Microsoft's Maia 300 Puts AI Inference on Its Own

Microsoft wants the chips answering many future Copilot requests to bear its own design. On October 12, the company announced plans to make its Azure Maia family the primary inference system for Microsoft products and services. Maia 200 is already serving production workloads; Maia 300 is scheduled to begin production deployments in 2027, when Microsoft says it will start scaling the family toward multi-gigawatt deployments.

That is a substantial ambition, with a careful verb attached: will. Microsoft is proposing to control more of the machinery that runs trained AI models each time someone asks for an answer, writes code with an assistant or sends a task through a Copilot workflow. The announcement does not say that Maia will handle every request. Nor does the start of a multi-gigawatt rollout mean those gigawatts will be operating on day one.

The useful question is what Microsoft can move onto Maia, how quickly, and what stays on other suppliers' hardware. A chip can be excellent at a chosen job and still leave a cloud company buying an extraordinary number of other chips.

Why the Job is Inference

Training builds a model by adjusting its parameters over repeated examples. Inference puts that trained model to work. For an AI assistant, inference happens when it processes a request and generates a response. Microsoft is designing Maia 300 for long-context reasoning, conversational responses that cannot keep people waiting, and multi-step agent tasks. Those jobs make memory movement, networking and power use consequential alongside raw computing speed.

That choice fits Microsoft's immediate business. Its Maia 200 already serves Copilot and runs OpenAI, Microsoft and open-source models. Microsoft plans to add GitHub Copilot and Microsoft Foundry workloads by the end of 2026. If the company can tune hardware and software together for requests it handles repeatedly, it has a route to greater control over how those services run. Customers might eventually encounter the result as faster or more dependable features, but the announcement supplies no measured user-facing improvement from Maia 300.

Microsoft describes Maia 300 as a system spanning chips, racks, memory and networking. It specifies 96 accelerators in a rack with a 167-kilowatt power envelope and says the system is designed to keep data moving efficiently across those components. That is why a comparison of individual chips would miss much of the proposal: a response may depend on getting model data to the right place as well as performing calculations once it arrives.

The company expects Maia 300 to produce 40% more tokens per megawatt than what it calls leading GPU-based systems. It also claims that a full 96-accelerator configuration delivers up to seven times Maia 200's rack-scale token throughput. These are Microsoft's comparisons, not independently demonstrated results across customers' models and workloads. Tokens per megawatt measures output against power; it does not, by itself, tell a Copilot user how long a response will take or tell an Azure customer what a deployment will cost.

The Rollout is Part of the Test

Microsoft introduced Maia 100 in 2023 and Maia 200 in January 2026. Maia 300 is therefore the next step in a program with production experience, rather than a newly sketched chip awaiting its first workload. But a production workload and a primary infrastructure system are different scales of undertaking.

As of this summer, Maia 200 was operating in two US data centers, The Information reported in August, as relayed by GeekWire. Microsoft now says Maia 200 is deploying across the United States, Europe and Australia. The statements describe different times and stages: operating in two facilities then, deploying across regions now. They make the pace of installation and the volume of work shifted onto Maia better tests of the 2027 plan than the number printed on a prototype.

The timetable also separates two promises that are easy to collapse. Microsoft says Maia 300 will begin production deployments in 2027 and says scaling toward multi-gigawatt Maia deployments will start that year. Neither statement supplies a date by which Maia will process most Microsoft inference requests. A company can begin installing a system well before it becomes the system doing the largest share of work.

Microsoft has a strong reason to pursue that share. It designs Maia while Taiwan Semiconductor Manufacturing Co. makes the chips, rather than Microsoft buying finished Nvidia processors for those jobs. At sufficient scale, choosing more of the hardware and software together could give Microsoft more room to match capacity to its own services. That is a strategic possibility, not a quantified saving: the cost of building and running the wider system still counts.

Nvidia remains a supplier to Microsoft's data centers. Microsoft executive Alistair Speirs also said the company works with Nvidia and AMD for model training, in remarks obtained by CNBC. Those relationships do not make Maia's inference goal incoherent. Training a model and serving its answers place different demands on hardware; moving serving workloads onto Maia would not, on its own, move training there too. Microsoft has also described its broader platform as supporting a choice of models and silicon. Its word primary leaves room for other hardware when a model or workload calls for it.

The current sales model adds another boundary. Unlike Google and Amazon's rentable AI accelerators, Maia cannot presently be rented by outside companies by the hour, CNBC reported. Microsoft can still use Maia to serve products those companies or their employees buy. But internal deployment and a customer selecting Maia as a cloud chip are different tests of demand. For now, Microsoft's most direct opportunity is to improve the economics and performance of services it operates itself.

A Rival's Answer Uses Two Chips

Amazon Web Services offers a useful comparison because its reported inference plan does not ask one accelerator to do the entire job. AWS plans to pair Trainium with Cerebras hardware: Trainium handles prefill, the computation involved in processing the input, while Cerebras handles decode, generating the output as memory access becomes a constraint. A kitchen analogy only gets you so far, but this is closer to assigning preparation and service to different stations than asking one station to optimize both.

That plan illustrates a choice Microsoft may face as Maia grows: which stages and models belong on its own system, and when another accelerator is a better fit. It does not show that Microsoft will copy AWS's two-chip arrangement, or that the companies have deployed comparable amounts of hardware. It does show why a cloud provider can invest heavily in custom silicon while continuing to work with specialist suppliers.

For people buying or using AI services, the chip logo inside a rack is rarely the deciding factor. Response time, reliability, price and access to the right models are. Microsoft's announcement gives those readers a concrete set of questions for 2027: where Maia 300 enters production, which workloads move to it, and whether its promised efficiency shows up in services rather than only in a rack-level comparison. The name on the chip will matter most if Microsoft can make that move at scale.