Microsoft announced the Azure Maia 300 on Monday, October 12, and CEO Satya Nadella did not undersell it. In his words, it is the company’s “biggest bet yet on our own silicon.” The deployment plan backs up the language: Microsoft will scale the Maia family to multi-gigawatt deployments starting next calendar year, and Reuters reports the company plans to deploy potentially more than one million of the chips in 2027 to power its internal AI services.
The Maia 300 is the third generation of Microsoft’s custom AI accelerators. The first, Maia 100, appeared in 2023; the Maia 200 followed in January 2026. Now, barely nine months later, Microsoft has shown its hand again. Production deployments of the Maia 300 begin in 2027, and the company says it will become “the primary AI inference system for Microsoft products and services”, the infrastructure behind Copilot, Microsoft 365, Teams, and everything else hundreds of millions of people use every day.
What is inside the Maia 300
The chip uses a multi-chiplet design, and the manufacturing split is the interesting part. The compute chiplets are built on TSMC’s N2P process, a 2-nanometer-class node, while the memory-controller and I/O chiplets use a cheaper 3-nanometer process. The logic is straightforward: put the leading-edge silicon only where the math happens, and use older, cheaper processes for everything else.
| Specification | Maia 300 | Maia 200 |
|---|---|---|
| Compute performance (FP4) | 13.8 PFLOPS | Not disclosed |
| Compute performance (FP8) | 6.9 PFLOPS | Not disclosed |
| Memory | 288 GB HBM4 | 216 GB HBM3E |
| Memory bandwidth | 18 TB/s | 7 TB/s |
| SoC power | 950 W | 750 W |
| Accelerators per rack | 96 | Not disclosed |
| Rack power | About 167 kW | Not disclosed |
| Scale-up clusters | Up to 12,000 accelerators | Not disclosed |
| Production deployment | 2027 | 2026 |
The memory system is the headline: eight stacks of HBM4, 288 GB in total, pushing 18 TB/s of bandwidth at 9 Gb/s per pin. Against the Maia 200, that is roughly a third more capacity and about two and a half times the bandwidth.
Why the memory system is the real story
AI model inference is increasingly a memory problem, not a compute problem. Raw FLOPS look good on a slide, but when a model is actually answering questions, the bottleneck is moving model weights and the KV cache, the model’s working memory of the conversation, between the processor and memory, over and over, for every token.
That is why the Maia 300’s design brief reads like a wish list for long-context reasoning and AI agents: big memory capacity, enormous bandwidth, and networking (3.2 TB/s bidirectional scale-up bandwidth over two-tier Ethernet) built for chaining thousands of accelerators together. Agent workloads, where a model takes dozens of steps and each step needs the full context of the last, are exactly the kind of traffic that punishes memory-starved hardware.
There is an irony here worth noting. The same AI-driven memory hunger is now showing up in consumer prices: Google just raised the Pixel 10a to $599 without changing a single spec, and the culprit is AI’s appetite for RAM. The Maia 300 is Microsoft’s answer from the other side of the same equation: if memory is the scarce resource, design the chip around it.
The metric Microsoft wants you to remember: tokens per megawatt
Microsoft’s headline claim is efficiency, not raw speed: the Maia 300 is expected to deliver 40% more tokens per megawatt than “leading GPU-based systems.” Rani Borkar, president of Azure hardware systems and infrastructure, told Reuters the goal is “delivering more intelligence for every watt, every dollar, and every byte of memory by optimizing across the entire system.”
Take the number with the usual vendor-claim caution. Microsoft has not named the GPU systems it compared against, and independent testing will be needed before the 40% figure means much. The same applies to the claim of up to seven times the rack-scale token throughput of the Maia 200. What is concrete is the rack math: 96 accelerators drawing about 167 kW, against Nvidia’s most recent GB300 racks, which pack 72 GPUs into about 142 kW, according to Reuters.
What this means for Nvidia
The Maia 300 lands in the middle of an industry-wide move away from Nvidia’s grip on AI compute. Google has been building its own TPUs for years (and is now putting AI data centers in orbit), OpenAI recently debuted its own in-house silicon, and Microsoft’s third generation arrives on a nine-month cadence that suggests this is now a standing program, not an experiment.
The economics explain the urgency. Every token your Copilot generates costs Microsoft real money in power, cooling, and silicon. If the Maia 300 genuinely delivers 40% more tokens per megawatt at scale, across a million-plus chips, the savings compound fast. And with frontier models like Meta’s new Muse Wonder pushing inference demand higher every quarter, the hyperscalers cannot afford to keep paying Nvidia’s margins forever.
What is still unproven
Two caveats matter. First, nothing about the Maia 300 has been independently tested; the efficiency and throughput figures are Microsoft’s, measured against baselines Microsoft chose not to name. Second, the chip does not enter production deployment until 2027, and the 950 W SoC power rating (up from 750 W on the Maia 200) shows the efficiency story is a rack-level one, not a per-chip one. A hot chip in an efficient rack is still a hot chip.
Still, the direction is unmistakable. The AI race is moving from training to inference, from FLOPS to tokens per dollar per watt, and every hyperscaler is now designing its own answer. Whether the Maia 300 dents Nvidia will depend on 2027 deployment numbers, but Microsoft has made its bet clear: the future of AI infrastructure is custom silicon, measured in tokens per megawatt.
