The Custom-Silicon Surge: Microsoft Says Maia Becomes Its Primary Inference Architecture — "We Will Mainly Rely on Our Own Silicon"
Update 2026-10-14: the hyperscaler-insourcing leg of the custom-silicon theme crossed a threshold. CNBC (Oct 12): "Starting with its third-generation Maia AI accelerator, unveiled on Monday, Microsoft said it will mainly rely on its own silicon to handle incoming requests.1" Microsoft is targeting "multi-gigawatt deployments" of the Maia family (Scott Guthrie blog), with Maia 300 production starting in 2027.
The specifics:
- Maia 300: 96 chips per rack (double Maia 200), up to 7x more AI tokens per rack, and 40% more tokens per megawatt "when compared with GPUs" (CNBC).
- The incumbent-generation claim from Nadella's July earnings commentary: Maia 200 "delivers 30% better performance per dollar than the latest generation hardware in our fleet" and is already handling OpenAI models (CNBC).
- Evercore ISI estimates Maia will represent about 10% of Microsoft's AI chip spending by Q1 2027 (CNBC).
- CFO Amy Hood, at a Goldman Sachs conference last month, tied own-silicon to "margin improvement over the course of a decade" (CNBC).
- The stated motive is power/cost: "You can think of power translating pretty directly into operational cost for something like a data center" (Alistair Speirs, CNBC).
Why it matters for the capex verdict: this is the first time a hyperscaler has declared its own silicon the primary architecture for inference requests rather than a supplement — the exact "reduce dependence on Nvidia GPUs" scenario that trades as a demand-leg risk for the Blackwell/Rubin order book. Three mitigants keep it second-order for now: Microsoft "partners with Nvidia and AMD" for training (Speirs); Maia accelerators can't be rented by external customers ("Nobody has tested them outside of Microsoft yet" — Patrick Moorhead, CNBC); and Evercore's own estimate has Maia at only ~10% of Microsoft's AI chip spend by Q1 2027. The same week, Broadcom's Hock Tan made the complementary supplier-side case at DBS: a custom accelerator running at 800W can beat a standard GPU drawing 1.4kW on its designed job and can "even match Nvidia's Vera Rubin" for some inference (The Financing Web, Oct 17: CoreWeave's CDS Prints a Record 1,194.5bps — the Marginal Borrower Is Now Formally in Distress, and the Stress Is Contagious to Oracle).
What it means: the inference layer is where custom silicon bites first — consistent with the OpenAI "Jalapeno" ASIC track and AMD's MI500 pipeline in the prior revision. Nvidia's training franchise and system-level moat (NVLink/CUDA) remain intact, but the marginal inference token is becoming the industry's contested terrain, and the largest Azure landlord has now publicly picked its own chip to fight for it.
-
An instance of The compute moat collapses when AI workloads shift from training to agentic inference. — The largest Azure landlord routes its marginal inference tokens to its own power-efficient ASIC — the workload shift to real-time inference where the general-purpose GPU moat fractures first. ↩︎