Microsoft's Maia 300 chip will run its AI services from 2027

On October 12 Microsoft unveiled Maia 300, its third-generation in-house AI accelerator. It will be deployed at multi-gigawatt scale in Azure from 2027 and become the main inference hardware for Microsoft services including Copilot. Microsoft still works with Nvidia and AMD for training.

Maia 300 is built from chiplets. Compute dies use TSMC's N2P process, with some supporting silicon on a cheaper 3 nm process. Eight HBM4 stacks give 288 GB of memory and 18 TB/s of bandwidth. It delivers 13.8 petaflops at FP4 and 6.9 at FP8 and draws 950 W. A rack holds 96 chips, draws 167 kW and carries 27.6 TB of memory.

Microsoft says Maia 300 delivers up to seven times the rack-scale token throughput of Maia 200, more than 2.5 times the performance per watt, and 40% more tokens per megawatt than GPU systems. These are company figures, not independent benchmarks.

Inference is a trained model producing answers, now the bulk of AI spending. HBM is high-bandwidth memory stacked next to the processor. Tokens are the small text pieces models process and most AI services price by.

Google, Amazon and now Microsoft are building their own chips to cut reliance on Nvidia and lower inference costs, which feeds into the price of services like Copilot.

Microsoft announced its first chip, Maia 100, in late 2023 and tested it on limited internal workloads. Maia 200 broadened use. Maia 300 is the first designed for gigawatt-scale deployment, roughly the output of several large power plants, showing Microsoft now treats the chip as the backbone of its infrastructure rather than an experiment. It is also expanding energy deals and liquid cooling in parallel.

Do not treat AI costs as fixed: token prices fall every year. Prefer usage-based pricing over long fixed contracts, try office AI tools as they get cheaper, and measure the time each automation saves now so you know where to scale when prices drop.