Maia 300 Chip Moves Copilot and Teams Inference In-House From 2027

What Microsoft announced

On October 12, Microsoft said it will run most of its own AI services on chips it designs itself. Scott Guthrie, head of the company's Cloud + AI division, introduced Maia 300, the third generation of Microsoft's in-house AI accelerator. He said the Maia family will become the main inference system for Microsoft products starting in 2027, in deployments sized at several gigawatts of power.49 Satya Nadella called it the company's "biggest bet yet on our own silicon."23

The announcement had two parts. The first is about the present: Maia 200, launched in January in a single data center region near Des Moines, Iowa, is now running production workloads, including Microsoft Copilot. It is being rolled out across the United States, Europe and Australia.29 GitHub Copilot and Microsoft Foundry, the Azure service businesses use to deploy models and agents, are due to move onto Maia 200 by the end of 2026.46 The second part is the roadmap: Maia 300 will start production deployments in 2027. Microsoft expects it to produce 40% more tokens per megawatt than leading GPU-based systems.69

For enterprise buyers, the useful part is not the chip specs. Microsoft is saying it wants to lower the cost of every Copilot answer, meeting summary and agent step its customers trigger. That matters a lot right now, because Copilot seat sales are growing much faster than the infrastructure behind them can be built.

The hardware, briefly

Coverage of the specs is consistent across outlets. Each Maia 300 chip has eight stacks of HBM4 memory, for 288 GB of capacity and 18 TB/s of bandwidth. That is 2.5 times Maia 200's bandwidth and 33% more capacity.28 The compute chiplets are made on TSMC's N2P process. Power per chip rises from 750 W to 950 W.2 A full rack holds 96 accelerators within a 167 kW power budget, and Microsoft says that rack produces up to seven times the token output of a Maia 200 rack.710

That sevenfold figure is easy to misread. One outlet points out that a Maia 300 rack holds twice as many chips as a Maia 200 rack, so the comparison includes the doubled density, not just better chips.2 Raw compute also grows much less than memory: about 36%, from just over 10 to 13.8 petaFLOPS at FP4.2 Microsoft describes the design as built first for inference, aimed at long contexts, memory-heavy work and decode-heavy tasks.4 In plain terms, it is built for long conversations and multi-step agent tasks, which are taking up a growing share of enterprise AI usage.

Where the coverage diverges: speed versus capacity

The biggest split in the reporting is over what users will actually notice. Some outlets framed Maia 300 as something that will make Teams and Copilot feel quicker, saying Microsoft hopes to cut the wait for automated responses.18 Others were more careful. They noted that the 40% and sevenfold numbers measure total throughput, not how fast any one person gets an answer, and that Microsoft has not published measured customer-facing improvements for Copilot or Teams on the new chip.6 One report stated it directly: Copilot looks the same whichever chip serves it, and the gain shows up as a lower cost per answer for Microsoft.2

The careful reading is the right one. Inference chips tuned for tokens per watt are mainly a capacity and cost tool. They let a data center with a fixed power supply handle more simultaneous requests or heavier features.610 Response time still depends on the model, the task and the surrounding system.6 For CIOs, the realistic hope is fewer throttling limits and room for more demanding agent workflows, not snappier chat.

Several outlets also stressed that the efficiency claims are Microsoft's own. The company has not named the GPU systems it compared against or explained how it measured the 40% and 1.7x figures.25 One analysis described the Maia 200 number as a present-tense claim with no published method behind it.5 Until independent benchmarks or customer-visible changes appear, these numbers should be treated as projections.

Why the timing lines up with Copilot demand

The chip news lands just as Copilot has moved from pilots into routine budgets. Microsoft has disclosed paid Microsoft 365 Copilot seat counts on three straight earnings calls: about 15 million in January, 20 million in April and more than 30 million at the July 29 report for fiscal Q4 2026.1920 Adding ten million seats in one quarter equaled the entire paid base Microsoft had reported only two quarters before.19 Microsoft also said the number of customers with more than 50,000 seats rose more than sevenfold year over year.19 In April it reported that queries per seat had climbed nearly 20% from the previous quarter.16

The early seat figures differ slightly by source. One analysis cites 16.1 million paid seats as of December 31, 2025, while others round that quarter to 15 million.1714 The overall trend is the same either way. More seats, each sending more queries, mean much more inference work, and that is exactly the workload Maia is built for. Reading Maia as a response to Copilot's growth is analysis, but it fits the numbers.

The large deployments that make up the case for Copilot's return on investment add to that demand. Accenture committed to more than 740,000 seats, the largest single Copilot deal Microsoft has disclosed.1618 Bayer and Johnson & Johnson each have more than 90,000 seats, according to Microsoft's Q3 commentary.16 Infosys, TCS and Wipro together account for more than 300,000 seats; Infosys reports monthly active use above 91%, and TCS reports daily AI use at 86%.12 Microsoft describes a recurring pattern in which customers pilot with hundreds of seats and come back for tens of thousands once they see productivity gains.19

The ROI gap that hardware can't close

Those headline deployments are not the whole picture. Even at 30 million seats, Copilot covers only about 6.5% of Microsoft's commercial Microsoft 365 base, which independent analysts estimate at around 464 million seats.1119 Estimates of that base vary, from 415 million to 464 million depending on the source and date, so penetration figures move with them.1719

The data on how much people actually use Copilot is less flattering. Recon Analytics found that only about 35.8% of employees with Copilot access actively use it.15 It also reported a negative accuracy net promoter score of -19.8, and found that 44.2% of lapsed users cited distrust.11 Forrester reports that most enterprises are still in pilot mode, 12 to 18 months from scaled deployment, with data readiness and ROI measurement as the main obstacles.15 Some analysts estimate that common enterprise discounts of 40% to 60% push real Copilot revenue well below what list prices suggest.14

This is where Maia's limits show. A cheaper token helps Microsoft's margins and its ability to serve heavy users. It does nothing for adoption quality, which one analysis says now depends on customers' IT operating models: who gets licenses, which data is ready, and which workflows actually produce different output.18 Seat counts tell you about purchases, not results, and usage depth is the measure most enterprises are not yet tracking.20

The Nvidia question and the pricing question

Maia is also a step toward depending less on Nvidia.67 Reuters reported that Microsoft could deploy more than one million Maia 300 chips in 2027.6 Evercore ISI estimates custom Maia silicon will make up about 10% of Microsoft's AI hardware budget by early 2027.7 Both figures point to a significant but not dominant share. Microsoft has not said Maia will replace Nvidia hardware altogether, and GPUs are staying in its data centers.26 One analyst described the progression this way: Maia 100 was a proof of concept, Maia 200 fixed its problems and validated the design, and Maia 300 is where Microsoft chose to "go big."4

The question that matters most to buyers is still open. Microsoft has not said whether any of its inference savings will show up in Copilot pricing or usage limits.2 It also has not said how many gigawatts "multi-gigawatt" means, which regions will get the capacity, what the program costs, or whether Azure customers will be able to rent Maia 300 directly.2510

The takeaway

Maia 300 strengthens Microsoft's position on cost and supply in enterprise AI. It does not, by itself, deliver a product upgrade. The company is preparing for a future where agentic, long-context workloads grow the way seat counts already have, and owning the inference hardware gives it more control over margins and capacity.10 For enterprises in the middle of AI projects, the practical point is that Microsoft's per-query costs are likely to fall faster than their own adoption problems get solved. The chip can make Copilot cheaper for Microsoft to run, but customers still have to show that it is worth paying for.