Skip to content

Tech

Microsoft Maia 300: 7x the throughput, and Nvidia gets a side-eye

Microsoft's third-generation Maia chip lands in 2027, with multi-gigawatt ambitions and a plan to run its own products mostly on its own silicon. Maia 200 only launched in January, so the roadmap is moving fast.

TL;DR

  1. Maia 300 hits production in 2027, Microsoft's third-generation in-house AI chip.
  2. Microsoft expects 40 percent more tokens per megawatt than leading GPU-based systems.
  3. Maia becomes the primary inference system for Microsoft products, at multi-gigawatt scale.

What happened

Microsoft announced Maia 300, the third generation of its in-house AI chip. It is aimed at inference, the part where a model answers you rather than learns. Production rollout starts in 2027.

The bigger claim is about scale. Microsoft says it will grow the Azure Maia family to multi-gigawatt deployments from 2027 and make Maia the primary inference system for its products and services.

Scott Guthrie, who runs Microsoft's cloud group, put numbers on it. He says Maia 300 should deliver 40 percent more tokens per megawatt than leading GPU-based systems.

That 7x is also Guthrie's claim, and it is measured against Maia 200, not against a rival. It is a vendor figure for a chip nobody outside Microsoft can run yet.

Speaking of Maia 200: it was only announced in January. It already serves Microsoft production workloads, including Microsoft Copilot, and helps the MAI team build next-generation models.

GitHub Copilot and Microsoft Foundry are due to follow by the end of the year. Microsoft says it runs OpenAI models, its own MAI models and leading open-source families.

Why it matters

Inference is where the running costs live. Every Copilot answer burns compute, and Microsoft sells a lot of Copilot. Owning the chip means owning more of the cost curve.

The Verge reads this as Microsoft becoming less reliant on Nvidia's GPUs, and says it should soon lean mainly on its own chips instead of those from Nvidia and AMD.

Microsoft's pitch is co-design: it sees how its AI services behave at scale, so it can tune hardware and software together.

Maia 300 is also pitched at newer workloads: long-context reasoning, latency-sensitive chat and sustained multi-step agent tasks. That is the shape of where AI products are heading, and the chip is built for it.

Microsoft wants its own silicon doing most of its own thinking.

Keep the caveats in view. The 40 percent figure is an expectation, not a benchmark anyone has independently run. The chip arrives in 2027. Slides are not racks.

What it means for Europe

Microsoft's rollout is described as starting in 2027, and European customers buy their AI through Azure like everyone else. If Maia delivers cheaper inference, that could show up in what enterprises pay for Copilot and Foundry workloads.

Nothing in the announcement gives a regional schedule, so treat the European timing as open. Ask your account rep where capacity lands, not just when.

What to do with it

  • Ask about pricing.If you run Azure AI workloads, ask your Microsoft contact how Maia-backed inference will affect your costs.
  • Watch the end-of-year milestones.GitHub Copilot and Foundry are due on Maia 200 by year end, so that is the first real signal.
  • Keep your stack portable.Maia runs OpenAI, MAI and open-source models, so avoid locking yourself to one model family.
  • Discount vendor numbers.Treat 40 percent and 7x as targets until independent tests show up.

Keep up, sapiens. The chips are now building their own pipeline.