Rent the Machines, Keep the Model

For a while the advice to enterprises was blunt: do not train your own model, you cannot afford the GPUs or the team to babysit them, just call a frontier API. That advice is quietly expiring.

A recent Microsoft write-up walks through a large telecom building its own domain-specific models, the kind that actually understand its networks, standards, and operations, at a scale of roughly a trillion tokens. The notable part is not the model; it is how they got the compute. Instead of owning and operating a GPU fleet, they leaned on managed compute, dedicated capacity without the deployment, scaling, and operational overhead that used to make this a non-starter for anyone but the labs.

That changes the build-versus-buy math in a specific way. The reason most enterprises rented a general model was never that a narrow, domain-tuned model would be worse; for their actual work it is often better. The reason was that owning the model meant owning the infrastructure, and the infrastructure was the part nobody wanted. Collapse that operational tax and the calculus flips for a whole class of regulated-industry customers who care about a model that speaks their domain and stays under their control.

I would not read this as “everyone should train models now.” Most should not. But the default answer is no longer automatically “rent a frontier model and move on.” For a global SI advising a client with deep domain data and real compliance constraints, there is a third option worth pricing out: rent the machines, keep the model, and own the thing that encodes your business.

The frontier labs will keep making the general models better. That is not the same as making them yours. The interesting enterprise question for the next year is not which model is smartest; it is which parts of your AI stack you can afford to actually own.