Cloud AI is a tenancy
Three documented cases from 2026 in which AI providers unilaterally retired models, shifted costs and rerouted data — and what follows from that.
If you run your AI in the cloud, you are a tenant. Only without any of the protections a tenant normally enjoys.
You build your business on infrastructure whose rules the landlord can change at any time. Models disappear, costs rise through the back door, data takes a different route. This is not speculation. Here are three episodes from spring 2026, each documented by the provider itself.
A model disappears
On 29 January 2026, OpenAI announced it would remove GPT-4o, GPT-4.1, GPT-4.1 mini and o4-mini from ChatGPT. The shutdown date was 13 February. Fifteen days.
In fairness: that concerned the ChatGPT interface, not the API. For the API, OpenAI states longer notice periods, as a rule six months for generally available models. So anyone building against the interface was not facing a fifteen-day deadline.
The point stands regardless. GPT-4o had already been announced for retirement once before and brought back after users protested. Sam Altman had promised at the time that there would be plenty of notice before a final shutdown. Whether fifteen days counts as plenty is not for the customer to decide, but for the provider. That is precisely the structure of a tenancy: the notice period is written into the landlord’s contract.
And for everyone working in the ChatGPT interface it meant something concrete: workflows that had settled around a particular model’s behaviour had to be rebuilt in two weeks.
The price stays, the invoice grows
Anthropic released Claude Opus 4.7 on 16 April 2026. The list price was unchanged at 5 dollars per million input tokens and 25 dollars per million output tokens, identical to its predecessor.
Except that a million tokens is no longer the same amount of text. Opus 4.7 ships with a new tokenizer that splits the same text into more tokens. Anthropic says so in its own migration documentation: the new tokenizer may consume up to 35 per cent more tokens for the same text. Independent measurements across more than a million real requests came out at 32 to 45 per cent depending on content, with code and structured data at the upper end.
The result: same request, same workflow, same rate card, a noticeably larger invoice.
I would not call this a trick. It is in the documentation, and anyone who reads it will find it. But the announcement led with the unchanged price, and the tokenizer change sat in the migration guide. Anyone who read only the headline planned their budget wrong. The rate is a price per unit. The unit got smaller.
The data takes a different route
On 17 April 2026, Microsoft introduced what it calls Flex Routing for Microsoft 365 Copilot. When European data centres are under heavy load, requests may be processed outside the EU — specifically in the United States, Canada or Australia.
Two details matter. First, it is an opt-out: do nothing and you get it. For some customers, such as larger organisations and the public sector, Microsoft says it set the default differently, so the starting position varies with how the tenant is classified. Second, turning it off takes time: the setting change takes about a week to propagate through the tenant.
The data continues to reside in the EU. It may be processed outside it. For anyone who has to demonstrate under GDPR, NIS2 or DORA where personal data was processed, that is a difference that ends up in the documentation, not in a footnote.
The pattern
Three providers, three entirely different mechanics, the same pattern: the provider changes the terms, the customer adapts or rebuilds.
This is not malicious. It is the logic of an industry that has invested enormous sums and has to refinance them. Keeping models available costs money. Capacity in Europe costs more than elsewhere. It is foreseeable that this pressure will grow rather than ease, precisely because switching gets harder for customers with every year of more deeply integrated use.
So the point is not that these providers do poor work. Their work is technically impressive. The point is that decisions about availability, cost and place of processing do not sit with the party carrying the business risk.
What that means in practice
Not that cloud AI is wrong. For peak load, for experiments, for anything that doesn’t touch data worth protecting, it is often the most sensible choice. The question is what you build into that tenancy.
Three things can be settled before the next announcement lands:
- Which processes come to a standstill if a particular model is no longer available tomorrow?
- What does your current operation cost if the effective cost per request rises by thirty per cent?
- Can you demonstrate, for any given period, where your requests were processed?
Anyone who can answer all three is making a deliberate decision. Anyone who cannot has not made a decision at all — they have made an assumption.
The practical conclusion is rarely “run everything yourself”. It is usually this: identify the processes that cannot absorb an outage or a change, and make exactly those independent. The rest can happily stay rented.
You should just know that you are renting.
This piece goes back to a series of LinkedIn posts from spring 2026. Information current as of August 2026.