Microsoft is reshaping its Maia chip family around the economics of AI inference, aiming to reduce reliance on Nvidia and deliver more tokens per unit of power as data-centre capacity tightens.
Microsoft is recasting its Maia programme as a chip family built chiefly for AI inference, a shift that underscores how cloud economics are changing. Inference, the phase in which a trained model answers prompts, generates text or creates images, now accounts for a growing share of the cost of running AI services. Microsoft says that is precisely where its custom silicon can deliver the biggest return, with Maia 200 already designed to improve the economics of token generation and cut the cost per unit of useful work.
That focus is reflected in the hardware itself. According to Microsoft’s own materials, Maia 200 is built on TSMC’s 3-nanometre process and uses lower-precision arithmetic aimed at inference rather than the heavier computation associated with training. The company says the accelerator delivers around 30% better performance per dollar than the hardware it already deploys, and it has positioned the chip as a first-party alternative to the Nvidia-dependent systems that still dominate much of the market. Microsoft has also claimed competitive results against Google’s TPU line and Amazon’s Trainium family, though those comparisons are based on its own benchmarks and should be read in that light.
The next step is Maia 300, which Microsoft says will become the main inference platform across its products and services when production begins in 2027. Axios reported that the move is part of a wider push among large technology companies to gain more control over AI infrastructure, reduce dependence on Nvidia and ease pressure on supply chains for advanced packaging and high-bandwidth memory. TechRadar, citing Microsoft’s launch material, said Maia 300 is being built with an inference-first design and is intended to support workloads such as long-context reasoning and multi-step conversational systems.
Microsoft says the new chip will lift efficiency rather than simply add raw capacity. The company has claimed Maia 300 can produce 40% more tokens per megawatt than leading GPU-based systems, while a full rack of 96 accelerators can reach up to seven times the throughput of Maia 200. TechRadar reported that the chip uses a chiplet-based architecture and supports newer low-precision data formats, part of an effort to make the system denser, cooler and easier to scale across data centres. Microsoft also says the rollout will expand to multi-gigawatt deployment from 2027, a sign that power availability, not just chip availability, is now a binding constraint on AI growth.
Even so, the shift does not mean Microsoft is abandoning outside suppliers. Daniel Newman, chief executive of The Futurum Group, told reporters that Maia is an important step for Microsoft’s in-house silicon efforts but will not fully replace Nvidia or AMD hardware. That caution is consistent with the broader industry picture: Microsoft’s own infrastructure still mixes chip types, Nvidia GPUs remain central for training large models, and third-party Azure customers such as OpenAI may continue to rely on Nvidia systems. The clearest implication is not replacement, but diversification, as Microsoft tries to control the cost base of AI services it sells and operates itself.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
