Maia 300: Microsoft's 2nm AI Chip Faces a 2027 Supply Test

What Microsoft announced

On October 12, 2026, Microsoft formally introduced Maia 300, the third generation of its in-house AI accelerator. The company describes it as an inference-first chip that will become the primary engine answering requests across its products starting in 2027 12. The plan has three parts: the chip itself, its role as the main serving system for products like Copilot and Teams, and a push to take the Maia family to multi-gigawatt deployments from next year 1.

The headline claim is efficiency. Microsoft says Maia 300 produces 40% more tokens per watt than leading GPU-based systems 1. Other coverage frames the same figure as tokens or useful compute per megawatt 245. The company also says the chip delivers seven times the token output of Maia 200 2.

At the rack level, Microsoft is packing 96 accelerators into roughly 167 kilowatts 24. For comparison, Nvidia's GB300 racks hold 72 GPUs at about 142 kilowatts 4. The chip carries 288 GB of HBM4 1. It uses a chiplet layout, with compute dies on TSMC's 2-nanometer process and data-movement dies on cheaper 3-nanometer silicon 4. Microsoft has said it intends to install more than one million units in 2027, against expected 2026 capital spending above $175 billion 4.

The leaks got a lot wrong, or the plan changed

The official specs differ from the picture that circulated over the summer. In August, TechTimes reported a monolithic-sounding 3nm design with more than 140 billion transistors and 216 GB of HBM3e at 7 TB/s, plus 272 MB of on-chip SRAM 3. RuntimeWire and TrendForce reported that Microsoft was negotiating TSMC capacity for more than 300,000 Maia 300 chips for 2027 delivery 3. Coverage at the time stressed that most of those details came from industry reporting rather than Microsoft datasheets 3.

The announced chip moves to a newer memory generation with more capacity, and to a split-node chiplet design. The stated volume is also more than three times the August figure 14. Two readings are possible. The leaks may have described an earlier or interim configuration, or Microsoft may have expanded its ambitions between August and October. Either way, the gap between 300,000 and "over one million" is the most consequential number in this story.

Why the efficiency claim matters

Microsoft's pitch rests on power, not raw speed. One outlet notes that the benefit is more AI output within the same electricity budget. That could mean more capacity for complex, multi-step tasks, but it does not guarantee faster individual responses 5. That framing fits a datacenter buildout constrained by grid power more than by floor space.

The per-rack numbers support the density story. Microsoft fits 33% more accelerators per rack for about 18% more power than a GB300 rack 24. Still, the 40% figure is Microsoft's own estimate against unnamed "leading GPU-based systems" 1. Until independent workload benchmarks appear, it should be treated as a vendor claim. The comparison point also matters: by the time Maia 300 ships at volume, Nvidia's next platform will be the relevant baseline, not GB300.

The supply-chain fine print

The efficiency claim is not the real test for a 2027 ramp. Manufacturing is. Placing compute dies on TSMC's 2nm process puts Microsoft in competition for leading-edge wafers. Apple, Nvidia, AMD and other hyperscalers' custom-silicon programs all want the same capacity. Analysts already framed Maia 300 as a challenger to Nvidia Blackwell, Google's TPU v7 and AWS Trainium 3 3, and every one of those programs draws from the same small pool of advanced-node and advanced-packaging capacity.

HBM4-based chiplet accelerators also depend on sophisticated packaging and on memory suppliers ramping a new HBM generation. None of the coverage details how much packaging or HBM4 supply Microsoft has locked in. That is notable given how aggressively the volume target appears to have grown since August 34. The analytical takeaway is that "one million units" is best read as an intent, not a delivery schedule.

The bottom line

Maia 300 is Microsoft's clearest statement yet that it wants to own the hardware running its own AI traffic. It positions the chip as a way to cut GPU dependence and energy costs for Copilot and Azure workloads 25, building on Maia infrastructure that already supports Copilot 5. The design looks credible and the rack math is competitive. Whether Maia meaningfully dents Nvidia's position in Microsoft's datacenters in 2027 will depend less on tokens per watt than on how many 2nm wafers, HBM4 stacks and packaging slots Redmond can actually secure.