Microsoft says its third-generation Maia AI accelerator will deliver 40 percent more tokens per megawatt than leading GPU-based systems, the company's most concrete claim yet in its effort to reduce dependence on Nvidia GPUs for AI inference. The chip, called Maia 300, enters production rollout in 2027.

Microsoft announced the chip today and positioned it as the "primary AI inference system" for its products and services. The company is not framing Maia 300 as a research project or a niche offload engine; it is the platform Microsoft intends to run its own AI workloads on, at cloud scale, inside its own datacenters.

What does Microsoft claim for Maia 300?

The headline number comes from Scott Guthrie, executive vice president of Microsoft's cloud group. "We expect Maia 300 to deliver 40 percent more tokens per megawatt compared to leading GPU-based systems," Guthrie said in remarks accompanying the announcement.

That metric matters more than raw peak performance. Tokens per megawatt measures how much inference throughput a datacenter extracts per unit of power, and power — not chip supply — has become the binding constraint on AI infrastructure buildouts. A 40 percent efficiency gain on that metric translates directly into either lower cost per query or more served traffic from the same energy envelope, whichever a hyperscaler prioritizes.

Microsoft also says Maia 300 improves the cost and efficiency of running AI workloads in the cloud broadly, and that it is more capable of handling the latest reasoning models and agentic tasks than its predecessors. Reasoning models generate far more tokens per request than conventional chat models, and agentic workloads run long, multi-step chains of inference calls. Both trends push operators toward accelerators optimized for sustained, high-volume token generation rather than one-shot training bursts — which is exactly the design center Microsoft describes.

The company has not yet disclosed the full rack-level figures; its announcement references performance in a complete 96-accelerator configuration, with detailed specifications to follow as the 2027 rollout approaches.

Where does Maia 300 fit in Microsoft's silicon strategy?

Maia 300 is the third generation of Microsoft's in-house AI chip line, and the cadence itself signals intent. Microsoft has now committed to a multi-generation accelerator roadmap, placing it alongside Google's TPU program and Amazon's Trainium and Inferentia families in the small group of hyperscalers designing their own AI silicon.

For Microsoft, the commercial logic runs on two tracks:

• Cost. Inference is the recurring, high-volume side of AI economics. Every point of efficiency gained on custom silicon compounds across the billions of tokens Microsoft's services serve.
• Supply. Nvidia's GPUs remain the industry default and are capacity-constrained. An internal inference platform gives Microsoft negotiating leverage and a hedge against allocation shortfalls.

None of this means Microsoft is abandoning Nvidia. The company frames Maia as its primary inference system for its own products and services, while GPU-based systems continue to run workloads across Azure. The realistic picture is a mixed fleet, with Maia 300 absorbing the inference traffic Microsoft controls end to end — Copilot-style services, reasoning model serving, and agentic pipelines.

Why 2027 matters

The production date sets the competitive clock. Maia 300 arrives in 2027, by which time Nvidia and its hyperscaler rivals will have shipped additional GPU and custom-silicon generations of their own. Microsoft's 40 percent tokens-per-megawatt claim is measured against today's leading GPU-based systems; whether the advantage holds against the systems actually deployed in 2027 is the open question.

Still, the direction is unambiguous. Microsoft's stated goal for Maia 300 is to become the primary AI inference engine for its own cloud, and the efficiency target suggests the company believes purpose-built silicon can beat general-purpose GPUs on the specific workloads it runs at scale. If Maia 300 lands near its 40 percent claim in production, expect Microsoft to shift a meaningful share of its inference traffic onto the chip — and to use that efficiency edge to press harder on AI service pricing.