Overview
- Microsoft announced Maia 300 on Monday and said it will begin scaling the Maia family to multi-gigawatt deployments starting in 2027 to make Maia the primary system for its inference workloads.
- The company claims Maia 300 uses a chiplet design with compute chiplets on TSMC’s 2nm (N2P) node, eight HBM4 stacks and a 96-accelerator rack that it says delivers about 40% more tokens per megawatt and up to seven times the rack throughput of Maia 200.
- Maia 200 is already running production inference for services such as Microsoft 365 Copilot and some OpenAI models, and Microsoft says Maia 300 will target long-context reasoning, low-latency conversational tasks and multi-step agent workloads.
- Microsoft will continue to use Nvidia and AMD for parts of its infrastructure, especially model training, and it has not announced broad public rental of Maia capacity to outside customers.
- Large-volume manufacturing depends on TSMC capacity and advanced packaging; reports of talks for hundreds of thousands to over a million chips exist but the Maia 300 efficiency and throughput figures remain company claims pending independent benchmarks and outside customer tests.