Microsoft Maia 300 Uses TSMC N2P as Azure Targets Gigawatts of AI Capacity in 2027

A 96 accelerator rack anchors Microsoft's next AI push

Microsoft has unveiled Maia 300, an AI accelerator whose computing chiplets will use Taiwan Semiconductor Manufacturing Co.'s N2P process, with production deployments beginning in 2027. The announcement puts a specific design behind Microsoft's ambition to make its own Maia hardware the primary system for running AI inference across its products and services.

The planned rack combines 96 accelerators within a 167 kilowatt power envelope, with 27.6 terabytes of high bandwidth memory and about 1.7 petabytes per second of aggregate memory bandwidth. Microsoft expects Maia 300 to deliver 40% more tokens per megawatt than leading systems based on graphics processing units, or GPUs, and up to seven times the output of a Maia 200 rack.

Those are company performance claims, not independent benchmark results. They nevertheless show where Microsoft is directing its investment: increasing the amount of useful AI output that can fit within a data center's electricity, cooling and space limits, rather than simply increasing the computing power of an individual chip.

Scott Guthrie, Microsoft's executive vice president for Cloud + AI, set out the deployment plan in the company's October 12, 2026, Azure Maia announcement.

We will scale our Azure Maia family to multi-gigawatt deployments starting in 2027, making Maia the primary inference system for Microsoft products and services.

The commitment concerns the Maia family, rather than Maia 300 alone. Reports that Microsoft could deploy more than one million Maia 300 chips in 2027 go beyond the published announcement, which confirms the starting year but does not specify a chip count or a deadline for completing the broader expansion.

What TSMC N2P will make

Microsoft's confirmation is more precise than a general description of Maia 300 as a 2 nanometer chip. The computing chiplets use N2P, a process within TSMC's 2 nanometer family. Memory controllers and input and output functions sit on separate dies, allowing different parts of the package to use manufacturing technologies suited to their tasks.

A chiplet is a smaller piece of silicon that works with other dies inside a package. In Maia 300, this division allows Microsoft to place the computing functions on an advanced process while treating connectivity, memory control and cost as separate design choices. It does not mean every die in the package uses N2P.

Microsoft's October 12 Maia 300 architecture analysis, authored by Rani Borkar, president of Azure Hardware Systems and Infrastructure, and Saurabh Dighe, corporate vice president of Systems Planning and Architecture, describes that division explicitly.

The document says the architectural changes increase system computing throughput, measured in floating point operations, by 36%, provide 2.5 times the HBM bandwidth and add 14% more bandwidth for communication among accelerators. These figures describe different resources within the system. None, by itself, explains the claimed sevenfold increase in rack token output.

Why memory matters as much as computing power

Maia 300 is designed primarily for inference, the process of using an already trained model to generate answers or perform tasks. Its intended workloads include reasoning over long documents, conversations that require quick responses, and AI agents that execute several steps and use tools over an extended period.

Such workloads can spend substantial time moving model data and stored conversation information rather than performing arithmetic. Microsoft therefore built each accelerator around eight HBM4 memory stacks, providing 288 gigabytes of capacity and 18 terabytes per second of peak bandwidth within a 950 watt thermal design power envelope. Thermal design power is a design specification for power and cooling, not a measurement of constant electricity consumption.

Compared with Maia 200, Microsoft says Maia 300 provides 33% more memory capacity and 2.5 times the memory bandwidth. The capacity comparison is consistent with the reported increase from 216 gigabytes to 288 gigabytes, a gain of 72 gigabytes per accelerator.

Bandwidth units require care. Some reports express the earlier chip's bandwidth in terabits per second and the new rack's bandwidth in petabits per second. Microsoft's Maia 300 documentation uses terabytes and petabytes per second. Bits and bytes differ by a factor of eight, so those differently expressed figures should not be treated as interchangeable.

Software controls where the data goes

Maia 300 extends Microsoft's Software Defined Locally Accessed Dataflow Architecture, or SDLA. In practical terms, this gives software explicit control over where data is stored, how it moves, when its format changes and how operations are coordinated across computing engines, memory and the network.

The new design uses a nonuniform memory arrangement across chiplets. Some memory is closer to a particular computing region than other memory, so software can reduce movement by placing data near the engines that need it. SRAM, a fast memory located on the die, also serves as a temporary workspace managed by software, reducing demands on HBM.

Microsoft combines those changes with expanded support for compact numerical formats, including MXFP4 and storage formats using fewer than three bits. These representations reduce how much memory data occupies and how many bytes must move during inference. Their usefulness depends on software and model support, rather than the smaller format alone.

The design also targets the KV cache, which stores information from tokens a model has already processed so it does not have to recompute that information at every step. Keeping this cache compact matters when conversations become longer or many users need service simultaneously.

Microsoft AI chief executive Mustafa Suleyman connected the hardware program to the company's model development work, explaining why his team wants control over the platform beneath its models.

At that scale we need a platform that we own, and that we co-design with our models. That's why running our Microsoft AI models on top of the Maia chip is a critical focus for our team.

From one rack to 12,288 accelerators

A Maia 300 rack contains 24 trays, each carrying four accelerators alongside the associated host processor, memory, storage and cooling. This doubles the accelerator count from the 48 in a Maia 200 rack. Integrating the host equipment into those trays removes the need for separate host trays and frees space for additional accelerators.

Microsoft's networking design can connect up to 128 racks into a superpod containing 12,288 accelerators. It uses electrical connections within racks and optical connections between racks, with a custom AI transport built on Ethernet.

The network presents a single communication model to software across the superpod. Accelerators can initiate transfers to other accelerators without routing them through a host processor or a separate network interface for communication between racks.

This matters when a model is divided among multiple chips. Tensor parallelism splits parts of a model's calculations across accelerators and requires frequent communication between them. Microsoft says software can keep traffic that needs the shortest response times within a rack while distributing other work more widely.

The same network can separate the initial processing of a prompt from the subsequent generation of an answer. Stored KV data moves directly between the workers handling those stages. Spare accelerators in another rack can also take over when capacity becomes unavailable, although the longer communication path introduces additional latency.

What the efficiency claims actually compare

Microsoft gives two headline comparisons that should be kept separate. The expected 40% increase in tokens per megawatt compares Maia 300 with leading GPU systems. The claim of up to seven times the token output per rack compares a fully configured Maia 300 rack with Maia 200.

A token is a unit of text processed by a language model, often a word fragment rather than a complete word. Token throughput measures output volume, but does not by itself establish answer quality, response speed for each user or performance on every model.

The architecture document also uses differing language for efficiency relative to Maia 200. Its rack discussion says up to 2.5 times more performance per watt, while its concluding comparison says more than 2.5 times. Without detailed test conditions, those descriptions should not be compressed into a single guaranteed result.

The accelerator count explains part of the rack output gain: Maia 300 doubles the number of chips. A sevenfold throughput claim therefore extends beyond density alone, combining additional accelerators with changes to memory, software, networking and execution efficiency. It should not be read as a sevenfold improvement for each individual chip.

The 950 watt accelerator specification is also distinct from the 167 kilowatt rack envelope. Multiplying 96 by 950 watts gives 91.2 kilowatts. The rack figure covers a broader system that includes hosts, memory, networking and power delivery, so those two values describe different boundaries.

Cooling and electricity shape the deployment

At 167 kilowatts per rack, Maia 300 requires an integrated approach to power and heat. Microsoft says liquid cooling removes more than 80% of the rack's thermal load from accelerators, voltage regulators and network switch chips. Air cooling handles memory, storage and other components.

The rack can connect directly to facility water where that is available or use a heat exchanger in a data center relying on air cooling. Dual power feeds and an automatic transfer switch support continued operation when one feed is taken offline for servicing.

Power measurements at rack and row level feed into data center operations. Microsoft says that information helps its algorithms allocate power, place workloads and adjust deployment density to match the electricity and cooling capacity available.

Alistair Speirs, a Microsoft general manager, explained the operating cost connection when discussing the new architecture.

We need new architectures that will give us essentially better performance, better use of the power that we have.

The gigawatt deployment target is therefore more than a chip purchasing plan. It describes a large infrastructure commitment involving racks, networks, electrical capacity and cooling. The published announcement does not identify the sites, regional allocation or completion schedule for that capacity.

Maia 200 provides the starting point

Microsoft says Maia 200 is already serving production workloads, including Microsoft Copilot, and supporting the MAI team as it develops new models. It runs OpenAI models, Microsoft's own MAI models and leading open source model families, with deployment across the United States, Europe and Australia.

Reports differ on the rollout of GitHub Copilot and Microsoft Foundry. One describes both as already running production workloads on Maia 200. Microsoft's October 12 statement says they are expected to follow by the end of 2026. That official schedule is the firmer basis for describing their status.

Microsoft reports up to 1.7 times more tokens per watt for built in AI functions in Teams, Copilot and Cowork experiences compared with leading GPU infrastructure. It also identifies future uses spanning coding assistance, cloud model serving, inference and data generation for MAI, and clinical workflows through Dragon Copilot.

The sequence of announcements and planned milestones is:

• 2023: Microsoft introduces Maia 100, its first custom AI accelerator.
• January 2026: Microsoft announces Maia 200.
• October 12, 2026: Microsoft unveils Maia 300's architecture and confirms Maia 200 production workloads.
• By the end of 2026: GitHub Copilot and Microsoft Foundry are expected to follow onto Maia 200.
• 2027: Maia 300 production deployments and the Maia family's expansion toward multiple gigawatts begin.

Microsoft technical fellow Jim Kleewein says parts of Microsoft 365 Copilot should become faster with Maia 300. No quantified improvement in user response times has been announced.

A larger role for TSMC and Taiwan's server industry

Maia 300 gives TSMC a confirmed role in another major cloud company's advanced AI hardware. Microsoft designs the accelerator, while TSMC manufactures its N2P computing chiplets. Reducing dependence on purchased GPUs therefore does not remove dependence on semiconductor manufacturing capacity.

TSMC has said five facilities using its 2 nanometer process will ramp during 2026, with two in Hsinchu and three in Kaohsiung. N2P volume production is scheduled for the second half of 2026, preceding Microsoft's planned 2027 deployments.

Market estimates put TSMC's monthly 2 nanometer capacity at about 90,000 to 100,000 wafers by the end of 2026 and above 110,000 by mid 2027. Estimates that the process could contribute close to 10% of second half revenue are forecasts, not confirmed revenue from Maia orders.

TSMC's stated targets include 45% more wafer output in the first year of its 2 nanometer process than in the first year of 3 nanometer production in 2023, and a 70% compound annual growth rate in 2 nanometer capacity between 2026 and 2028. Neither target establishes Microsoft's share of that output.

Broader market forecasts identify Foxconn, Quanta, Wistron, Wiwynn, Pegatron, Inventec, Compal and MiTAC as potential beneficiaries of rising AI server demand. Forecasts put AI server shipments at 3.5 million in 2026 and 5.2 million in 2027. Those rounded totals imply growth of about 49%, slightly above the separately quoted 46% forecast.

Forecasts also project 96% growth in major US cloud providers' capital expenditure in 2026, with about half the increase coming from data center construction. Foxconn expects annual AI rack shipments to more than double, while Quanta maintains a target of at least 100% growth in AI servers. These are broader industry expectations, not confirmation of Maia 300 supply contracts.

Nvidia remains part of the plan

Microsoft's goal is to make Maia the primary inference system for its own products and services, not to replace every GPU throughout Azure. Training new models remains a separate requirement, and Microsoft continues to work with Nvidia and AMD for that purpose.

On its July earnings call, Microsoft said it expected to be among the first cloud providers to deploy Nvidia Vera Rubin and AMD Helios systems. Evercore ISI estimates that Maia could represent about 10% of Microsoft's AI chip spending by the first quarter of 2027. Spending share is different from the share of inference work performed, so that estimate does not directly measure progress toward Microsoft's operational goal.

Maia accelerators are not currently available for other companies to rent by the hour. That limits outside evaluation and means the published performance claims do not establish what an independent customer would achieve with a particular model.

The remaining uncertainties include the deployment count, the timing of full gigawatt capacity, detailed benchmark conditions and the eventual division of workloads between Maia and purchased GPUs. Microsoft says it will share more strategy information at Ignite 2026 in San Francisco and online in November.

Key Points

• Maia 300's computing chiplets use TSMC N2P, with production deployments beginning in 2027.
• Each accelerator has 288 GB of HBM4 memory and 18 TB/s of peak memory bandwidth.
• A full rack contains 96 accelerators within a 167 kW power envelope.
• Microsoft expects 40% more tokens per megawatt than leading GPU systems and up to seven times Maia 200's rack output.
• Maia family deployments will expand toward multiple gigawatts starting in 2027, without a published completion deadline.
• Reports of more than one million Maia 300 deployments are not confirmed in Microsoft's announcement.
• Microsoft continues to plan deployments of Nvidia and AMD systems alongside its own hardware.