Microsoft has set 2027 as the deployment date for its Azure Maia 300 accelerator, a chip the company says will serve as the "primary inference system" for its products and services.

Scott Guthrie, executive vice president for cloud and AI, announced the timeline in a blog post covering Microsoft's in-house silicon. He said the company plans to scale the Azure Maia family to "multi-gigawatt deployments" starting next year.

Maia 200, the current generation, is now in production. The chip runs on TSMC's 3nm process and delivers approximately ten petaflops of FP4 and five petaflops of FP8 compute within a 750W system-on-chip thermal design power envelope.

Deployments of Maia 200 already span the US, Europe and Australia, and the silicon currently carries workloads including Microsoft Copilot, OpenAI models and Microsoft's own MAI models, according to Guthrie. In short: Microsoft is already routing production AI traffic through its own hardware.

What does Maia 300 change?

Guthrie described Maia 300 as a "complete intelligent system" built for new classes of AI workloads. He named three:

• Long-context reasoning
• Latency-sensitive conversational experiences
• Sustained, multi-step agentic tasks

The chip moves to TSMC's N2P process node and adopts a new multi-chiplet architecture that separates compute, memory-controller and I/O functions. Microsoft says this split improves efficiency, scalability and cost structure.

At the rack level, the numbers are concrete. Each Maia 300 rack holds 96 accelerators within a 167kW power envelope and provides:

• 27.6TB of high-bandwidth memory capacity
• 1.7Pbps of aggregate HBM bandwidth across high-speed scale-up domains

Microsoft projects Maia 300 will deliver up to 40 percent more tokens per megawatt than leading GPU-based systems. In a full 96-accelerator configuration, the company expects up to seven times the rack-scale token throughput of Maia 200.

The takeaway: Microsoft is claiming both an efficiency and a raw-throughput advantage over merchant GPU vendors at rack scale.

What does Maia 200 bring today?

Maia 200 carries a redesigned memory system sized for inference workloads:

• 216GB of HBM3e memory running at 7Tbps
• 272MB of on-chip SRAM
• Dedicated data movement engines

The memory configuration matters for inference economics, where bandwidth and capacity often constrain throughput more than raw compute.

Guthrie tied the two generations together directly. "Experience running our AI models and services on Maia 200 directly informed the design of Maia 300," he said. "Together, the two generations are designed to support a broad set of additional Microsoft AI services."

How large can the footprint grow?

The "multi-gigawatt" scaling target starting next year marks the most aggressive commitment in Microsoft's in-house silicon roadmap to date. A gigawatt-class footprint dedicated to Maia-family inference would place Microsoft's custom accelerators alongside, and potentially ahead of, many national-scale AI compute installations.

With Maia 300 racks drawing 167kW each, a one-gigawatt deployment would imply roughly 6,000 racks, or close to 575,000 accelerators, assuming full allocation to that power budget. Microsoft did not break down how it will split capacity between Maia 200 and Maia 300 across the scaling period.

What remains open?

Microsoft has not published pricing, detailed benchmark methodology for the 40 percent tokens-per-megawatt claim, or named the GPU-based systems used for comparison. The company also did not specify which Azure regions will receive Maia 300 first when deployment begins in 2027.

What is clear is the cadence: production silicon now, a gigawatt-scale ramp next year, and a next-generation platform with a dated arrival. Microsoft has committed its own silicon roadmap in public, against a clock.