Microsoft Unveils Maia 300 AI Chip: 40% Inference Gain, Gigawatt-Scale by 2027

After Google's TPU, another trillion-dollar cloud giant is putting custom silicon front and center. Microsoft has officially unveiled its in-house AI chip Maia 300 and committed it to mass production: targeting a 40% inference efficiency gain and multi-gigawatt Azure deployment starting 2027, with one explicit goal — reducing dependence on, and procurement costs of, Nvidia GPUs.

From Pilot to Gigawatts

This is not Microsoft's first custom chip, but likely its biggest bet ever — described by observers as the company's largest custom-silicon wager. Maia 100 already serves Copilot-type workloads at small scale; Maia 300 is built for volume. According to multiple reports, Microsoft plans large-scale deployment from next year, with total capacity measured in gigawatts — implying on the order of a million chips at typical accelerator power draws.

Why Inference, Why ASIC

Microsoft anchoring its custom chip in inference rather than training follows clear logic: training demands peak performance where Nvidia remains dominant, while inference dominates AI operating costs and is acutely sensitive to efficiency and cost-per-token — exactly where ASICs shine. From Google TPUs serving Gemini inference at scale, to AWS Trainium, to non-GPU accelerators outselling GPUs in China, "purpose-built silicon eating inference" is a global trend.

From Google TPU to Microsoft Maia, the AI inference era has ignited an ASIC compute expansion spree.

Implications for Compute Supply and Demand

In the short run Nvidia remains irreplaceable for frontier training. But as leading cloud operators migrate inference workloads to in-house ASICs at scale, GPU demand will tilt from "training plus inference" toward "training-led" — a real challenge to Nvidia's valuation narrative and product strategy, and part of why it is opening Vera Rubin rack designs and standardizing power and liquid cooling for system-level competition.

For enterprise users, hyperscaler custom silicon means inference costs keep falling, and multi-cloud, multi-chip scheduling becomes a foundational AI capability. NineZenith applies the same principle in its full-domain AI platforms: decoupling the compute layer so applications and model services never lock into a single chip ecosystem — letting clients continuously capture the cost dividends of compute competition.

(Synthesized from public coverage by Cailianshe, Wallstreetcn, Sina Finance, EET-China, MyDrivers, Phoenix News and others)