Microsoft has unveiled Maia 300, its next custom accelerator for running artificial intelligence models in Azure. The important detail is easy to miss: this is not a graphics card or a consumer processor. It is a complete inference system designed for the data centers that answer AI requests after a model has been trained.
Production deployments are scheduled to begin in 2027. Microsoft says Maia 300 will become a primary inference platform across its products and services, with multi-gigawatt deployments planned. That scale makes it relevant even if you never see a Maia chip on a store shelf.
What does Maia 300 actually do?
AI computing has two broad jobs. Training teaches a model by processing huge datasets. Inference is what happens when that trained model responds to a prompt, generates an image, summarizes a document, or completes another task for a user.
Maia 300 is built for the second job. Microsoft can tune the silicon, memory, networking, software, racks, and cooling around the workloads running in Azure. The company says the result delivers 40 percent more tokens per megawatt than leading GPU-based systems. That is a Microsoft performance claim, not an independent benchmark, but it reveals the metric the company cares about: useful AI output per unit of power.
The same cost and energy problem is pushing the wider industry toward more specialized infrastructure. It also explains why efforts such as Microsoft’s Super Intelligence Task Force need more than better models. They need enough efficient hardware to serve those models at enormous scale.
Why Maia 300 uses chiplets and HBM4
Rather than building one enormous piece of silicon, Microsoft combines multiple chiplets in each accelerator. Compute chiplets use TSMC’s N2P process, while specialized interface chiplets handle other functions. This approach can improve manufacturing flexibility and lets different parts of the design use the technology best suited to them.
Each accelerator has eight stacks of HBM4, or high-bandwidth memory, providing 18 terabytes per second of bandwidth. AI inference repeatedly moves model parameters and intermediate data between memory and compute units, so memory speed can be as important as raw arithmetic performance. Microsoft says Maia 300 provides 33 percent more memory capacity and 2.5 times the bandwidth of Maia 200.
The accelerator is rated at 950 watts. That sounds extreme by desktop standards, but Maia belongs in liquid-cooled data-center racks, not PCs.
The rack is the real product
A single Maia 300 chip is only one layer of the design. Microsoft packages 96 accelerators into a rack with 27.6 terabytes of HBM and 1.7 petabytes per second of aggregate memory bandwidth. A full rack is rated at 167 kilowatts and is engineered around power delivery, cooling, networking, and serviceability.
Microsoft says a full rack can produce seven times the token throughput of a Maia 200 rack. Larger deployments can connect up to 12,288 accelerators in a superpod using Ethernet-based transport. In other words, Maia 300 is best understood as a data-center system whose parts were designed together, not as a standalone chip competing on a single benchmark chart.
Does Maia 300 replace Nvidia GPUs?
Not immediately, and probably not across every workload. Nvidia’s GPUs support a mature software ecosystem and remain useful for both training and inference. Microsoft also buys systems from Nvidia and other vendors. Maia gives the company another option for workloads it controls closely, especially high-volume inference where efficiency translates directly into operating cost.
Reuters reports that Microsoft could deploy more than one million of its own AI chips next year. If that plan holds, Maia would reduce some dependence on outside accelerator suppliers without eliminating them. A mixed fleet also lets Azure match different models to the hardware that runs them most efficiently.
What Maia 300 could change for users
Consumers will not buy Maia 300, but they may feel its effects through cloud products. More efficient inference can help Microsoft serve more requests within the same power envelope, control costs, and make advanced features available to more customers. It does not guarantee lower subscription prices or faster answers, because software, networking, model size, and demand also matter.
The bigger takeaway is strategic. Microsoft is no longer treating the AI chip as a component it simply purchases. It is designing the silicon and the surrounding system as part of the cloud service itself. That vertical integration could matter as much as any individual performance number, especially while ambitious research efforts, including work on hard mathematical problems such as the Riemann hypothesis, keep raising the demand for specialized compute.
Microsoft’s official Maia 300 announcement and its architecture deep dive provide the full technical specifications. For now, Maia 300 is a 2027 infrastructure story, not a product you can order today.