Microsoft to rely more on homegrown AI chips, reducing dependence on Nvidia

Microsoft unveiled its third-generation Maia AI accelerator and plans to rely more on its own chips for incoming AI requests, while continuing to use Nvidia and AMD for model training. Microsoft says Maia 300 can support up to seven times more AI tokens per rack and 40% more tokens per megawatt than GPUs; each rack is expected to hold 96 chips, double the Maia 200 count. Production of Maia 300 is slated to start in 2027, and Evercore ISI analysts estimate Maia will account for about 10% of Microsoft’s AI-chip spending by Q1 2027.

Analysis

Custom silicon is a cost-control lever for Microsoft, not yet evidence of a broad Nvidia displacement. If Maia delivers the claimed throughput per watt at scale, Azure can improve inference economics or preserve pricing flexibility as compute demand rises; utilization, yields, and software compatibility—not peak chip specifications—will determine whether savings reach margins. The near-term read-through to NVDA revenue looks limited: Maia is estimated to be only about 10% of Microsoft AI-chip spending in early 2027, and Microsoft still relies on external silicon for training. AMD faces a mixed outcome: potential inference substitution, but continued opportunity in training and general-purpose compute.

The more important competitive effect is strategic: Azure gains another lever in negotiating accelerator supply and differentiating cloud unit economics. Amazon and Google have analogous in-house options; Microsoft’s lack of external Maia rentals means it cannot yet validate demand or monetize the chip as a standalone platform. Company performance claims and the multi-gigawatt ambition remain unverified; rack-level benchmarks do not establish fleet-wide cost savings.

Over days, avoid treating this as a break in Nvidia’s demand outlook. Over 1–3 months, watch Azure capacity commentary and evidence of Maia workloads moving into production. Over 6–18 months, monitor Maia deployment pace, inference cost per token, and Azure gross-margin trajectory. The thesis weakens if deployment slips, utilization disappoints, or Microsoft increases external accelerator purchases despite Maia availability.

Key Decisions for Investors

• Prefer a measured MSFT-over-NVDA relative position only on a pullback or after confirming deployment and utilization; size it as a multi-quarter margin thesis, not an immediate earnings catalyst. The main risk is that Maia’s claimed efficiency does not translate into fleet economics.
• Do not short NVDA on this announcement alone. Training demand and Microsoft’s continued use of external accelerators limit the near-term substitution signal; reconsider only with evidence of sustained share loss in Microsoft’s actual accelerator purchases or a change in Nvidia demand commentary.
• Track Azure margins, AI capacity additions, and inference cost per token alongside Maia production and deployment timing. A delay, weak utilization, or no improvement in cloud economics would falsify the MSFT margin-upside thesis.
• Treat AMD as a watch item rather than a directional trade: inference substitution could be a headwind, while continued training demand may offset it. Verify workload mix and purchase trends before taking a position.