COMPUTE
Unified memory is redrawing the local inference budget
Large unified memory pools change which models a single workstation can hold. That expands the local option for small organisations, but capacity still has to be tested against speed, power, and the real workload.
What is verified
Apple offers Mac Studio configurations with up to 512 GB of unified memory. In its launch material, the company says that capacity can hold language models with more than 600 billion parameters entirely in memory. The important part is not the vendor superlative. It is the shape of the machine: one large memory pool available to the CPU and GPU in a desk-side system.
That changes the first constraint in local inference. Before tokens per second, the model and its working state have to fit. A larger shared pool makes model classes available on one machine that otherwise require several discrete GPUs or a rented server.
What it does not prove
Holding a model is not the same as running it well. Quantisation level, context length, memory bandwidth, prompt processing, power draw, and runtime support all change the result. A large model that generates too slowly for the workflow is still the wrong deployment, however impressive the capacity number looks.
Purchase price also matters. Local compute exchanges a variable service bill for capital cost, maintenance, and utilisation risk. The comparison only becomes useful when those costs are measured per unit of finished work.
Our next test
The $10K Sovereign AI Lab will publish the exact configuration, model format, context, power measurement, and workload before publishing a conclusion. The signal is the expanded option set. The lab has to establish where that option is actually economical.