Microsoft unveiled Maia 300 on October 12, pitching the custom accelerator at 40% more tokens per megawatt than leading GPU-based systems and committing to make the Azure Maia family the primary inference substrate for Microsoft products and services beginning in 2027. The announcement, posted by EVP of Cloud + AI Scott Guthrie, pairs a scheduled multi-gigawatt rollout with Reuters reporting that Microsoft plans to field "potentially more than a million" of the chips next year.

The efficiency number does the political work. Microsoft didn't identify what it benchmarked against, but SiliconANGLE read the 40% figure as a reference to Nvidia, and the comparison tables Microsoft provided make the subtext obvious.

Specs land in that register. Each accelerator pairs four compute chiplets on TSMC's 2-nanometer N2P node with auxiliary chiplets on cheaper 3-nanometer tech, flanked by eight HBM4 stacks totaling 288 GB. That matches Nvidia Rubin's HBM capacity with roughly 20% less memory bandwidth, which Microsoft's own figures put at 18.0 TB/s peak and 950 watts TDP per unit. Racks hold 96 accelerators inside a 167-kilowatt envelope, against Nvidia's GB300 racks running 72 GPUs at 142 kilowatts. Superpods group 12,288 accelerators across a custom Ethernet-based transport. At rack scale, Microsoft claims a 7x token-throughput improvement over Maia 200.

Maia 200 is already in production serving Microsoft Copilot, with GitHub Copilot and Microsoft Foundry scheduled to migrate by year-end. It runs OpenAI models, Microsoft's own MAI family, and open-source weights, which is the detail that matters most to anyone paying an Azure invoice. The same workloads already powering Copilot, including the inference under products like MAI-Transcribe 2's 2.5% WER streaming stack, will shift onto silicon Microsoft designed, fabricated under contract, and owns the depreciation schedule for.

Rani Borkar, president of Azure hardware systems and infrastructure, framed the pitch plainly: "It's basically delivering more intelligence for every watt, every dollar, and every byte of memory by optimizing across the entire system." That's the vocabulary of a hyperscaler that has decided renting compute is a losing position in a market where the frontier capability bar keeps rising, as Gemini 4 Argon's AutomationBench run made evident last week.

The precedent is Amazon's Graviton arc, which took roughly a decade to translate custom silicon into visibly cheaper EC2 pricing and reshape the AWS margin profile. Microsoft is attempting the same maneuver on a compressed timeline, with a workload mix that didn't exist when Graviton shipped. If the 2027 production ramp hits, downstream Azure OpenAI, Copilot, and Foundry pricing is the dependent variable.

A chip that doesn't ship in volume is a slide. The slide now has a date.

Microsoft's Maia 300 Targets 40% More Tokens Per Megawatt, Named Primary Inference System for Copilot in 2027

Sources