Microsoft targets gigawatt-scale Maia 300 AI infrastructure from 2027
Microsoft has outlined plans for Azure Maia 300, a custom AI accelerator designed as part of an integrated system spanning chips, software, racks, networking and data centres. Production deployments are planned for 2027, when Microsoft says it will begin scaling Maia infrastructure to multi-gigawatt capacity, giving the company greater control over the computing stack behind its AI services.
Microsoft claims a rack with 96 Maia 300 accelerators could deliver up to seven times the token throughput of a Maia 200 rack and 40 per cent more tokens per megawatt than leading GPU-based systems. Each accelerator will have eight HBM4 memory stacks and up to 18 TB/s of peak memory bandwidth, aimed at long-context and memory-intensive inference. Microsoft says Maia 200 is already serving production workloads and plans to share further details at its November 2026 Ignite event.
- Microsoft plans Maia 300 production deployments in 2027.
- A 96-accelerator rack is claimed to offer seven times Maia 200 throughput.
- The rollout is planned to reach multi-gigawatt scale.
New here? Start with this
AI models need large amounts of computing power to learn from data and to respond to users. An accelerator is a specialised chip built to handle those calculations; Microsoft’s Maia chips are designed for use in its Azure cloud services, which provide computing capacity to businesses and other customers.
Microsoft is developing Maia 300 as part of a wider system that includes software, networking and data centres. Its earlier Maia 200 is already being used for some production workloads, while other cloud providers rely heavily on chips from suppliers such as Nvidia.
The plans matter because the cost and availability of computing power can affect how quickly companies expand AI services and what those services cost to run. Building its own chips could give Microsoft more control over its infrastructure, though the performance figures cited so far are the company’s own claims.
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
Building its own AI infrastructure could give Microsoft more control over performance, supply and energy use as demand for AI services grows. If its claims about throughput and tokens per megawatt hold up in production, Maia 300 could make computing more efficient and reduce reliance on scarce third-party accelerators, potentially supporting more capable services at scale.
The case against
Scaling custom AI infrastructure to multi-gigawatt capacity could lock in substantial spending and energy demand before the benefits are proven at production scale. The performance figures are Microsoft’s own claims, and a system designed around its chips and software may offer less flexibility than relying on a wider market of hardware suppliers, while concentrating more control over AI computing in one company.
Read the full article at the source →
Originally published by The Register as “Microsoft plans multi-gigawatt rollout of next-gen Azure Maia AI accelerator racks”.