🚨 Microsoft's Maia 300 Chip Claims 40% More AI Per Megawatt Than GPUs — and It's Headed for Gigawatt Scale
Microsoft's third-gen Maia 300 AI chip claims 40% more tokens per megawatt than GPU systems, with multi-gigawatt deployment from 2027.
Microsoft just unveiled Maia 300, its third-generation homegrown AI chip, and it says the thing squeezes out 40% more tokens per megawatt than leading GPU-based systems (TechRadar).
The announcement landed on Monday, October 12. Deployment starts in 2027, and Microsoft says it plans to scale the Maia family to multi-gigawatt deployments (The Information).
Translation: one of the biggest buyers of AI chips on the planet wants more of its AI answering done on its own silicon.
⚡ What Happened
Maia 300 is built for inference, the part of AI where a trained model actually answers your prompt. Microsoft calls it an “inference-first architecture” (TechRadar).
The headline specs:
- Compute chiplets on TSMC’s N2P process, with memory and I/O split onto separate chiplets (Notebookcheck)
- 288GB of HBM4 across eight stacks, with 18 TB/s of bandwidth, in a 950W power envelope (Notebookcheck)
- 96 chips per rack, and up to 7x the token throughput of a Maia 200 rack (TechRadar)
For context, the Maia 200 came with 216GB of slower HBM3E and 7 TB/s. So this is a big jump in memory, which is exactly what long-context reasoning and AI agents eat for breakfast.
Maia 200 is already running production work, including Microsoft 365 Copilot. Maia 300 is slated for GitHub’s coding agents, Microsoft Foundry model serving, Copilot, Teams and Cowork.
🧠 Why It Matters
Every answer Copilot gives you costs Microsoft electricity and chip time. At Microsoft’s scale, 40% more output per megawatt is real money, and power has become the scarcest resource in AI.
It also changes the shopping list. CNBC framed it plainly: Microsoft will rely more on homegrown chips and less on Nvidia (CNBC). Google has TPUs. Amazon has Trainium. Microsoft, late to this party, is now bringing a much bigger dish.
Nvidia isn’t going anywhere, and Maia isn’t pitched as a training chip. But inference is where usage grows, and that’s the market Maia is aimed at.
🔜 What To Watch
- Independent benchmarks. Microsoft hasn’t said which GPU system its 40% claim is measured against, so nobody outside the company can check it yet.
- The gigawatt number. “Multi-gigawatt” is the plan. An actual figure, and a chip count, would tell you how serious it is.
- Nvidia’s reaction, on the stock chart and in its next earnings call.
- Whether Azure customers get access. For now, Maia mostly powers Microsoft’s own services.
Some details, including full compute specs, are still emerging.
Sources
- TechRadar: Microsoft unveils Maia 300, its most powerful AI chip yet
- Notebookcheck: Microsoft Maia 300 arrives with 288GB HBM4 and 18 TB/s
- CNBC: Microsoft to rely more on homegrown AI chips, reducing dependence on Nvidia
- The Information: Microsoft debuts new homegrown AI chip, plans to deploy multiple gigawatts next year
- Microsoft: Microsoft to expand Azure Maia system to multi-gigawatt deployments
❓ Quick answers
What is Microsoft Maia 300?
It's Microsoft's third-generation in-house AI accelerator, unveiled on October 12, 2026 and built for inference, with 288GB of HBM4 and 18 TB/s of bandwidth (Notebookcheck).
When will Microsoft deploy Maia 300?
Deployment begins in 2027, and Microsoft plans to scale the Maia family to multi-gigawatt deployments (The Information, TechRadar).
Is Maia 300 faster than Nvidia GPUs?
Microsoft claims up to 40% more tokens per megawatt than leading GPU-based systems, but it hasn't named the comparison system, so the claim isn't independently verified yet (TechRadar).
