Loading page…
The shift targets running AI models, not training them. Microsoft promises denser racks and better power efficiency, but outside testing remains a missing piece.
Listen to this story
Microsoft is positioning its Maia family as an inference architecture, not a broad replacement for Nvidia or AMD, which remain part of its training infrastructure. Maia 300 is designed for 96-chip racks—double Maia 200’s density—and Microsoft claims up to seven times as many tokens per rack and 40% more tokens per megawatt than GPUs, without naming the comparison GPU. Production is scheduled for 2027; because outsiders cannot rent or test Maia, customers cannot independently verify those efficiency claims.
Microsoft says parts of Microsoft 365 Copilot should become faster with Maia 300, but that improvement is an expectation, not a result measured on available hardware.
Evercore ISI analysts forecast Maia will account for about 10% of Microsoft’s AI-chip spending by Q1 2027; that is not a deployment commitment or share of requests served.
Microsoft is targeting multi-gigawatt deployments across the Maia family but has provided no timeline for the broader transition.
Microsoft intends to put its own chips at the center of serving AI requests, reducing its reliance on Nvidia for that work. The company introduced its third-generation Maia 300 accelerator on October 12, with production scheduled for 2027. The shift targets inference—running AI models to handle incoming requests—rather than training them.
Maia 300’s design starts with fitting more processors into each data-center rack. Microsoft expects a rack to hold 96 chips, twice the number for Maia 200. That gives the company a denser system for processing requests, rather than simply another chip to add to its existing fleet.
Microsoft says the new version can process up to seven times more AI tokens per rack. It also claims 40% more tokens per megawatt than GPUs, or graphics processing units. Those are company-supplied performance figures; CNBC’s account does not identify the GPU model used for the power comparison.
The power measure is central to Microsoft’s business case. General manager Alistair Speirs told CNBC that new architectures need to deliver better performance and make better use of available electricity. Power, he said, translates fairly directly into a data center’s operating cost. The pitch is more AI processing from the power Microsoft already has.
Microsoft also points to a result users might notice: technical fellow Jim Kleewein said parts of Microsoft 365 Copilot should become faster with Maia 300. That is an expectation for the forthcoming hardware, not a measured improvement available today. The company has not yet reached the chip’s scheduled production year.
Maia is moving toward becoming Microsoft’s primary architecture for running models, but Nvidia and AMD remain part of its training infrastructure. Speirs said Microsoft partners with both companies for that work. The distinction limits the scope of the announcement: Microsoft is changing how it serves AI requests, not declaring an end to outside chip purchases.
The new chip builds on a program Microsoft began with Maia 100 in 2023. Maia 200 followed in January 2026. In July, CEO Satya Nadella said Maia 200 was handling OpenAI models and delivered 30% better performance per dollar than the latest-generation hardware in Microsoft’s fleet. That earlier claim concerns Maia 200, not the newly introduced processor.
The near-term spending picture is smaller than the long-term ambition. Evercore ISI analysts estimated that Maia would represent about 10% of Microsoft’s AI-chip spending by the first quarter of 2027. Their forecast covers Maia spending, not the share of requests it will process, and is not a deployment commitment from Microsoft.
Microsoft is targeting multi-gigawatt deployments across the Maia family, according to executive vice president Scott Guthrie’s blog post, cited by CNBC. Speirs declined to give a timeline. He said Microsoft wants to preserve flexibility to deploy the right chips for the models it uses as the industry changes, leaving the broader transition without a fixed completion date.
There is also a limit on outside scrutiny. Unlike Google’s and Amazon’s AI accelerators, Microsoft’s cannot be rented by the hour by other companies. Patrick Moorhead, CEO of Moor Insights and Strategy, told CNBC that nobody outside Microsoft had tested them yet. For now, buyers can assess Microsoft’s efficiency promises, but cannot directly rent Maia hardware to check those promises themselves.
Story updates
Microsoft’s Maia 300 is designed to keep AI data close to the hardware that needs it, using software-controlled data movement and on-chip memory. The Azure accelerator targets longer reasoning tasks, responsive conversations and multi-step agents. Microsoft’s aim is to use computing, memory and networking resources more efficiently—not simply add more processing capacity.
Those architectural details are part of the Maia 300 introduction reported by TechRadar on October 12. Microsoft plans to begin deployment in 2027. The chip is designed primarily for inference: running AI models to respond to requests. Its predecessor, Maia 200, already runs production workloads including Microsoft 365 Copilot and models from Microsoft’s AI team, according to the company.
Microsoft describes the design as a software-defined dataflow architecture. In plain terms, software helps control how data moves through the system. Combined with AI-assisted optimization of kernels—the software routines that perform computing work—the approach is intended to keep data near the parts of the hardware that use it. Reducing that movement is central to Microsoft’s efficiency pitch.
The memory design gives that software another tool. Specialized software can use on-chip SRAM, a form of memory, as a scratchpad: a working area for data needed during a task. Microsoft says this can ease pressure on high-bandwidth memory, or HBM, during demanding workloads. The design therefore combines substantial HBM resources with software-managed use of memory on the chip itself.
Maia 300 separates computing, memory-controller and input/output functions into different chiplets, the smaller components within its multi-chiplet architecture. Microsoft says the compute chiplets will use TSMC’s N2P manufacturing process. This division sits alongside the software-controlled dataflow design; it is not a separate product announcement or a change to the deployment timetable.
The accelerator also supports low-precision formats, including an enhanced MXFP4. These represent values with less precision. Microsoft says that reduces memory and bandwidth demands, helping the system handle larger context windows—the information a model can work with at once—and increase output within the same power envelope. Those are intended benefits of the design, not demonstrated improvements to individual Microsoft applications.
One accelerator: eight HBM4 memory stacks, with Microsoft citing 18 TB/s of peak bandwidth and a 950-watt thermal design power.
One rack: 96 accelerators, 27.6 TB of HBM capacity and 1.7 petabytes per second of aggregate HBM bandwidth, within a 167-kilowatt power envelope.
Networking is designed as part of the system’s dataflow, rather than a separate layer. Maia 300 will use high-bandwidth connections between accelerators for tensor parallelism, which spreads model computation across accelerators. It also uses a custom AI transport built on Ethernet infrastructure. Microsoft says this networking approach can scale from a single rack to data-center deployments.
Microsoft says the rack’s density will allow more simultaneous AI interactions and larger workloads. That is the service-level ambition behind the hardware specifications. The two headline performance comparisons measure different things: output for a given amount of power against GPU-based systems, and total rack output against Maia 200. Neither figure, by itself, establishes how much faster a particular conversation or agent task will finish.
techradar.comLoading discussion...
Join the conversation
Explain what would make those claims credible to you.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.