The spending data and the valuation do not agree

Two numbers crossed this week and almost nobody put them side by side.

Ramp, which sees the actual card transactions of thousands of companies, reported that AI spending is flattening — not collapsing, but flattening — right as Anthropic moves toward a listing. In the same news cycle, OpenAI reportedly raised another $11 billion at a $1.25 trillion valuation.

One number describes what companies are actually paying for. The other describes what investors expect them to pay for later. They are allowed to disagree. But if you are the person signing the renewal, the disagreement has a cost, and it lands on your side of the table.

What a plateau actually looks like

A spending plateau in software rarely looks like a decision to stop. It looks like three quieter things happening at once.

Seats get consolidated. A team buys twelve licences, uses four heavily, and renews eight. The vendor books a renewal, the buyer books a saving, and the aggregate number goes flat.

Usage shifts to the free tier or the cheap model. This is the one vendors do not talk about. Once a tool is embedded, the marginal task migrates to whatever is cheapest that still works, and the expensive plan becomes a seat for the person who needs the good model occasionally.

Pilots end without a second act. A company runs six experiments in Q1, keeps two, and the other four quietly expire. Nobody announces it. The line item just disappears from next year's budget.

None of these show up in a funding round. All of them show up in Ramp's data, because Ramp sees the card statement rather than the press release.

Why the valuation is still rational

It would be lazy to say the investors are wrong. They are pricing something different: not current spend, but the option value of being the default layer if AI usage compounds for another decade. At $1.25 trillion, you are not buying this quarter's revenue, you are buying the assumption that this quarter's revenue is a rounding error on the eventual one.

That is a coherent bet. It is also completely irrelevant to whether your team should pay 40% more per seat in March.

The mistake buyers make is treating a vendor's funding as evidence of the vendor's durability, and therefore as a reason to accept longer terms. The logic runs: they raised $11 billion, they will be around in three years, so a three-year deal is safe. But the same funding is what gives them the pricing power to make that three-year deal worse for you. Both things come from the same cheque.

The part that should worry a sourcing desk

Here is the specific mechanism, and it is not hypothetical. When spend flattens, a vendor has two ways to keep revenue growing: raise price per unit, or raise the number of units each customer consumes. The first is visible and invites churn. The second is engineered into the product — more seats required for the same work, more credits consumed per task, more surfaces that need a licence.

We have watched three tools this year move features behind a tier that used to be included. None of them raised the list price. All of them increased the cost of the same workflow by 20–35% for a mid-sized team.

That is what a plateau looks like from the buyer's chair. The headline price is stable. The bill is not.

Four things we now put in a renewal conversation

1. Ask for the last twelve months of your own usage, in the vendor's billing units. Not seats — tasks, credits, documents, whatever they meter. If they cannot produce it, you are negotiating blind and they are not.
2. Cap the unit, not the price. A 5% price cap is worth very little if the units per task can double. Ask for a contractual definition of what one unit buys, and a cap on consumption per workflow.
3. Right-size before you renegotiate. Drop seats you do not use before asking for a discount. A vendor will rarely cut your rate, but they will almost always accept a smaller contract, and the smaller contract is the real saving.
4. Put a second supplier in the room. Not a full migration — one alternative you have actually run on a real task, mentioned by name. It changes the conversation more than any negotiation tactic.

What this week actually told us

The interesting story is not that AI spending is stalling. It is that the companies selling AI know it, and are pricing for a future that requires your spend to stop stalling. Those two facts can coexist for a while. They cannot coexist forever, and when they resolve, the resolution will not be a lower valuation — it will be a tighter definition of what you get for the money.

Which means the useful thing to do this quarter is not to predict the market. It is to make sure your contract can survive either outcome: short enough to reprice, specific enough to audit, and small enough that walking away is a Tuesday, not a project.

---

We publish the plan limits and unit definitions for every tool in the library, and we flag it when a vendor changes one without changing the price. That change is usually the first honest signal you will get.