Microsoft unveiled Maia 300, its third-generation in-house AI chip, on October 12 -- in a post by Scott Guthrie, the company's Executive Vice President of Cloud + AI, announcing a plan to scale the whole Azure Maia family to multi-gigawatt deployments starting in 2027. The goal, stated plainly: Maia becomes "the primary inference system for Microsoft products and services," not a side project next to the Nvidia GPUs that still run the bulk of Microsoft's AI today.
Maia 300 is built inference-first: 8 HBM4 memory stacks per chip (roughly 288 GB), 18.0 terabytes per second of peak memory bandwidth, a 950-watt power envelope, and on-die SRAM used as a software-managed scratchpad to keep workloads off the memory bus when possible. It's a multi-chiplet design -- compute, memory-controller, and I/O functions physically separated -- with the compute chiplets built on TSMC's N2P process and an enhanced MXFP4 low-precision format meant to shrink memory and bandwidth needs for both computation and KV-cache storage. A rack holds 96 of the accelerators, draws 167 kilowatts, and carries 27.6 terabytes of total HBM across 1.7 petabytes per second of aggregate bandwidth -- which is just the per-rack math on the 8-stack, 18 TB/s figures stated per chip, not a separately disclosed number.
What Microsoft's Maia 300 figures actually cover
- 27.6 TB · HBM4 per rack
- Total memory across a 96-accelerator rack
Includes: Microsoft's own disclosed per-chip memory (8 HBM4 stacks) multiplied by 96 -- works out to roughly 288 GB per accelerator, matching Notebookcheck's independently derived figure.
Excludes: Compute throughput (FLOPS) -- Microsoft's post references a spec chart it did not publish in full text. - 40% · more tokens per megawatt
- Microsoft's headline efficiency claim for Maia 300
Includes: A comparison Microsoft itself ran.
Excludes: The name of the 'leading GPU-based system' being compared against, and any independent confirmation. - 1,000,000+ · chips, reported 2027 target
- Reuters' reported deployment plan for Microsoft's own AI services
Includes: A target Microsoft is reportedly negotiating capacity for with TSMC.
Excludes: A signed purchase order -- Microsoft publicly disputed the specific figures reported in August without offering its own number. - $35.80B · Microsoft capex, quarter ended June 2026
- Property and equipment spending, more than double the $17.08B a year earlier
Includes: The financial backdrop driving Microsoft toward a cheaper, owned alternative to Nvidia GPUs.
Excludes: Any Maia-specific cost breakout -- Microsoft does not report chip-program spending separately.
Every comparison number in that list that makes Maia 300 look good is Microsoft's own, and every one of them is missing the baseline a reader would need to check it. The 40%-tokens-per-megawatt figure doesn't name the GPU system it beats. A separate claim -- up to 7x more rack-scale token throughput than Maia 200 -- at least names its comparison point, but it's still Microsoft grading its own newer chip against its own older one. And Maia 200 itself, already in production and running Microsoft Copilot across the US, Europe and Australia, is credited with 1.7x more tokens per watt than "leading GPU infrastructure" for the built-in AI features in Teams, Copilot and Cowork -- again, no named comparison system.
- Maia 300 delivers 40% more tokens per megawatt than leading GPU-based systems.
- A 96-accelerator Maia 300 rack delivers up to 7x the token throughput of the equivalent Maia 200 configuration.
- Microsoft will deploy over one million Maia 300 chips in 2027.
None of that is happening in a vacuum. Microsoft's capital spending on property and equipment hit $35.80 billion in the quarter ended June 2026, more than double the $17.08 billion spent a year earlier; trailing four-quarter spending topped $115.95 billion, up 80% year over year. Gross margin has fallen for three straight quarters, from 69.05% to 67.20%. A chip that cuts the cost of running Microsoft's own AI workloads is a direct lever against exactly that pressure -- which is also why skepticism about an unnamed-baseline efficiency claim matters more here than it would on an ordinary spec sheet: the 40% figure is doing real work in the story Microsoft is telling investors about where the capex finally pays off.
The more basic question is whether any of this threatens Nvidia's position the way the framing implies.
The nearest real test of how fast any of this actually moves is already on the calendar: Microsoft says GitHub Copilot and Microsoft Foundry should be running on Maia 200 -- the chip already in production, not Maia 300 -- by the end of 2026. Maia 300 itself doesn't reach production deployment until 2027, with Microsoft promising more specifics at Ignite in San Francisco next month. Until then, the chip that might matter is a spec sheet and a pair of uncredited comparisons.
- Microsoft unveiled Maia 300, its third in-house AI chip, on October 12.
- Each rack holds 96 chips, draws 167 kW, and carries 27.6 terabytes of HBM4 memory.
- Microsoft claims 40% more tokens per megawatt than rival GPU systems, naming no comparison system.
- Production deployments start in 2027; Reuters reports Microsoft may deploy over 1,000,000 chips that year.
- Caveat: Maia 300 runs only Microsoft's own services for now -- Azure customers still rent Nvidia GPUs.