Here's what published — published is not granted. Application US20240169463A1, "Mixture-of-Experts Layer With Dynamic Gating," published May 23, 2024, assigned to Microsoft Technology Licensing, LLC, with a large inventor list including Changho Hwang and Yongqiang Xiong. The CPC codes are G06T 1/20 (GPU/graphics compute) and G06N 5/043 (inference engines).
The mechanism is the gate. A mixture-of-experts model contains many expert sub-networks but routes each input token to only a few of them, so the model has huge capacity while activating only a fraction per token. The gating network is what decides which experts handle which tokens. "Dynamic gating" points at making that routing decision adaptively at runtime — adjusting how tokens are dispatched rather than using a fixed scheme — which affects both quality and the efficiency of how work spreads across hardware.
MoE is one of the hottest AI subfields, and the gating mechanism is its crux: it determines load balance, throughput, and whether the sparse-activation promise actually pays off on real hardware. The G06T 1/20 code signals the claims reach toward GPU execution, which is consistent with Microsoft's interest in serving large MoE models efficiently.
Because this is a publication, the framing is intent. The claims as filed describe what Microsoft seeks; until a grant issues with allowed claims, there is nothing enforceable, and the scope is undetermined. Calling this "Microsoft's MoE patent" misstates both the status and the breadth.
The takeaway: US20240169463A1 is a published marker in the MoE routing race — dynamic gating, tied to GPU execution — and a reminder that this is a fast-moving, crowded subfield where the published-versus-granted distinction matters enormously for any ownership read.