Microsoft Books TSMC Capacity for 300,000 Maia 300 AI Chips for 2027
The software giant is accelerating its custom silicon roadmap to reduce reliance on Nvidia and lower the cost of AI inference.
Microsoft has secured manufacturing capacity with TSMC for more than 300,000 next-generation Maia 300 AI chips scheduled for delivery in 2027. The move signals an aggressive push to scale its proprietary hardware ecosystem to support the massive compute demands of Azure, Copilot, and OpenAI-linked workloads.
This procurement comes amid a volatile development cycle for Microsoft's custom silicon. The Maia 200 chip, known internally as "Braga," faced significant setbacks, with mass production delayed from 2025 to 2026. To maintain its competitive edge during these shifts, Microsoft is reportedly considering the introduction of an interim "Maia 280" chip in 2027 before transitioning to the Maia 400 in 2028.
The Silicon Race
Microsoft entered the custom AI chip market later than rivals like Google, with its Tensor Processing Units (TPUs), and Amazon, which utilizes Trainium and Inferentia. While the company launched the Maia 100 in the 2023-2024 window, the rapid evolution of large language models has forced frequent design changes and roadmap adjustments.
Despite these internal efforts, Microsoft remains deeply tethered to external suppliers. Reports from The Information, citing Nvidia employees, state that Microsoft was Nvidia's largest customer by revenue last year, though Nvidia does not officially disclose individual customer rankings. This dependency creates a strategic urgency to migrate high-volume workloads to in-house hardware.
The Economics of Inference
This shift is primarily a battle over "inference economics." Every query processed by Copilot or an AI agent incurs a direct compute cost. By owning the underlying hardware, Microsoft can bypass the high margins charged by Nvidia and gain tighter control over data center power consumption and capacity.
Reducing the cost per token is critical for the company's long-term margins as it scales AI services to millions of enterprise users. Success in custom silicon would allow Microsoft to optimize its hardware specifically for the architectural needs of the models it hosts, rather than relying on general-purpose GPUs.
Expanding the Ecosystem
Microsoft is also looking beyond its own services to find utility for its hardware. Anthropic has been in early-stage talks to utilize Microsoft's custom Maia chips—specifically the Maia 200—to power its Claude models.
Industry observers will be watching whether Microsoft can resolve the production delays that plagued the Maia 200. If the company can successfully execute the delivery of the Maia 300 and the potential Maia 280, it will significantly diminish its vulnerability to the current GPU supply chain bottlenecks.