Microsoft will mainly use its own Maia 300 chips for inference once production starts in 2027

The 40% is a forecast about power. "We expect Maia 300 to deliver 40 percent more tokens per megawatt compared to leading GPU-based systems," said Scott Guthrie, executive vice president of Microsoft's cloud group ^[13]. If it holds, a site that needs 100 megawatts of GPU systems for a given output would need about 71.4 megawatts of Maia 300 for the same output ^[19], a saving of 28.6 megawatts ^[20]. Cryptobriefing noted that it is a target, not a benchmark anyone has independently verified ^[16].

Power is one line of the cost. The only per-dollar figure on record belongs to the previous chip: CEO Satya Nadella told analysts in July that Maia 200 "delivers 30% better performance per dollar than the latest generation hardware in our fleet" ^[3]. The sources do not include one for Maia 300. Alistair Speirs, a Microsoft general manager, tied power to the bill: "You can think of power translating pretty directly into operational cost for something like a data center as well" ^[17].

We think the near-term effect on Nvidia is small and the later effect depends on volume. Evercore's 10% estimate comes from analysts who recommend buying Microsoft shares ^[7], and it falls in the same year Maia 300 production starts ^[2]. Volume is the open variable. Cryptobriefing, citing unnamed "research findings", reported that Microsoft is negotiating with TSMC to make more than 300,000 Maia 300 chips for delivery in 2027 ^[21]. At 96 chips a rack that is about 3,125 racks ^[22], after only tens of thousands of Maia 200 units ^[5]. The figure is not attributed to Microsoft.

The case against us is that 10% is a starting share: if the 300,000-chip figure is right, the 2027 ramp is several times Maia 200's. Speirs, who declined to give a timeline for the multi-gigawatt deployments ^[11], said "We reserve a little bit of optionality there, because this is an industry that's changing very fast, and we want to deploy the right silicon for the right models that we're using at the right time as well" ^[10].

We would be wrong about the pace if Microsoft confirms the 300,000-chip volume or if analyst estimates of Maia's spending share rise well above 10% during 2027. The efficiency claim, meanwhile, has no outside test. "Nobody has tested them outside of Microsoft yet," said Patrick Moorhead, CEO of advisory firm Moor Insights and Strategy ^[15].