Who actually owns this subfield? Across the portfolio, the 2024 mixture-of-experts record has a clear volume leader: Intel, which by the assignee facet holds a striking share of MoE-related filings, with Microsoft, NVIDIA, and Baidu appearing as the next named players over a smaller unassigned bucket than you see in newer subfields. The structural read: MoE IP is consolidating fast, and it's consolidating around the companies that have to run these models efficiently on hardware.
A representative exemplar is Microsoft's US20240273346A1, "Self-Balancing Mixture of Experts," with CPC G06N 3/098, G06N 3/042, and G06N 3/063. Its title points at the load-balancing problem — keeping experts evenly utilized — which is the recurring engineering theme across the 2024 MoE cluster, alongside gating and routing.
The CPC clustering reveals where the contest is. MoE records concentrate in G06N 3/045/0455 (architectures), G06N 3/098 (distributed training), and notably G06T 1/20 (GPU compute) and G06F 9/30xx (instruction-level execution). That hardware-proximate footprint explains why Intel and NVIDIA loom large: a big part of the MoE problem is making sparse expert routing run well on real silicon, which is their home turf.
The caveats are sharp here. Intel's lead is by filing volume, much of it applications rather than grants, so it reflects investment and intent more than settled, enforceable position — application counts are not granted rights. And the counts are filing-date sensitive: MoE filings spiked in 2023-24, so the landscape is still actively forming and any leader board will shift as prosecution resolves.
For the strategist: treat MoE as a fast-consolidating, hardware-flavored subfield where the silicon vendors are well represented, and anchor any ownership claim to specific records like US20240273346A1 rather than to the volume facet, which overstates settled position.