Microsoft Maia 300 AI chip: 288 GB of HBM4, 2nm compute and a 40% tokens-per-megawatt claim

On October 12, 2026, Microsoft unveiled the Maia 300, the third generation of its in-house AI accelerator. In a post on the Official Microsoft Blog, Scott Guthrie, executive vice president of Cloud + AI, said the chip will go into Microsoft's multi-gigawatt AI datacenter deployments starting in 2027 and will become the primary inference platform for Microsoft's products and services. Microsoft's headline claim is that Maia 300 delivers 40% more tokens per megawatt than leading GPU-based systems. That is a vendor claim, and Microsoft did not say which GPU systems it compared against.

This story is based on Microsoft's announcement as reported by several outlets on October 12, plus earlier Reuters reporting.

The short version

• What it is: an inference-first AI accelerator, with compute chiplets on TSMC's N2P (2nm-class) process and eight HBM4 stacks, 288 GB in total.
• Bandwidth and power: 18 TB/s peak memory bandwidth and a 950 W TDP, up from 7 TB/s and 750 W on Maia 200.
• Rack: 96 accelerators per rack in a 167 kW envelope, with 27.6 TB of HBM.
• Claims (vendor-reported): 40% more tokens per megawatt than leading GPU systems; up to 7x the rack-scale token throughput of Maia 200.
• When: production deployments start in 2027. Microsoft hasn't said when, or whether, Azure customers can rent Maia 300 directly.

What is Maia 300?

Microsoft describes the chip as built around "an inference-first architecture optimized for long-context, memory-intensive and decode-heavy AI workloads," according to Mobile World Live's account of Guthrie's post. Those are the workloads that dominate when a chatbot or agent generates long answers or works through multi-step tasks: generating each token means reading the model's weights and its growing key-value (KV) cache from memory. Techzine quotes Microsoft's announcement making the same point: "Moving data efficiently, rather than raw compute power, is increasingly the bottleneck in large inference systems."

That explains where the spec sheet puts its emphasis. Memory bandwidth more than doubles over Maia 200, while power rises by 200 W.

Maia 300 specs

Last verified: October 12, 2026. Maia 300 figures are from Microsoft's October 12 blog post and its Azure architecture post. Maia 200 figures are from Microsoft's January 2026 launch, as summarized by Unite.AI.

Notebookcheck also reports 13.8 petaFLOPS of FP4 compute, 6.9 petaFLOPS of FP8, and 3.2 TB/s of bidirectional scale-up bandwidth per chip. Other outlets' accounts of the announcement don't include those figures, so treat them as reported, not confirmed.

For networking, Microsoft uses its own Ethernet-based AI transport protocol for scale-up, which it says can extend from a single rack across a datacenter. Microsoft says a two-tier network builds superpods of up to 12,288 accelerators.

How good is the 40% claim?

Microsoft says Maia 300 delivers "40% more tokens per megawatt" than "leading GPU-based systems." Tokens per megawatt measures how much model output a fixed power budget produces, which matters more as power, not chips, limits how fast new AI capacity comes online (see our tracker of gigawatt-scale datacenter projects).

What the claim doesn't tell you:

• No named baseline. Microsoft didn't name the GPU, and the post did not mention Nvidia, according to Mobile World Live. Notebookcheck notes that Microsoft hasn't disclosed enough about the GPU configurations to check the claim.
• No workload details. Tokens per megawatt depends on the model, batch size, context length and precision. Microsoft didn't publish those conditions.
• No independent results yet. No third party has published Maia 300 measurements, and the chip won't be in production until 2027. By then it will face Nvidia's Vera Rubin, which already runs customer workloads (see our Vera Rubin coverage).

The 7x rack-throughput figure compares a fully populated 96-accelerator Maia 300 rack with Maia 200. It is also vendor-reported.

Where Maia 200 stands

Maia 300's predecessor is already in service. Microsoft says Maia 200 runs production workloads including Microsoft Copilot and OpenAI models on Azure. TechRadar reports that GitHub Copilot and Microsoft Foundry are expected to follow before the end of 2026. Microsoft also claims Maia 200 delivers up to 1.7x more tokens per watt than leading GPU infrastructure for AI features in Teams, Copilot and Cowork, per Techzine. Again, that is a vendor figure without a named baseline.

How many chips?

Microsoft hasn't given a volume. Two sets of reports have:

• August 10, 2026: Reuters, citing The Information, reported that Microsoft planned to unveil Maia 300 "this fall." TrendForce's August 11 summary of those reports said Microsoft was "eyeing capacity for more than 300,000 Maia 300 chips for 2027 delivery" and was in talks with TSMC. That describes capacity Microsoft was seeking, not confirmed orders.
• October 12, 2026: Mobile World Live reported that, according to Reuters, Microsoft could deploy more than 1 million of the chips next year. Notebookcheck also attributes that figure, and the use of 3nm silicon for some supporting chips, to Reuters.

Neither figure comes from Microsoft. The two may also count different things, such as wafer capacity versus total units across chiplets, so don't read them as a revision from 300,000 to 1 million.

Why it matters

Every major cloud now designs its own accelerators: Google with TPUs, Amazon with Trainium, and Microsoft with Maia. The goal is less dependence on Nvidia and a lower cost per token for services they run at enormous scale. Maia 300 is Microsoft's clearest statement yet that its own silicon, not bought GPUs, will serve most of its first-party inference from 2027. The chip is aimed at inference, as CNBC's headline put it; Microsoft hasn't said it will stop buying GPUs for other work.

For readers, the practical effect is indirect. If Microsoft's efficiency claims hold, the cost of running Copilot-style features falls, which affects pricing and usage limits. If you want a refresher on what dedicated AI silicon does differently from a general-purpose processor, start with our NPU explainer.

What's confirmed, reported and uncertain

• Confirmed by Microsoft (October 12 blog and Azure architecture posts): the name and generation; N2P compute chiplets; eight HBM4 stacks with 288 GB in total; 18 TB/s; 950 W; the 96-accelerator, 167 kW, 27.6 TB rack; deployments from 2027; Maia as Microsoft's primary inference platform.
• Vendor claims, not independently verified: 40% more tokens per megawatt than leading GPU systems; 7x rack throughput over Maia 200; Maia 200's 1.7x tokens per watt.
• Reported, not confirmed by Microsoft: the FP4/FP8 figures; capacity for 300,000+ chips (The Information/Reuters, August); possible deployment of 1 million+ chips (Reuters, October).
• Unknown: the GPU baseline, whether Azure customers will be able to rent Maia 300 directly, and which datacenter regions get it first.