Fentner

Markets · Opinion

Your AI vendor is selling you a price its own supplier cannot hold

Whatever you make of Gioia's crash numbers, B2B vendors that built roadmaps on cheap model access are now selling flat-fee AI features they may not be able to afford to keep serving.

Ted Gioia says Oracle was at $265 when he warned of a bubble top and is now at $114, a fall of 56%. He puts CoreWeave at $42 against $138, Meta down 22% from $769, and a $2 trillion NASDAQ wipeout in October. I can't verify any of that, and I'd treat the figures as his. I'm not interested in whether he called the top. I'm interested in what the numbers would do to the people who buy and sell software.

I expected to skip most of his piece. A good half of it is a holiday wish list addressed to Spotify, Amazon and the colleges, and I have no opinion on whether Amazon should feel guilty about its book business. What held my attention was one count: of 41 data centre initiatives on the ballot this week, he says 40 went against AI. If that is accurate, the companies that rent out compute will find it slower and dearer to build, and they will also find the people who fund them less patient.

Follow the money one step down the chain. Your software vendor does not run its own models. It buys them from a lab, the lab buys capacity from landlords like the ones Gioia lists, and the landlords live on cheap capital. I think the cheap capital is what let model access be sold below what it costs to serve. A vendor that built its roadmap on that price, and then wrote it into a flat per-seat fee, holds a margin that is the gap between what you pay and a metered bill it does not control. When the supplier reprices, the vendor can eat it, pass it on to you at renewal, or quietly make the product worse.

The vendor's likeliest response is to bundle. Put the AI feature into the base price, raise the base price, and call it value. Gioia's complaint about Microsoft is that it pushes AI on customers who have not opted in. I'd put it differently, in procurement terms: bundling is how a seller defends revenue when it can no longer defend the unit economics. You end up paying for a feature your staff may not use. By Gioia's account, hostility to AI rises the more people use it, so a share of your own employees and customers may be actively resentful of the thing you are being charged for.

The best objection is that token prices keep falling and open-weight models such as DeepSeek exist, so any vendor can swap in a cheaper engine and the squeeze never arrives. I accept half of it. A vendor that built to switch models will survive a supplier price rise, and I'd pay a premium for one. But look at what the swap does to the negotiation. If your vendor can move to a cheaper model overnight, you are entitled to ask why your price has not moved with it, and the margin it was counting on goes to whoever pushes hardest. Cheaper tokens also do not shrink the bill when an agent burns many more of them per task than a chatbot reply did. That is my bet rather than a measured figure, and your vendor should be able to tell you the real number for its own product.

There is a second group exposed here, and it is the one I'd worry about most: the vendors that replaced human customer service with bots to cut cost. Gioia tells chief executives to ask customers first. I'd add that a support team that was cut to fund an AI line cannot be rehired quickly, and a vendor in that position has the least room to absorb a price rise from its own supplier.

Before your next renewal, put one question to every vendor that sells you an AI feature at a flat fee, and get the answer in writing: what happens to my price if your model supplier raises its rate, and is there a cap in the contract. A vendor with a good answer will give it to you in a day. I expect most of the large ones to move their flat-fee AI add-ons to metered pricing or drop them by the end of 2027, and the buyers who asked early will have the better terms.

Prompted by The AI Business Model Just Collapsed, honest-broker.com.