MAIA 300 AT A GLANCE
● 288 GB of HBM4 at 18 TB/s. Eight 36 GB stacks running 9 Gb/s per pin.
● 13.8 PFLOPS FP4, 6.9 PFLOPS FP8. Main compute die on TSMC N2P — 2 nm — in a multi-chiplet design.
● 950 W per accelerator, up from 750 W on Maia 200. 96 per rack, scale-up clusters near 12,000.
● Not available yet. Deployment begins 2027. Reuters reports Microsoft could field more than a million chips next year.
Microsoft unveiled Maia 300 on 12 October, the third generation of the accelerator it designs for its own data centres. It is an inference chip, and the design choices make that obvious.
Full specifications
| Specification | Maia 300 |
|---|
| Generation | Third, succeeding Maia 200 |
| Compute (FP4) | 13.8 PFLOPS |
| Compute (FP8) | 6.9 PFLOPS |
| Memory | 288 GB HBM4, 8 stacks of 36 GB |
| Memory speed | 9 Gb/s per pin |
| Memory bandwidth | 18 TB/s |
| Process | TSMC N2P (2 nm) compute die; some supporting silicon on 3 nm |
| Packaging | Multi-chiplet, separate memory and I/O chiplets |
| Power | 950 W SoC TDP |
| Scale-up bandwidth | 3.2 TB/s bidirectional |
| Network | Two-tier Ethernet, custom AI transport protocol |
| Rack density | 96 accelerators |
| Cluster size | Around 12,000 accelerators |
| Availability | Deployment from 2027 |
What the numbers actually tell you
288 GB of memory at 18 TB/s is the whole point. Inference at scale is bound by how fast you can move weights and KV cache, not by raw arithmetic. A chip with this much high-bandwidth memory holds larger models and longer contexts resident without spilling across devices, which is where most of the latency goes in production serving.
FP4 at twice the FP8 rate says the same thing from another direction. This is built for quantised inference, not for training runs.
950 W, up from 750 W, is a 27% power increase per accelerator. At 96 per rack that is a meaningful change in what a rack draws, and it is why the efficiency claim below is the number Microsoft leads with rather than raw throughput.
The performance claims, and how much to trust them
Microsoft says Maia 300 delivers up to 40% more tokens per megawatt than leading GPU-based AI systems, and up to 7x the rack-scale token throughput of Maia 200 in a fully populated 96-accelerator rack.
The second claim is internally comparable and reasonable to take at face value — Microsoft is measuring its own two generations under conditions it controls.
The first is not checkable. Microsoft has not specified which GPU configurations it benchmarked against, at what precision, on what model, or with what software stack. "Leading GPU-based systems" is doing a lot of work in that sentence. Tokens per megawatt is also the metric that most favours a fixed-function inference part against a general-purpose GPU, which is the honest reason to choose it.
WHAT THIS IS REALLY ABOUT
Microsoft is not trying to sell you a chip. Maia 300 exists to reduce what Microsoft pays Nvidia and AMD for the inference that Copilot, Teams and Azure OpenAI consume. The relevant comparison is not Maia against a GPU on a benchmark — it is Microsoft's cost per served token in 2027 against what it would otherwise have paid. On that measure, a chip that is merely competitive is already a win.
What it means if you are buying AI, not building data centres
Nothing immediately — deployment starts in 2027 and you cannot rent one. The second-order effect is the one to watch: if hyperscalers meaningfully cut their inference costs, that shows up as cheaper API pricing or higher rate limits on the services built on top. Google has been running TPUs on this logic for years, and Amazon has Trainium and Inferentia doing the same job.
The chip to compare it against is not announced yet either. Nvidia's roadmap for 2027 will land in the same window, and that comparison — not the one against "leading GPU-based systems" today — is the one that decides whether Maia 300 was a good bet.
Sources
Notebookcheck — full specifications ·
UPI — Microsoft announces Azure Maia 300