Microsoft Debuts Maia 300 Chip and Pledges Multi-Gigawatt Inference Fleet
In one of its most significant in-house silicon moves yet, Microsoft unveiled Maia 300, a next-generation AI accelerator built to run the inference workloads behind Microsoft 365 Copilot and other services. The company says the new part can pack 96 chips into each rack, handle up to seven times more tokens per rack than before, and deliver roughly 40% more tokens per megawatt than GPU-based systems.
Microsoft has unveiled Maia 300, the third member of its custom Maia line of AI accelerators, and has signalled a major shift in how it will run its artificial intelligence services. The chip, announced on Monday 12 October, is explicitly aimed at inference — the ongoing process of answering prompts and generating output — rather than the large-scale training runs that have dominated headlines. Production is scheduled to begin in 2027.
What Microsoft announced
According to reporting on the announcement, Microsoft expects to fit 96 Maia 300 chips into each data-center rack, double the count for the previous-generation Maia 200. Microsoft says the new accelerator can run through up to seven times more AI tokens per rack, delivering about 40% more tokens per megawatt when compared with graphics-processing-unit systems — a figure that maps directly onto the dominant cost driver for a data center: power.
- Maia 300 is Microsoft's third-generation Maia AI accelerator, following the Maia 100 (2023) and Maia 200 (announced January 2026).
- Each rack will hold 96 Maia 300 chips, roughly double the previous generation.
- The design can handle up to seven times more AI tokens per rack than before.
- Microsoft claims about 40% more tokens per megawatt versus GPU-based systems.
- Maia 300 will become the primary inference silicon for Microsoft products and services, beginning with parts of Microsoft 365 Copilot.
Why it matters
The clearest strategic signal is a deliberate move to reduce reliance on Nvidia for inference workloads. Microsoft's Azure cloud already runs a broad set of models, and its in-house parts have grown steadily: executives said the previous Maia 200 delivered around 30% better performance per dollar than the latest hardware in its fleet and was already handling some workloads. By building the inference stack on its own silicon, Microsoft can control both the hardware and the models it hosts, and stand to capture more of the margin from the inference traffic that now makes up a growing share of cloud AI demand.
Analysts have already begun to factor the shift into forecasts. A note from Evercore ISI estimated that Maia will represent roughly 10% of Microsoft's spending on AI chips by the first quarter of 2027 — a meaningful slice of a budget that has been dominated by GPU purchases. Microsoft executives framed the timing as a balancing act, with one general manager saying the company is reserving "optionality" in a market that "is changing very fast."
If you depend on Microsoft AI services — Microsoft 365 Copilot, GitHub Copilot, or Azure-hosted models — expect a quieter efficiency push rather than a headlining feature launch: Microsoft says parts of Microsoft 365 Copilot should get faster on Maia 300 hardware. For builders, the more important takeaway is that inference economics at the cloud layer are about to get more competitive, which historically puts downward pressure on per-token pricing.
Microsoft has not committed to an exact schedule for at-scale deployment, and one spokesperson declined to provide a timeline, noting that the company wants to "deploy the right silicon for the right models ... at the right time." If the multi-gigawatt roadmap holds, Maia 300 would be the most concrete evidence yet that Microsoft intends to run a substantial share of its own AI workload on chips it designs and controls.
Sources: Microsoft: Microsoft's latest custom chip: Maia 300 (official announcement); CNBC: Microsoft will mainly use its Maia chips for AI inference; HPCwire: Microsoft Announces Maia 300, Its Next-Gen Custom AI Chip.
Corrections: Report an error at corrections@wireshaman.com.