Microsoft Maia 300 unveiled — TSMC 2nm, eight HBM4 stacks, its main AI inference chip from 2027
Microsoft introduced Maia 300, its third-generation in-house AI chip, on October 12, 2026. The compute chiplets use TSMC's 2nm-class N2P process, the chip carries eight HBM4 memory stacks with 18 TB/s of bandwidth, and it draws 950 W. A 96-chip rack uses about 167 kW. Microsoft claims 40% more tokens per megawatt than leading GPU systems and up to seven times the rack-level throughput of Maia 200, without naming the comparison systems. From 2027, Maia will be deployed at multi-gigawatt scale as the primary inference system for Microsoft's AI services.
The three lines
- Chip — Maia 300: TSMC N2P compute chiplets, eight HBM4 stacks at 18 TB/s, 950 W (Maia 200: 750 W)
- Claims — 40% more tokens per megawatt than leading GPUs; 96-chip rack up to 7x Maia 200 throughput. Comparison unnamed, not independently verified
- Timeline — multi-gigawatt deployment from 2027 as Microsoft's main inference system; TSMC talks for 300,000+ chips reported in August
Key questions
- Microsoft Maia 300 specs
- **A TSMC 2nm-class, 950 W inference chip with eight HBM4 stacks.** | Item | Maia 300 | |---|---| | Generation | Third | | Process | Compute chiplets on TSMC N2P; some support silicon on 3nm | | Memory | 8 HBM4 stacks, 288 GB (reported), 18 TB/s | | Compute | Up to 13.8 PFLOPS FP4 (reported) | | Power | 950 W | | Rack | 96 chips, ~167 kW, 27.6 TB HBM |
- Maia 300 vs Nvidia
- **Microsoft claims 40% more tokens per megawatt but did not say which GPU system it compared against.** | Item | Maia 300 rack | Nvidia GB300 rack (Reuters) | |---|---|---| | Chips | 96 | 72 GPUs | | Power | ~167 kW | ~142 kW | | Efficiency claim | +40% tokens per MW | Baseline | | Verification | Company claim only | — |
- When will Maia 300 be available
- **Production deployments start in 2027. Maia 200 is running today.** | When | What | |---|---| | Now | Maia 200 serves Copilot and MAI model work in the US, Europe, Australia | | End of 2026 | GitHub Copilot and Foundry move to Maia 200 | | November 2026 | More detail at Ignite | | 2027 | Maia 300 production deployment as primary inference system |
From 2027, Microsoft's AI services will run mainly on its own silicon. In an October 12, 2026 blog post by cloud and AI chief Scott Guthrie, Microsoft unveiled Maia 300, the third generation of its in-house AI chip. The compute chiplets are made on TSMC's 2nm-class N2P process. Starting in 2027, Microsoft plans "multi-gigawatt" deployments with Maia as the "primary inference system" for products such as Copilot and Teams.
The same day, The Information reported Microsoft is on track to pay Anthropic more than $2 billion a year for Copilot models. Microsoft may pay others for models, but it intends to cut the power and chip cost of running them itself.
1. What Maia 300 is
| Item | Maia 300 | Maia 200 |
|---|---|---|
| Generation | Third | Second |
| Design focus | Inference-first | Inference (Copilot, MAI development) |
| Compute process | TSMC N2P (2nm-class) | — |
| Structure | Chiplets separating compute, memory control and I/O | — |
| Memory | 8 HBM4 stacks, 18 TB/s | — |
| Power | 950 W | 750 W |
| Low precision | Enhanced MXFP4 | — |
| Deployment | From 2027 | Now: US, Europe, Australia |
The chiplet design puts only the compute section on the costly 2nm node; memory control and I/O go on cheaper nodes (3nm, per Reuters). Our explainer on 2nm processes covers what actually changes at that node. On-die SRAM acts as a software-managed scratchpad, and 4-bit MXFP4 formats cut memory and bandwidth needs — aimed at the "KV cache" that piles up in long conversations and multi-step agent work.
2. Rack-level numbers and claims
| Item | Maia 300 rack | Nvidia GB300 rack (Reuters comparison) |
|---|---|---|
| Chips | 96 | 72 GPUs |
| Power | ~167 kW | ~142 kW |
| HBM capacity | 27.6 TB | — |
| Aggregate HBM bandwidth | 1.7 PB/s | — |
| Efficiency claim | 40% more tokens per MW vs. leading GPU systems | — |
| Generational claim | Up to 7x Maia 200 rack throughput | — |
Microsoft did not name the GPU system behind the 40% figure, and there is no independent test yet. Still, the choice of metric — tokens per megawatt, not raw chip speed — is telling: the data-center bottleneck has shifted from chip supply to electricity. Rani Borkar, who runs Azure hardware, told Reuters Maia 300 is "delivering more intelligence for every watt, every dollar, and every byte of memory."
3. Why now, and what changes
| Issue | Detail |
|---|---|
| Nvidia dependence | After Google (TPU), Amazon (Trainium) and OpenAI (with Broadcom), Microsoft moves main inference to its own chip |
| Scale | Multi-gigawatt from 2027; 300,000+ chips discussed with TSMC (The Information, August) |
| Models served | OpenAI models, Microsoft MAI models, major open-source families |
| Korean memory makers | Eight HBM4 stacks per chip; supplier undisclosed, but a demand driver |
| Next disclosure | Microsoft Ignite, San Francisco, November 2026 |
The news is less "Microsoft makes a chip" than a public date for making it the default. As in-house chips multiply, the hard part becomes getting everyone's models to run well on them — the same reason Amazon is reportedly negotiating to buy Decart.
4. What remains unconfirmed
- Comparison baseline: The "leading GPU systems" behind the 40% claim are unnamed.
- Detailed specs: 13.8 PFLOPS FP4 and 288 GB are press figures, not in Microsoft's blog.
- Memory supplier: Undisclosed.
- Volume: 2027 output is based on reports and estimates; Ignite or TSMC's earnings may offer clues.
Sources
- Official Microsoft Blog — Microsoft to expand Azure Maia system to multi-gigawatt deployments (2026-10-12)
- Azure Blog — Deep dive into Azure Maia 300 architecture
- The Next Web — Microsoft says Maia 300 chip will run its AI services from 2027
- Notebookcheck — Microsoft Maia 300 AI chip arrives with 288 GB HBM4 memory and 18 TB/s bandwidth
- Interesting Engineering — Microsoft Maia 300 96-chip rack hits 40% more tokens per megawatt