Microsoft Maia 300 arrives with a number that does not sound efficient: 167 kilowatts. That is the power envelope for one rack containing 96 of Microsoft’s new AI accelerators, scheduled to enter production deployments in 2027.
The apparent contradiction matters because Microsoft is not promising low-power AI. It is promising more AI work inside a fixed power budget—and more control over the hardware bill behind Copilot, GitHub Copilot and its own models.
How Microsoft Maia 300 measures efficiency
Microsoft says Maia 300 will deliver 40% more tokens per megawatt than “leading GPU-based systems.” Tokens are the small units AI models process when they read prompts and generate answers. The metric therefore asks how many useful model outputs a data center can serve for a given amount of electricity.
That claim comes from Microsoft, and the company did not identify the competing GPU systems in its public comparison, Reuters reported. It should be read as a vendor benchmark until independent workloads provide a clearer comparison.
The architecture shows where Microsoft expects the gains to come from. Each 950-watt accelerator uses eight HBM4 memory stacks with 18 TB/s of peak bandwidth. A full rack combines 27.6 TB of high-bandwidth memory with 1.7 PB/s of aggregate memory bandwidth, according to Microsoft’s announcement.
Maia 300 also splits compute, memory control and input-output functions across specialized chiplets rather than forcing every job onto one monolithic die. Microsoft says the compute chiplets use TSMC’s N2P process. The broader system—from networking to cooling and workload placement—is co-designed to reduce expensive data movement during long-context reasoning and multi-step agent tasks.
Efficiency does not mean less total power
Microsoft says a 96-chip rack can produce up to seven times Maia 200’s token throughput and more than 2.5 times its performance per watt. But it also plans Maia deployments measured in multiple gigawatts. Reuters reports that the company could deploy more than one million Maia 300 chips next year.
Those statements can both be true. Better efficiency allows Microsoft to fit more inference into constrained data centers; enormous deployment can still increase total electricity demand. Maia is an answer to scarcity, not an end to it.
The strategic consequence reaches beyond engineering. Microsoft wants Maia to become the primary inference platform for its own products and services. Axios notes that this does not necessarily extend to third-party Azure customers, while Nvidia remains dominant in AI training. Still, a company that buys Nvidia hardware is now designing a rival system for the fastest-growing part of AI computing.
Maia 300’s real test will not be whether 167 kilowatts sounds large. It will be whether Microsoft can turn each of those watts into enough useful answers—and enough cost advantage—to justify building the next gigawatt.
