flat for the other 493. Profit margins are climbing for the seven. Flat for the 493. And on the third chart, built from the San Francisco Fed's utilization-adjusted total factor productivity series, there is still no sign that AI is lifting productivity at all.

That is the shape of the argument. Underneath it sits the spending data, which makes his case sharper than his prose does.

What the buyers are actually spending

Two of Apollo's charts lean on the Ramp AI Index, which tracks card and bill-pay transactions across more than 85,000 American businesses. The headline number is the one everyone quotes: 55.9% of US businesses on Ramp paid for AI in September, up 12.1 points from a year earlier. That sounds like a stampede.

Then you get to the spend. Among firms that pay for AI at all, the median spent $15.70 per employee in September. Per month. At the blended price Ramp recorded for the week ending 11 October, 64 cents per million tokens, that is roughly 25 million tokens per employee per month. Call it $188 a year per head.

The top 1% of those firms spent a median of $9,567.53 per employee per month. That is a gap of about 609 to 1 between the median AI buyer and the heaviest one. Ara Kharazian, Ramp's lead economist, puts the concentration more bluntly: 1% of firms drive 80% of combined OpenAI and Anthropic spending.

So "half of American business pays for AI" and "half of American business barely pays for AI" turn out to be the same sentence.

The trend line is worse for the people selling it. Token spending was flat through August and September even as volume kept climbing, because the blended price per million tokens fell 25.4% to $0.64 between the weeks ending 13 September and 11 October. Volume up 52.4%, spend up 23% over the same comparison. AI's share of total business spend on Ramp, excluding payroll, slipped to 5.77% from 5.96%.

Kharazian reads that as paid adoption hitting the top of its S curve, with growth from here having to come from existing customers using more rather than new ones signing up. Slok reads it as a demand curve that is flat and highly price sensitive, and he said so in a note on 20 October: Silicon Valley's premise of unlimited compute demand only holds while compute is effectively free.

Three notes in five days

Slok has been making a version of this argument since the summer, and Apollo published a similar note on 9 August with a line that has aged well: "The AI capex boom is so far only showing up in the sellers' margins, not the buyers'." Saturday's note is a culmination, not a discovery.

What changed last week is that he stopped arguing about margins in the abstract and started arguing about cash.

On Tuesday he pointed at declining return on invested capital across the five largest hyperscalers, Amazon, Google, Meta, Microsoft and Oracle, as invested capital grew faster than the earnings it generated. On Thursday he raised circularity, arguing that most of the heaviest AI spenders are venture backed and are therefore paying token bills with investors' money rather than with earnings. Fortune reported his figure as 80% of the top 1% of AI spenders. On Saturday he closed the loop on the customer side.

The line from Tuesday is the one that will get quoted for a while: "The core economic flaw in the AI bull case is that the marginal cost of producing compute exceeds the marginal benefit it delivers to users, raising questions about claims of unlimited demand for data centers."

The bill arrives before the revenue

That argument would be an academic curiosity if the buildout were being paid for out of pocket. It isn't.

The five big hyperscalers issued roughly $223 billion of bonds between January and late September, according to CPRAM, about $117 billion more than they raised in all of 2025. More than 40% of this year's issuance came at maturities of 20 years or longer, against 15% for the US investment grade market as a whole. Hyperscaler capital spending is running at around $800 billion this year, per Reuters, with next year's plans reported above $1 trillion.

You do not borrow at 20 years against a demand curve nobody has validated. Or rather, you do, and then the credit market starts asking questions. Meta's five-year credit default swaps widened 7.6 basis points to a record 114.79 on Thursday, according to Fortune. CoreWeave, the neocloud most exposed to the same trade, is down more than 30% this month with its five-year swaps at 1,145 basis points.

Mark Habib, S&P's global head of digital infrastructure, described the disclosure problem at a roundtable in London, as Reuters reported it: "It is safe to assume that there is more debt out there that we are not seeing."

Meta is the clearest test case

If you want one company that shows both halves of this, it is Meta. It reported on 21 October and the top line was excellent: revenue of $64.60 billion, up 26% year over year.

The rest of the statement was harder to read. Costs and expenses rose 68% to $51.61 billion. Income from operations fell 37% to $12.99 billion. Operating margin went from 40% to 20%. Capital spending including finance lease payments was $34.91 billion against operating cash flow of $32.79 billion, so free cash flow came in at negative $2.12 billion. The company narrowed its 2026 capex range by lifting the floor to $135 billion from $130 billion and leaving the ceiling at $145 billion, which is an increase dressed up as a tightening.

Mark Zuckerberg's line was: "Our investments are paying off, the core business is strong, and we're seeing velocity across the board." Meta says AI is already contributing incremental advertising revenue through stronger engagement, better recommendations and improved monetization. It has not put a dollar figure on that contribution.

That is the whole problem in a single filing. Meta can point at a mechanism. It cannot point at a number. And Meta is the seller here, with the best data in the industry, a clear causal chain from model to ad ranking, and 26% revenue growth to absorb the spending. The 493 have none of those things. A regional distributor or a contract manufacturer buying copilots and agents has no way to attribute a margin point to a model, which is why the honest answer to "what did we get for the tokens" is so often a shrug. WebProNews, summarizing EY-Parthenon's CEO Outlook, noted that half of UK chief executives rank AI tools among their top three productivity drivers while only 14% had full real-time visibility of what those tools cost.

The fair counterargument

Slok is not obviously right, and the note is more fragile than most of the coverage suggests.

His charts are aggregates. They show that margins outside tech are flat, not that AI failed to move them. As TokenPost noted in its write-up, the comparison does not isolate AI's effect at individual companies, and it does not establish that AI caused the gap. A firm could be getting real gains from AI while its competitor gets none, and the sector median would sit exactly where it sits.

There is also a serious measurement literature arguing that this is simply what a general purpose technology looks like early on. Jim Tebrake and Erich Stassner's October paper on the limits of productivity measurement in national accounts points out that micro level studies keep finding productivity improvements of 14% to 56% across customer support, professional writing, software development, consulting and legal analysis, while the aggregate statistics show almost nothing. The ILO made a related point in April and called it the aggregation paradox: task level gains of 10% to 70%, firm level results mixed, and no clear sectoral or macroeconomic signal yet. A St. Louis Fed study of nearly 490,000 earnings calls, reported by Fortune in July, found no measurable aggregate bump, and one of its authors raised the possibility that AI is producing real gains that stay structurally invisible because it destroys the value of the thing it makes abundant.

Both things can be true. Gains can be real at the task level and absent at the margin, because a technology that makes a task 40% cheaper does not help a company whose actual constraint is a customer relationship, a regulatory filing, or a machine that takes eighteen months to build.

Which is roughly where Slok lands, and where his argument is strongest. His claim is not that AI doesn't work. It is that intelligence got cheap faster than judgment did, and that the gap between the two is where the margin lives.

What this does to the demand signal

The practical consequence gets lost in the bubble argument.

The buildout has been planned on the assumption that demand is a technology question. Better models produce more usage, and more usage justifies more capacity. Slok's series makes it a procurement question instead. The marginal buyer is a finance director at a company with a 6% operating margin, staring at a line item that grew while the margin did not.

That matters for anyone selling hardware into this. Enterprise AI budgets and industrial automation budgets come out of the same pool, and the pool is set by the same CFO. If the 493 pull back on tokens, the demand shock does not land on Nvidia first. It lands two or three steps down the chain: the server refresh that gets deferred, the edge deployment that stays a pilot, the capacity that was booked for 2027 on the strength of a business case nobody has finished writing.

The supply side is already pricing some of that in. Meta's record credit default swaps and CoreWeave's 1,145 basis points are the bond market doing arithmetic the equity market has been avoiding. If next year's trillion dollars of capex has to be funded at those spreads, the cost of capacity rises for everyone, including the companies without a hyperscaler's balance sheet.

Wednesday is the test. Microsoft and Alphabet report after the close, Amazon on Thursday, and the question on all three calls is the one Apollo has been asking in public for four months. Not how much are you spending, but what is it earning. Reuters reports that S&P 500 earnings growth is on pace for more than 32% this quarter, so the headline numbers will probably be fine. Guidance on capex, and any first real disclosure of AI-attributable revenue, will matter more.

The uncomfortable possibility, for everyone in that supply chain, is that the answer takes years and the borrowing was priced for quarters.

Hardware companies are living with this on both sides of the ledger. Demand from enterprise buyers now hangs on AI budgets that are under CFO review, while capacity commitments for 2027 and 2028 were booked against a business case that is still unproven. At DMC we work with hardware companies on exactly that mismatch, from sourcing strategy and cost modeling to production ramp planning when the forecast moves underneath you. If your 2027 plan assumes a demand curve nobody has validated yet, let's talk.