The Flat Spend Problem: AI Usage Outruns the Business Model

Tue, 10 Nov 2026 15:00:00 +0000

Key points

Corporate token spend grew 1.5 per cent across August and September, down from 42 per cent in July, according to Ramp's card data.
Google's AI Mode produced the fastest query growth in five years while advertising revenue grew 12 per cent, the slowest rate in five quarters.
OpenAI is reported to have raised $11 billion at a $1.25 trillion valuation, and Fal doubled its valuation to $16 billion.
Microsoft is cutting its neocloud suppliers, and Crusoe sold Google an unfinished Wyoming campus.

Ramp’s corporate card data recorded something quietly strange this autumn. Token spending across its business customers grew 1.5 per cent in August and September combined, having grown 42 per cent in July, according to TechCrunch’s account of the figures. Usage did not fall; volumes climbed.

That gap is the week’s story, and it is not a story about demand cooling. It is about the distance between how much AI people use and how much anyone has worked out how to charge for it. The evidence kept pointing the same way, and the markets answered by concentrating their bets.

Queries up, revenue sideways

Forrester’s analysis of Big Tech earnings found that Google’s AI Mode drove the fastest year-on-year query growth in five years. Advertising revenue over the same period grew 12 per cent, the slowest rate in five quarters.

Twelve per cent is not a crisis. Read on its own, that is a healthy quarter. The problem is the pairing. The fastest query growth in five years produced the weakest advertising growth in five quarters, so the link between attention and income has loosened. An AI answer can settle a question without a results page, and without the advertising that used to sit beside one. Alphabet’s disclosure that AI summaries reduce advertising revenue put a number on that, and the market took more than $500 billion off its value in a day.

Reddit’s chief executive said Google referrals fell by more than a million visits a week in the third quarter. A system that answers directly serves the person asking and shortchanges whoever wrote the material.

When cheaper tokens look like a stall

The Ramp numbers need care, because the obvious reading is wrong. A near-flat spend line against rising usage does not mean firms lost interest. Ramp’s lead economist, Ara Kharazian, attributes it to price competition rather than falling demand, and the October rebound fits that better than a retreat.

Two years of cuts have made inference cheap enough that a modest budget buys work which would have been unaffordable in 2024. Buyers capture the saving; sellers absorb it. A provider can serve far more tokens and still disappoint its lenders.

Price cuts buy market share, and market share is worth having only if it converts into pricing power. That conversion keeps being postponed. The same tension showed up in the payments data on AI adoption, where spending ran ahead of the surveys.

Capital concentrates as spend flattens

If revenue is uncertain, the rational response is to own a piece of the layer that is not. The week’s funding news looks like exactly that.

OpenAI is reported to have raised about $11 billion at a $1.25 trillion valuation, with the total possibly reaching $20 billion, as TechCrunch reported the Bloomberg sourcing. The same week, inference provider Fal doubled its valuation to $16 billion in a round led by Bessemer. Money of that size is not a bet on a smooth revenue curve. It is a bet that a few positions become unassailable before the curve settles.

The infrastructure side moved the other way. CNBC reported that Microsoft is reducing its neocloud suppliers, concentrating on a few partners while expanding its own capacity. Crusoe sold an unfinished Wyoming campus to Google, a site Google had been courted to occupy as a customer rather than to buy, as our write-up noted. Buyers at this scale want terms set by contract.

The counter-position worth watching

Against all of that, Anthropic published a statement that Claude will carry no advertising, on the grounds that sponsored content inside a thinking assistant would undermine trust. That is a positioning choice as much as an ethical one, and it stays credible while subscriptions cover the difference.

Search advertising was a good business for twenty years because attention could be sold beside an answer. If the answer becomes the product, something else has to be sold.

What this means going forward

Inference prices will keep falling until weaker providers exit or are bought, which is what the Fal round finances. Consolidation among compute suppliers will continue, because buyers at scale prefer few relationships. The pressure will land hardest on anyone supplying the raw material for an AI answer without capturing the value, which describes most publishers and many universities.

The honest uncertainty is whether cheap inference is a phase of a price war or the market’s permanent shape. If permanent, companies raising at trillion-dollar valuations are buying into a utility, and utilities are low-margin and durable. If temporary, the survivors charge what the service costs.

Institutions have a more useful question to ask than which model is best this month: which suppliers will still exist, at what price, when this capital has to show a return. Usage has outrun the business model. The reckoning is one of timing, not direction.

Frequently asked questions

What does the Ramp data show about AI spending?

Token spend grew 1.5 per cent across August and September, against 42 per cent in July, then 8.9 per cent in October. It is a proxy for enterprise budgets, not an audited total.

Why can AI usage rise while spending stays flat?

Price per token falls faster than volume rises. Providers pass the savings to buyers, so a company runs more queries for a similar bill.

Is Google’s advertising business shrinking?

No. Revenue grew 12 per cent, healthy in isolation, but query growth was the fastest in five years.

Why does the AI Mode figure matter for publishers?

An AI answer can satisfy a search without a click, so the cost falls on the sites supplying the material. Reddit reported referrals down by more than a million visits a week.

What is Anthropic’s position on advertising in Claude?

Claude will carry none, on the argument that sponsored content inside a thinking assistant would compromise trust.