Microsoft Unveils Maia 300 AI Chip, Aiming to Replace Nvidia in Azure

Microsoft today announced the rollout of its third‑generation Maia AI chip, saying it will deliver 40% more tokens per megawatt than leading GPU systems and, in a full 96‑accelerator configuration, up to seven times the rack‑scale token throughput of Maia 200. The company plans to deploy the chip in production AI workloads in 2027, targeting a shift away from Nvidia and AMD hardware.

Key points

• Maia 300 promises 40% higher tokens per megawatt versus GPU competitors.
• 96‑accelerator builds deliver up to 7× the throughput of Maia 200.
• TSMC N2P process underpins a chiplet architecture separating compute, memory, and I/O.
• Azure will scale Maia to multi‑gigawatt deployments starting in 2027.
• Microsoft AI CEO Mustafa Suleyman says the chip is key to catching up with DeepMind and OpenAI.

What happened

Microsoft’s cloud group executive Scott Guthrie announced that Maia 300 will serve as the primary AI inference system for Microsoft’s products and services. The report quotes Guthrie saying, "We expect Maia 300 to deliver 40 percent more tokens per megawatt compared to leading GPU‑based systems." In a 96‑accelerator configuration, the chip can achieve up to seven times the rack‑scale token throughput of the earlier Maia 200.

The new chip will begin rolling out to production AI workloads in 2027, with Microsoft planning multi‑gigawatt deployments across Azure. Existing production services—including Microsoft Copilot, GitHub Copilot, Microsoft Foundry, and the company’s own AI models—are already running on Maia 200, the first generation of the line.

Maia 300 is built on TSMC’s N2P process node and adopts a chiplet architecture that isolates compute, memory controller, and I/O functions. This separation allows each block to use the technology best suited to its role, boosting overall silicon efficiency.

The backstory

Microsoft’s push into in‑house AI hardware began with the Maia 200, announced in January. The 200 series quickly became the backbone of Microsoft’s cloud‑based AI services, delivering the best real‑world performance in the company’s fleet. However, the rapid growth of reasoning models and agentic tasks exposed limits in GPU‑centric architectures, prompting the development of a more specialized chip.

The industry has seen a similar trend: companies like Google and Meta have invested heavily in custom ASICs to reduce latency and cost. Microsoft’s Maia 300 follows that pattern, aiming to lower dependence on Nvidia and AMD components.

Technical trade‑offs

The chiplet design offers significant power savings, but it also introduces inter‑chip communication overhead. Separating compute and memory can reduce latency for certain workloads, yet the need for high‑bandwidth links between chiplets may increase cost in dense data‑center deployments.

Power efficiency gains—40% more tokens per megawatt—translate into lower cooling and electricity bills, but the upfront cost of deploying a new silicon line remains high. Microsoft’s decision to scale up to multi‑gigawatt deployments suggests confidence that the long‑term savings outweigh the initial capital outlay.

Business impact

By moving to an in‑house chip, Microsoft can lock in pricing for silicon and reduce exposure to GPU supply constraints. The shift also positions Azure to better compete with AWS and Google Cloud, which rely heavily on Nvidia GPUs for AI inference.

The move could pressure Nvidia’s market share, but it also forces the company to innovate further. Businesses that currently depend on Nvidia GPUs for Azure workloads may face migration costs unless Microsoft offers a clear transition path.

What builders and businesses should do

• Evaluate current AI inference workloads for compatibility with Maia 300’s architecture.
• Plan migration timelines to align with Microsoft’s 2027 rollout schedule.
• Negotiate Azure pricing tiers that reflect the new chip’s efficiency benefits.
• Monitor Azure’s multi‑gigawatt deployment plans for potential cost‑saving opportunities.
• Prepare internal teams for potential changes in driver and firmware requirements.

The bottom line

Microsoft’s Maia 300 signals a decisive move toward silicon ownership for AI inference. Watch Azure’s deployment schedule in 2027 to gauge how quickly businesses can transition away from Nvidia GPUs.