Microsoft plans to scale its own Maia AI chips to gigawatt levels by 2027

Microsoft executive Scott Guthrie wrote that the company plans to deploy its in-house Azure Maia chips at multi-gigawatt scale starting in 2027 and make them the primary AI inference system across the company's products. This is a confirmed planning statement, but "multi-gigawatt" itself is a future target, and whether it can land on schedule is an execution-level uncertainty.

Running parallel to the far-off 2027 node is a deployment that has already happened: Maia 200 is currently in use in Copilot operations across the US, Europe, and Australia. GitHub Copilot and Foundry are expected to follow with Maia chips by the end of this year. In other words, Microsoft's inference chip swap is not a one-time switch — it starts on the single Copilot product line, then spreads to more products.

Putting these two nodes together, the 2027 multi-gigawatt target is not a jump out of nowhere — it builds on Maia 200 already running in actual products. But the report does not say the actual scale or share of Maia 200 usage in Copilot, so there is still a considerable distance between "already in use" and "becoming the workhorse."

Microsoft CEO Satya Nadella called this "our biggest bet yet on custom silicon." This is a qualitative statement from the company's top leadership, a source-relayed claim rather than a verifiable technical metric. The signal it sends is that, inside Microsoft, in-house inference chips have been placed at a very high strategic priority.

On scope of application, the chip can handle OpenAI, MAI, and open-source models. This is worth noting because it means Maia does not only serve Microsoft's own models, but is meant to cover multiple model sources hosted on Microsoft cloud. However, the report does not quantify the degree of support or performance for these models, so "can handle" can only be understood as a compatibility-level statement.

Successor Maia 300 will ship in 2027, targeting long inputs, fast conversations, and agentic tasks. These three directions point to scenarios in inference workloads that increasingly consume memory capacity and bandwidth: long context needs to fit, fast conversations need low latency, and agentic tasks often require multi-turn, sustained inference.

The confirmed specs are: each Maia 300 chip integrates 8 HBM4 memory stacks, 288 GB total capacity, 18 TB/s bandwidth, and 950 watts rated power. This set of numbers is the "hardest" information in this article, because they are concrete hardware parameters rather than relative performance comparisons.

But specs themselves do not equal actual performance. The 950-watt rated power means the per-chip power budget is quite high, and the report does not say the chip manufacturing process, foundry, or supply chain arrangements, so cost or final energy-efficiency outcomes cannot be inferred from these parameters.

Around energy efficiency and output, Microsoft gave a set of comparison figures. In the Teams scenario, Microsoft says Maia chips can output up to 1.7 times more tokens per watt than leading GPUs. For Maia 300, Microsoft claims it can output 40% more tokens per megawatt than leading GPU systems. At the rack level, a 96-chip rack can produce up to 7 times the tokens of a Maia 200 rack.

All three figures need to be treated with caution. They are all Microsoft's unilateral claims, the report provides no benchmarking details, and there is no independent third-party verification. Especially for the 96-chip rack 7x token output comparison, neither the comparison baseline nor the test conditions are disclosed — whether it is a comparison at equal power or at equal chip count would directly change the meaning of this multiple.

Therefore, these numbers are better read as targets and confidence Microsoft is projecting externally, not as verified conclusions. Maia 300's specs and performance targets are aimed at 2027 shipment and have not yet been realized; whether the claimed per-megawatt token advantage can be delivered under actual workloads is likewise still uncertain.

The information disclosed so far comes from a technical report, with more to be announced at Ignite 2026 in November 2026. This means what we see now is a staged disclosure: hardware specs are relatively complete, but the methodology behind the performance comparisons, actual deployment scale data, and horizontal comparisons with other acceleration options are all left for later.

This "conclusions first, details later" cadence itself constitutes an information constraint. Before Ignite 2026, any judgment about Maia 300's actual competitiveness can only be built on Microsoft's own statements.

It must be made clear that all the above facts come from a single reporting source, THE DECODER, whose content mainly relays Microsoft's statements. This means we are not dealing with information cross-verified by multiple parties, but a transmission chain from Microsoft to a single outlet.

On this chain, three layers can be distinguished: hardware specs (288 GB, 18 TB/s, 950 watts) and the deployment timeline (2027, Maia 200 already in use in Copilot across the US, Europe, and Australia, GitHub Copilot and Foundry to follow by year-end) are confirmed factual statements; Nadella's quote and all performance multiples (1.7x, 40%, 7x) are source claims; and "whether Microsoft can achieve multi-gigawatt-scale deployment on schedule in 2027" and "the impact of the custom silicon strategy on the GPU procurement landscape" are inferences, and the report itself does not provide answers.

Only by separating these three layers can we avoid treating Microsoft's targets as accomplished facts. The strategic intent of this roadmap is clear, but the verifiable evidence of execution is currently still concentrated in Maia 200's limited deployment.