Capex is a promise; placement is how you keep it for mixture-of-experts. Intel's application US20250086424A1 (“Deployment of resources in mixture of experts processing,” published 2025-03-13), reflecting 2024 research, goes after MoE placement. Assigned to Intel Corporation and classified CPC G06N 3/042 and 3/063, it covers how MoE resources are deployed across hardware.
The cost lever is where things sit. In a mixture-of-experts model, experts are distributed across devices, and the placement determines how much data has to move between them and how evenly the hardware is utilized. Poor placement means idle chips and expensive communication; good placement means the savings MoE promises actually materialize.
For the capex desk this is the unglamorous middle layer. Companies disclose how much hardware they buy; they don't disclose how well that hardware is orchestrated for the models they run. MoE placement IP sits in that gap — it's part of why two identical fleets can have very different effective serving costs.
The discipline: this is a published application, scope unsettled, and no filing attributes a figure to it. We model nothing off it. The point is that the placement-and-utilization layer of MoE serving was an explicit Intel research target, with dated IP behind it.
The reusable frame for payback: serving cost is compute plus communication plus utilization, and placement touches all three. Patents on deploying MoE resources efficiently are part of the software layer that decides whether the capex on accelerators turns into cheap inference or wasted silicon.