Custom Silicon: Microsoft's Maia 300 Claims 40% More Tokens per Megawatt Than "Leading GPU-Based Systems" — the Second First-Party Inference Chip in Three Weeks
The custom-silicon competition story this note tracks (Custom Silicon: Microsoft's Maia 300 Claims 40% More Tokens per Megawatt Than "Leading GPU-Based Systems" — the Second First-Party Inference Chip in Three Weeks — prior revisions: OpenAI's Jalapeño claiming the inference Pareto frontier against GB200/GB300, AMD's Helios ramp, Broadcom's custom franchise) escalated on October 12 when Microsoft unveiled Maia 300, its third-generation in-house accelerator — and, per GeekWire, made its "biggest bet yet" on its own silicon "to reduce reliance on Nvidia."
The claim, in Microsoft's own words: "Introducing Maia 300, Microsoft's next-generation accelerator. Building on the magic of Maia 200, it delivers 40% more inference tokens per megawatt compared to leading GPU-based systems123, and up to 7x the rack-scale token throughput of Maia 200." (Microsoft on X, via Techmeme) Scott Guthrie, Microsoft's Cloud and AI chief, repeated the figure: the Maia 300 "processes approximately 40% more tokens per megawatt of power consumed than 'leading GPU systems'" — a comparison Microsoft deliberately left unnamed. (BigGo) Deployment begins 2027, with Reuters reporting plans that "could involve" more than 1 million units next year and a multi-gigawatt footprint. (Blockspace, TechRadar)
The pattern this note is tracking. In three weeks, the two claims that inference can be done more efficiently than on Nvidia's flagship racks have come from OpenAI (Jalapeño, benchmarked against GB200/GB300 in its own appendix) and now Microsoft (Maia 300, "leading GPU-based systems" unnamed) — i.e., Nvidia's two largest customers are building the substitution path while remaining its biggest buyers4. The counterweight: both claims are vendor benchmarks against unnamed baselines, and both target inference specifically — training and frontier research demand still run through Nvidia's stack. Meanwhile the longer-dated vertical-integration threat also firmed: Texas filings point to a December 1, 2026 construction start for Terafab, the Tesla/SpaceX fab with Intel (14A process) as partner, with Musk saying his companies "will probably learn to make chips 'better than anyone'" — though Morgan Stanley sees first output no earlier than mid-2028 and the IPO filing calls Terafab "a general framework that either side can walk away from." (Admirals)
What it means for the brief. The capex story's volume risk now has a quantified efficiency counterparty: if Maia 300's 40%-per-megawatt claim survives independent benchmarks at 1M-unit scale, the same workload needs proportionally fewer accelerators — the mechanism by which capex growth and capability growth decouple. The tripwires that matter are deployment scale (watch registered) and whether any hyperscaler discloses a GPU order cut attributable to first-party silicon — the existing cancellation watch.
-
An instance of When megawatts ration the buildout, market share goes to the most tokens per watt — The customer's substitution challenge is priced entirely in tokens per megawatt against unnamed GPU baselines — share contested on watts, not raw capability. ↩︎
-
An instance of The compute moat collapses when AI workloads shift from training to agentic inference. — Nvidia's largest customer claims its custom inference silicon beats flagship GPUs on the agentic workload — the training-to-inference shift fracturing the general-purpose monopoly from inside the installed base. ↩︎
-
An instance of Custom, model-tailored silicon is breaking the general-purpose GPU monopoly. — Nvidia's largest customers are co-developing custom inference ASICs benchmarked against its own racks, building the substitution path for real-time workloads while remaining its biggest buyers. ↩︎
-
An instance of The compute moat collapses when AI workloads shift from training to agentic inference. — Microsoft's Maia 300 leads with an inference tokens-per-megawatt claim against unnamed GPU baselines — custom ASICs attacking exactly the power-sensitive, real-time reasoning side of the GPU franchise. ↩︎