Microsoft is in discussions with TSMC to secure manufacturing capacity for more than 300,000 units of its Maia 300 custom AI accelerator for delivery in 2027. The company plans to unveil the chip this fall, and a Microsoft manager indicated the ultimate goal is to produce 'gigawatts' worth' of Maia chips—a scaling ambition that echoes cloud peers Amazon and Google's custom silicon strategies.
The Maia 300 is designed to deliver over 30% cost savings per token compared to Nvidia's GPUs, positioning it as Microsoft's answer to CUDA vendor lock-in as capex pressures mount across hyperscalers. Microsoft hopes to court major customers including Anthropic, which has already begun designing its own semiconductor team for Claude models. Maia 200 (launched January 2026) remains limited in deployment to just a handful of Azure data centers, so the 300,000-unit TSMC order signals a massive acceleration in Microsoft's custom silicon roadmap.
For architects, the 2027 delivery timeline and tight TSMC CoWoS packaging capacity (cited by J.P. Morgan as constrained through 2027) mean Maia 300 availability will be supply-gated, not demand-gated. Microsoft's willingness to lock in quarter-megawatt-scale TSMC capacity for a chip not yet launched underscores how serious cloud providers are about de-risking Nvidia dependency. Watch whether Anthropic and other customers actually adopt Maia at scale or remain Nvidia-first.