Microsoft has unveiled Maia 300, a new AI chip designed to make its own AI products faster and cheaper to run - though it won't ship until 2027.
The chip is built specifically for inference rather than training, and Microsoft says a rack of 96 Maia 300 accelerators will deliver up to seven times the token throughput of an equivalent Maia 200 rack, alongside a claimed 40% more tokens per megawatt than competing GPU-based systems. It uses eight stacks of HBM4 memory with 18TB/s of peak bandwidth, draws up to 950W, and splits compute, memory-control and input/output functions across separate chiplets built on TSMC's N2P process. Microsoft says the design leans on software-defined dataflow and AI-assisted kernel tuning to cut data movement, rather than simply adding more raw silicon.
The real audience here isn't outside hardware buyers - it's Microsoft's own product teams. Maia 300 is meant to underwrite everything from GitHub Copilot's agentic coding features to Microsoft Teams and Cowork, areas where slow AI responses are a visible daily annoyance rather than an abstract benchmark number. That continues the pattern set by Maia 200, which Microsoft says already runs Microsoft 365 Copilot and claims a 1.7x tokens-per-watt edge over GPU infrastructure for those same products.
Every efficiency figure here is Microsoft's own, measured against unnamed leading GPU-based systems, and the chip is still roughly a year from shipping - so the real test is how Copilot and Teams actually feel once it arrives. Still, Maia 300 continues Microsoft's push to treat in-house AI silicon as core infrastructure, the same bet Google and Amazon have made with their own custom chips.