Microsoft’s Biggest Bet on Its Own Silicon, in a Bite

Microsoft says it will run most of its own AI on chips it designs, and it measured the new one in tokens per megawatt.

• Azure Maia 300 is the third generation, and Microsoft calls it its biggest bet yet on its own silicon. It packs 96 accelerators into one 167 kW rack.
• The claims are efficiency claims. Microsoft says 40 percent more tokens per megawatt than leading GPU systems, up to 7x the rack-scale token throughput of Maia 200, and up to 2.5x the performance per watt.
• Maia becomes the primary inference system for Microsoft’s products and services from 2027, at multi-gigawatt scale. Production starts that year.
• It is not a bid for Nvidia’s training market. Maia is not rented to outside customers, so this is a cost play, and the analyst read is that Nvidia and AMD keep shipping into Azure.

Read the full analysis for the three questions to put to any vendor who sells compute by the megawatt.