cloud

In depth

The decisive argument

For AI, the decisive argument is not price but access. The best-performing models are not distributed: they can be used only through the service of the vendor that makes them, and they cannot therefore be brought back in house. To this is added the question of hardware: the accelerators required are expensive, scarce and depreciate quickly, which makes acquiring them risky when needs change from one quarter to the next. The cloud turns that investment into an operating expense, which is often the only way to begin.

The five questions

The questions to ask before committing come down to five points. Where is the data processed, and where is it stored, which is not always the same place. How long is it kept, and by whom can it be read. Is it used to train models, which professional offers generally exclude and consumer offers rarely do. Is the provider subject to an extraterritorial law allowing a third state to demand access. And what happens at the end of the contract, in terms of retrieval and erasure.

Lock-in

The real long-term risk is lock-in, and it is subtler than an invoice. It does not lie in the data, which can be exported, but in what has been built around it: prompts tuned to the behaviour of one precise model, evals calibrated on it, tool chains particular to one provider. Changing then becomes a project, not a decision. The remedy is known: maintain a layer of abstraction over the calls, keep your test sets independent, and check from time to time that a competing model passes those tests.