J.P. Morgan: Microsoft Maia 300 Inference Performance Up 40%, Million-Chip Deployment Cuts Costs

per Chaoxiang Research, JPMorgan's research report dated October 12, 2026 indicates that Microsoft's Maia 300 achieves 40% more tokens per megawatt than leading GPU systems, with rack-level token throughput in a full 96-accelerator configuration being 7 times that of the Maia 200; the Maia 200 delivers 1.7 times more tokens per watt compared to GPU infrastructure; Reuters reported Microsoft's plan to deploy over 1 million Maia 300 chips in 2027; Microsoft's capital expenditure is projected to be approximately $210 billion for fiscal year 2027 and around $175 billion for calendar year 2026.

JPMorgan anticipates that once Maia becomes the primary inference system, it will boost Copilot performance and reduce cost per compute unit, with Microsoft's cloud gross margin poised to improve. The firm maintains an Overweight rating with a price target of $625.