Microsoft outlines plans for Maia 300 AI chip and multi-gigawatt deployment starting in 2027

Microsoft is laying out another step in its push to build custom chips for artificial intelligence workloads, unveiling plans for a new proprietary accelerator called Maia 300 and indicating that it intends to scale the chip across its data center footprint starting in 2027.

Maia 300 is described as a custom “inference accelerator.” Inference is the phase where trained AI models are used to run real tasks such as generating responses or producing recommendations, typically at very high volume. Microsoft’s emphasis on inference hardware reflects how companies often spend more on running models at scale than on training them.

According to the report, Maia 300 is intended to be manufactured using TSMC’s 2-nanometer process, a leading-edge semiconductor node associated with smaller transistors and, in principle, more efficient power and performance per chip. The choice of process technology matters for AI builders because data center energy and cooling constraints increasingly shape how quickly capacity can be expanded.

The announcement also points to a large rollout path, describing “multi-GW” deployment from 2027. “Multi-GW” refers to multi-gigawatt power levels, a scale that, if applied to AI compute, implies Microsoft expects to ramp electricity-hungry infrastructure in step with its chip strategy. The report does not provide a breakdown of how many Maia 300 devices would be involved, nor how that power target translates into compute capacity.

Microsoft has been building toward a broader mix of AI hardware and software stacks, aiming to reduce reliance on externally sourced accelerators for key parts of its AI services. Custom chips are often pursued to improve cost efficiency, manage supply constraints, and tailor performance to specific model-serving pipelines.

In this latest update, Microsoft did not provide additional detail in the cited post, such as target throughput, memory specifications, interconnect design, software compatibility layers, or whether Maia 300 is intended primarily for Microsoft’s internal deployments or also to support customer workloads. The report also does not specify which AI models or services will be prioritized for the early rollout window.

There is also no information in the cited material about the exact timeline for sampling, production ramp, or qualification testing, all of which are usually critical for chips destined for large data center deployments. Without those specifics, it is unclear how quickly Microsoft expects Maia 300 to reach full scale during 2027.

Investors and competitors will likely focus next on whether Microsoft provides more granularity around the multi-gigawatt target, including how it measures deployment progress and the role of Maia 300 relative to other AI accelerators in its portfolio. Additional disclosures on performance-per-watt, deployment geography, and software tooling would also help clarify how the chip fits into Microsoft’s longer-term AI infrastructure plans.