Microsoft Announces Maia 300: 40% More Tokens per Megawatt Than GPUs

Microsoft has announced the third generation of its in-house Maia AI accelerator. Maia 300 is meant to become the "primary inference system" for the company's products and services. The headline number: Microsoft expects the chip to deliver 40% more tokens per megawatt than leading GPU-based systems. Microsoft will begin rolling out Maia 300 in production in 2027. For the company, it is a step toward reducing its dependence on Nvidia, though not a move away from Nvidia's GPUs.

Megawatts Instead of Teraflops

The choice of metric is deliberate. Scott Guthrie, executive vice president of Microsoft's cloud division, measures efficiency in tokens per megawatt. For data centers, the main constraint is increasingly not the number of chips but the available electrical power. If the same megawatt yields 40% more generated tokens, then with a fixed energy budget, revenue per site rises. Or the cost of each model response falls.

There is a caveat. Guthrie talks about what Microsoft expects ("we expect"), not about results from independent tests. The comparison is against unnamed "leading GPU systems." The measurement conditions are not disclosed in the published part of the material. This is a vendor claim, not a benchmark.

Built for Reasoning Models and Agents

Microsoft stresses that Maia 300 handles modern reasoning models and agentic scenarios better. Such workloads generate many tokens per request: the model "thinks" in chains, and the agent makes a series of calls. Inference, not training, is becoming the biggest line item. A chip built for inference targets the most expensive part of cloud AI.

The scale of the plans is visible in the headline of the corporate blog post: Microsoft talks about expanding the Azure Maia system to multi-gigawatt deployments. Guthrie's quote about a full configuration of 96 accelerators cuts off in the available excerpt, so details of the rack architecture will have to be sought in the original source.

Key parameters of the announcement:

• Generation: third in Microsoft's line of in-house AI chips
• Role: "primary inference system" for the company's products and services
• Claimed efficiency: 40% more tokens per megawatt versus leading GPU systems
• Target workloads: reasoning models and agentic tasks
• Start of production rollout: 2027
• Scale: expansion of Azure Maia to multi-gigawatt capacity

What to Watch Before 2027

Mass launch is more than a year away, and there is nothing to buy for Maia 300 yet. The practical questions are these.

First: when Microsoft will open the chip to Azure customers, not just to its own services. For now it is described as internal infrastructure for the company's products. If it reaches public instances, expect fresh price pressure on inference.

Second: which models and frameworks will run without painful porting. Any custom chip depends on its software stack. Developers should check in advance how tightly their pipelines are tied to CUDA.

Third: independent measurements. The 40% figure needs to be checked against real tests on real models before it goes into any cost planning.

Teams that build products on reasoning models and agents should already be calculating cost in tokens per task and in energy per token. These are the metrics that will determine which cloud provider's inference gets cheaper first.