The invoice for cloud AI software has a single line. The real bill has three.
The first is the one you see: the subscription. Specialised AI platforms bill per user, per month, forever — with going rates of 300 to 500 dollars a month per professional on the high-end legal platforms. For a firm of ten people that is the equivalent of an extra salary a year, growing with every hire.
The second item never appears on the invoice: the data. Every contract uploaded, every question asked, every draft generated crosses servers the firm does not control, under jurisdictions it never chose. For an ordinary company that is a risk to be weighed; for anyone bound by professional secrecy and by the federal data protection act, it is a question you must be able to answer in writing: where exactly did my clients' documents end up?
With cloud AI, the honest answer is: I don't know for certain. With a machine in the office, the answer is: here.
The third item is dependence. When the provider changes prices, terms or models, whoever is in the cloud can only accept. 2026 has shown it more than once: price lists adjusted, models withdrawn, features moved into the higher tier. Whoever has the intelligence in-house upgrades when it suits them — or does not upgrade at all, and everything keeps working.
The logic, reversed
The reason the cloud seemed unavoidable — only a datacenter can run serious models — is no longer true. A desktop machine with unified memory now runs models of hundreds of billions of parameters (we write about it here, with the numbers and the sources). Professional open-source software is mature and public. What remains is the last mile: selecting, installing, specialising, maintaining. That is our trade.
The bill, reversed, becomes: you pay once for the machine, once for the tailored installation, and an annual maintenance fee. No per-user licences. No documents in transit. No terms changing over your head.