Microsoft has made its biggest move yet to run its artificial intelligence services on hardware of its own design. On Monday the company unveiled Maia 300, the third generation of its custom AI chip, and said it plans to deploy potentially more than a million of them starting next year in multi-gigawatt installations. Chief Executive Satya Nadella described the rollout as the company's biggest bet yet on its own silicon, and Azure executive Scott Guthrie said Maia is intended to become the primary system for running AI inference across Microsoft's products and services.
Inference — the work of running a trained model to answer a user's request — is where AI spending increasingly lives, and Maia 300 is built squarely for it. Microsoft says the chip handles coding, long-context reasoning, latency-sensitive conversations and the sustained, multi-step tasks performed by AI agents. The company claims it delivers about 40 percent more tokens per megawatt than leading GPU-based systems, though it has not named the systems used for that comparison, and the figures are Microsoft's own rather than independent benchmarks.
The design also departs from the dominant template. Where Nvidia builds its accelerators around one very large processor surrounded by high-bandwidth memory, Maia 300's central processor is assembled from chiplets — smaller pieces of silicon bonded together. Key computing functions use Taiwan Semiconductor Manufacturing Company's 2-nanometer process, while cheaper 3-nanometer chiplets handle jobs such as moving data on and off the processor. The chips will ship in racks of 96 accelerators drawing about 167 kilowatts; Microsoft says a full rack delivers up to seven times the token throughput of its predecessor, Maia 200. By comparison, Nvidia's latest GB300 racks hold 72 graphics processors and use about 142 kilowatts.
The timing shows how fast the program is moving. Maia 200 was only announced in January, and it is already running production workloads, including Microsoft's in-house models and OpenAI models delivered through Azure. Rani Borkar, who leads Azure's hardware systems and infrastructure, framed the goal as getting more intelligence out of every watt, dollar and byte of memory by optimizing the whole system rather than a single chip.
The strategic stakes are clear. Every hyperscaler is now designing custom silicon to loosen Nvidia's grip on AI computing, cut costs and secure supply — and Microsoft has said it expects to spend more than 175 billion dollars on capital expenditures this calendar year alone, a figure that helps explain why Broadcom's chief executive argued earlier Monday that capital, not chips, is the industry's binding constraint. Production deployment of Maia 300 is slated for 2027, so Nvidia and AMD hardware will keep carrying much of Microsoft's AI load in the near term, including model training. But if the efficiency claims hold up at scale, Maia 300 marks the point where Microsoft's own silicon stops being a side project and becomes the center of its AI business.