When AI gets expensive to run, guess who pays

In the first days of November 2026, people running some of the biggest AI services said the quiet part out loud. According to Fortune (opens in a new tab), Elon Musk said the running cost of Grok Bot is "unsustainably high relative to revenue." The head of ChatGPT's consumer business described OpenAI's new Dots feature as "on the compute heavy side." And Anthropic, in an updated filing for its public listing, told investors that not having enough computing power is a risk to its business.

It's easy to read all of this as industry news. At Reclassify AI, we are not convinced that's the right reading. It's a billing notice. It just won't always arrive looking like a price.

What each of them actually said

It's worth being precise, because the three statements aren't the same.

Musk was the only one who talked about cost against revenue. Fortune reported that he also suggested the company was throttling back the compute allocation, which means users hit their limits sooner. Fortune doesn't link to his original statement, so treat this as Musk's words as Fortune reported them.

OpenAI's comments were about one product. Adam Fry, who heads ChatGPT's consumer business, said Dots is "on the compute heavy side." It's limited to pricier plans: Fortune names the $100-a-month Pro plan, and OpenAI's own page lists Pro, Business Premium and an Enterprise beta. On November 3, OpenAI said (opens in a new tab) Dots "can draw on different models for different parts of its work," now including a model called GPT-6.1 Terra. Fortune describes Terra as "smaller and uses less computing power" than GPT-6 Astra.

Anthropic's filing with the SEC (opens in a new tab) is the most sobering, because of the scale. It lists about $895.4 billion in commitments, "substantially all of which is related to compute and infrastructure." It names failure to "secure adequate compute" as a key risk and warns of possible "limitations on customer and user access." It also notes that short-term compute "currently costs significantly more."

To be fair, the same filing says about $213.8 billion of those obligations carry termination rights or contingent conditions. Even so, a company with commitments on that scale still lists running short as a risk. That's a supply problem, and no customer is going to negotiate it away.

How a vendor's cost becomes your problem

None of this needs a press release to reach you. When compute is tight and expensive, a vendor has a handful of levers, and every one of them lands on the customer.

Limits get tighter. The work your team finished comfortably last month now hits a ceiling halfway through.

Features move up the price list. The capability you were counting on lives on the plan above yours.

Work gets sent somewhere cheaper. Some of your tasks are routed to a smaller model. That may be perfectly fine. You may also never know it happened.

Each of these is a reasonable business decision for the vendor. That's the point. They're the vendor's decisions, made for the vendor's reasons, on the vendor's schedule.

Regulated work has to be repeatable, explainable and on time. "Our vendor ran short of capacity this week" is not a sentence any firm wants to say to a client, and it's an even worse one to say to a regulator.

Questions to ask before any AI purchase

None of this means firms should stop buying AI services. It means asking a few more questions before signing, and writing the answers down.

1. Where does the compute behind this service actually run?
2. Who decides when it gets rationed, and how will we find out?
3. What happens to our work, mid-task, when a limit is hit?
4. Is the price tied to usage, and who controls the meter?
5. Where does our data physically sit while the AI works on it?
6. Can the model doing our work be changed, and will we be told when it is?
7. What does the contract say about limits, price changes and model changes?

If a vendor can't answer these clearly, that's an answer too.

Notice that none of these questions assume a particular kind of setup. A public cloud service, a dedicated deployment, your own environment: each has tradeoffs, and the right choice depends on the work, the data and the risk. What matters is that people at the firm make that choice deliberately, with the answers in hand, rather than inheriting it from whatever was easiest to sign up for.

Decide on purpose

The vendors have been unusually candid. Compute is scarce, it's expensive, and when there isn't enough, somebody gets less. I'd much rather a firm had that conversation during a contract negotiation than in the middle of a quarter-end close.

People should decide where their work runs. Do it on purpose, before a vendor's finance team decides for you.

Every AI service runs on someone's computers. Know whose, and know who gets to decide when there isn't enough to go around.