In October I watched a chipmaker launch a new AI laptop in Berlin, the fruit of four years with Microsoft. One petaflop, 128 GB. Then I read what happens when you put Linux on it. Plugged in, a 120 billion parameter open model runs at 13.9 tokens per second. On Windows, 65.4. The firmware holds an encrypted power table, and only the Windows driver has the key.
That bothers me. We tell clients that local means control: your hardware, your model, your data. Here you own the machine, and the vendor still decides how fast your open model runs, and on which system. It is one test, on laptops, not the datacentre line. The question carries over anyway. Before we call a box sovereign, we ask who holds the keys to its firmware. If it is not us, we rent the speed.