For four years Microsoft has run the largest rental business in computing on accelerators designed by Nvidia. On Monday it told customers and rivals where it intends to stand in 2027.
Scott Guthrie, the executive vice president who leads Microsoft’s cloud group, said the company’s third-generation in-house chip, Azure Maia 300, will enter production in 2027 and become the primary AI inference system for Microsoft’s products and services. The Maia family, he said, will scale to multi-gigawatt deployments beginning that year.
The numbers attached to the announcement describe a gain in efficiency rather than in raw speed, which is the trade Microsoft has chosen. Guthrie said Maia 300 is expected to deliver 40% more tokens per megawatt than leading GPU-based systems, and that a full rack of 96 accelerators will produce up to seven times the token throughput of Maia 200, the chip Microsoft announced only in January.
That comparison matters because of where the chip will be used. Inference is the work of answering requests with a model that has already been trained, and it runs continuously across Copilot, GitHub Copilot, Microsoft Foundry and the company’s own models. Training gets attention. Inference pays the electricity bill, and Microsoft has concluded that the bill is now the constraint it can control.
The design reflects the workloads the company says it is preparing for: long-context reasoning, conversational products that cannot tolerate delay, and agentic tasks that run unattended through many steps. Maia 300 is built on TSMC’s N2P process and uses a chiplet architecture that separates the functions of compute, memory control and input-output, so that each can be made with the process best suited to it. The point of the split is to spend more silicon on computation and less on everything around it.
Microsoft’s reliance on a single supplier has been an awkward feature of its position. The company buys GPUs from Nvidia at prices set in a market where demand has exceeded supply for three years, and it has spent billions on both the chips and the data centers to hold them. A working in-house accelerator changes the terms of that relationship even if it does not end it. Guthrie’s phrasing was careful: Maia will become the primary system, not the only one.
The competitive logic runs across the industry. Google has built its own tensor processors for a decade and rents them through its cloud. Amazon designs training and inference silicon for its data centers. Meta has begun publishing details of its accelerators. Each company reached the same conclusion from the same arithmetic: at sufficient scale, a chip designed for your own workload can cost less to run than one designed for everyone’s.
The timetable is the part to watch. Microsoft said production begins in 2027, which is an intention rather than a shipment, and the company declined to say how many chips it will build or how much of its fleet they will cover. Executives said they want to keep some discretion about scale, a remark that leaves room to slow the program if the economics change or to accelerate it if they do not.
Investors read the announcement quickly. Microsoft shares rose 1.14% on the news, and the company’s market value returned above $4 trillion, a level it had slipped below in the preceding weeks. The move was modest for a company of that size, which suggests the market treated the disclosure as confirmation of a strategy it had already priced rather than as a new fact.
The strategy has been visible in the spending. Microsoft has committed to capital outlays that run into the hundreds of billions of dollars across several years, most of it for data centers and chips, and it has signed power agreements to feed them. A chip that produces 40% more output per unit of power changes the return on that spending without changing the amount committed, which is the kind of improvement that shows up in gross margin rather than in a headline.
There is a version of this that disappoints. Custom silicon has a long record of arriving late, costing more than forecast and running software poorly in the first generation, and Microsoft’s own first Maia was widely seen as a cautious step. Maia 200 does have production workloads behind it, which gives the third-generation design something to build on that its predecessor lacked.
The larger point is about dependency rather than performance. Every large cloud operator now wants the option to serve its most valuable workload on silicon it controls, because that option changes what it pays and whom it negotiates with. Microsoft has said where it wants to be in 2027. Whether it gets there depends on a chip that has not yet been built, a process node that has not yet reached volume, and a demand forecast for AI services that no one outside the company can verify.
