Microsoft Maia 300 Chip Set for Multi-Gigawatt Rollout in 2027

Microsoft's third-generation chip gets a deployment plan

Microsoft has stopped treating its in-house AI silicon as an experiment. On October 12, the company announced the Maia 300, the third generation of its custom accelerators, and said it would start "multi-gigawatt deployments" of the chip next year.1112 Reuters went further than that and reported that Microsoft could deploy potentially more than a million of the chips in 2027 to run its internal AI services.1316

The chip is meant for inference-heavy work such as writing code and carrying out complex business tasks on behalf of users.11 The previous generation, Maia 200, already runs Microsoft's own models as well as OpenAI models served through Azure.16 Scott Guthrie, executive vice president of Cloud + AI, called the new system a "complete intelligent system" designed for long-context reasoning, latency-sensitive conversations and multi-step agentic tasks. Microsoft also said Maia will become the main inference platform behind Copilot, GitHub Copilot and Foundry.18

The reports describe the motive the same way: cut the cost of serving AI and loosen Nvidia's hold on the market. Where they differ is on how much Microsoft has actually committed to.

What the specs say, and what they don't

Maia 300 does not follow Nvidia's design. Nvidia's accelerators use one very large compute die surrounded by high-bandwidth memory. Microsoft instead built the Maia 300 processor from chiplets.16 The main compute chiplets use TSMC's 2-nanometer technology, and other chiplets that move data on and off the processor use cheaper 3-nanometer technology.19 Microsoft's system-level figures put eight HBM4 stacks on each chip, giving 18 TB/s of peak memory bandwidth at a 950-watt TDP. A rack holds 96 accelerators in a 167-kilowatt envelope with 27.6 TB of HBM.18

The main efficiency claim is roughly 40% more useful compute per megawatt than "leading GPU-based systems." Microsoft did not say which systems it compared against.1115 Rani Borkar, who runs Azure hardware systems and infrastructure, described the aim as more intelligence for every watt, every dollar and every byte of memory, achieved by optimizing the whole system.12

The rack figures need a careful reading. Reuters compared Maia's 96-chip, roughly 167 kW rack with Nvidia's GB300 racks, which hold 72 GPUs and draw about 142 kW.17 Those are different configurations. Power draw alone says nothing about how much work each rack does. The metric that tries to capture that, tokens per megawatt, comes from Microsoft.18 Microsoft's separate "up to 7x" figure is narrower still: it compares a full Maia 300 rack with Maia 200 on rack-scale throughput, not with Nvidia hardware.18

Where the coverage diverges

The biggest gap between reports is the chip count. Microsoft's own announcement promises "multi-gigawatt deployments" and gives no number of chips. The figure of more than a million comes from Reuters' reporting.1820 That matters for anyone trying to work out what this means for Nvidia's order book. A gigawatt target describes data center power, not a purchase order, and CNBC reported that a Microsoft general manager would not give a timeline for reaching it.20

Analyst estimates suggest a gradual shift rather than a sudden one. Evercore ISI expects Maia to account for about 10% of Microsoft's AI chip spending by the first quarter of 2027.18 Satya Nadella said back at the Maia 200 launch that Microsoft would keep buying chips from Nvidia and AMD even with its own silicon in hand.10 Taken together, Maia looks like a large and growing share of Microsoft's own inference fleet, while Nvidia remains its main supplier for training and general-purpose compute.

Some syndicated versions of the story mangled the timeline, with years that don't match Microsoft's 2027 deployment start or its 2026 capex guidance.13 Both points are clear in the primary reporting. Production deployments start in 2027, and Microsoft expects capital expenditures above $175 billion in calendar 2026.1718

The inference cost math

The cost argument rests on how the AI business is changing. One industry analysis estimates that inference now accounts for about two-thirds of AI compute spending.21 Running models, not training them, has become the recurring cost. In a data center, power turns fairly directly into operating cost, which is why Microsoft frames its claims per megawatt.20

Maia 200 shows where this is heading. When it launched in January, Microsoft said it delivered 30% better performance per dollar than the newest hardware then in its fleet.2 The chip was built on TSMC's 3-nanometer process with 216 GB of HBM3e at a 750-watt envelope.2 Microsoft now says Maia 200 is deployed across the United States, Europe and Australia and gets up to 1.7x more tokens per watt than leading GPU infrastructure on built-in AI features in Teams and Copilot. Again, it did not name the comparison system.18

That said, better efficiency does not automatically mean lower total spending, and it may never show up in customer prices. Microsoft has not published a Maia budget, a savings estimate or a payback period.20 No customer purchase path has been announced either, so Maia 300 is so far presented as infrastructure for Microsoft's own services.18 Analysts made a related point at the Maia 200 launch. If Microsoft moves heavy internal workloads like Copilot onto its own chips, it stops competing with its own Azure customers for Nvidia GPUs, which could indirectly increase the supply of GPUs available to rent.6

Nvidia's roadmap is the moving target

Nvidia keeps moving too. Its Vera Rubin NVL72 rack systems are listed as in production as of September 2026.26 Nvidia says the platform delivers far higher inference performance and much lower cost per token than Blackwell.8 Custom chips typically take two to three years to design, while Nvidia now releases new products every year. So hyperscalers can gain cost advantages and still fall behind on absolute performance.4

This is why the unnamed "leading GPU-based systems" in Microsoft's 40% claim matter so much. A lead over Blackwell-era racks in 2026 could shrink by the time Maia 300 hardware arrives in volume and has to compete with Rubin. Software is the other obstacle. One Forrester analyst compared Nvidia's CUDA and Microsoft's Maia SDK to two rail gauges that don't connect, so applications have to be adapted before they can move across.6

Everyone is building their own chips

Microsoft is following a path rivals opened years ago. Google has built TPUs since 2015, and its Ironwood generation came out of about a decade of work on the line.22 Google announced an eighth generation in April 2026, split into TPU 8t for training and TPU 8i for inference.26 One analysis of Google Cloud pricing concluded that Ironwood comes in about 18% cheaper than Nvidia's B200 on inference cost.28

Amazon's Trainium has the clearest external customer. Anthropic says it uses more than a million Trainium2 chips.26 Amazon is also reportedly close to buying the Israeli startup Decart for about $7 billion. That deal is aimed directly at the software gap that leads developers to choose Nvidia even when Trainium is better on price-performance.24 OpenAI joined the field with Jalapeño, an inference chip built with Broadcom.22 Meta has said it has deployed hundreds of thousands of its own MTIA chips.18

For a company that started from nothing about three years before Maia 200, Microsoft has moved fast.22 Neutral coverage still doesn't back the boldest prediction, that Nvidia's inference share will collapse. Forecasts that it could fall from around 90% to 20–30% by 2028 depend on hyperscalers moving their own internal workloads onto custom chips.21 Merchant GPUs are still the only option that runs across clouds.21

The takeaway

Maia 300 is best understood as Microsoft committing, at power-utility scale, to run its own AI products on chips it designs. It is not a direct challenge to Nvidia in the open market. That is a big deal for Microsoft's margins and a real dent in Nvidia's largest category of customers. Even so, every performance figure so far is Microsoft's own, measured against competitors it hasn't named, on hardware that won't ship until 2027. Microsoft says more details will come at Ignite next month.18 Until independent measurements exist, the 40% figure is a design target, and the million-chip number is a reported possibility, not a commitment.