For most of the last decade, the answer to "where does our compute live" has been a procurement decision. You picked a hyperscaler. You signed a contract. You got an account manager. The thing your data sat on was somebody else's building, in somebody else's region, on somebody else's terms, and the only operational question you were ever asked was how much of it you wanted to rent.
Quietly, that is changing. NVIDIA shipped a desktop unit this spring — a single appliance that runs frontier-scale models in the corner of a room — and is shipping a sibling box that handles trillion-parameter workloads without a cloud connection. They are calling them personal AI supercomputers. The marketing is aimed at developers. The implication is for operators.
The implication is this. The question is no longer where to rent. The question is what to keep.
Most mid-market operations we work with have not had to answer that second question in ten years. They have a colocation cabinet they have forgotten about, a server in a closet that runs payroll, and a long bill from a hyperscaler that nobody on the executive team can fully decompose. They have outsourced the placement decision because, until recently, there was no real placement decision to make. Anything serious had to live in a hall somewhere.
When a meaningful unit of frontier compute fits on a desk, the placement decision returns. It returns first for the workloads that never wanted to leave the building — the records legal will not let go of, the dispatch logic that cannot tolerate a vendor's latency, the model weights an operator does not want to be a tenant on. It returns next for everything else, more slowly.
Most boards will read this as a hardware story and look away. The operators who read it as a placement story will be the ones who, three years from now, can tell you exactly which of their workloads sit where, and why.
That is the work.