Microsoft plans to begin production deployments of its Maia 300 artificial intelligence accelerator in 2027, extending the company’s strategy of building custom silicon for Azure and its own AI services. Microsoft announced the chip on October 12, 2026, and said Maia will become the primary inference system across its products and services.

The company is positioning Maia 300 for long-context, latency-sensitive and multi-step agentic workloads. Its largest performance claim is at rack scale: a 96-accelerator configuration is designed to deliver up to seven times the token throughput of a comparable Maia 200 rack.

Microsoft Maia 300 raises rack-scale throughput

In its official announcement, Microsoft said Maia 300 should produce up to 40% more tokens per megawatt than leading GPU systems. The company also expects up to 2.5 times the performance per watt of Maia 200, making power efficiency as important as raw model throughput.

Each accelerator uses eight HBM4 memory stacks that provide 18 terabytes per second of peak bandwidth, while the chip carries a 950-watt thermal design power. Microsoft said TSMC’s N2P process will manufacture the compute chiplets. The specifications reflect the demands described in our AI inference explainer: serving models repeatedly requires memory bandwidth, predictable latency and efficient token generation.

A full Maia 300 rack operates at 167 kilowatts and connects 96 accelerators. Microsoft lists 27.6 terabytes of total HBM capacity and 1.7 petabytes per second of aggregate HBM bandwidth for that configuration. At a larger level, Azure can connect as many as 12,288 accelerators in a superpod, according to the company’s technical overview.

Why Microsoft is designing its own AI chip

Custom accelerators give Microsoft another way to tune infrastructure around the models and services it operates at scale. Maia 200 is already in production for Copilot, Microsoft’s MAI models and other services, with GitHub Copilot and Microsoft Foundry support expected by the end of 2026.

Microsoft said Maia 200 is deployed in the United States, Europe and Australia and can deliver up to 1.7 times the tokens per watt of leading GPU infrastructure for workloads including Teams, Copilot and Cowork. Maia 300 advances that design with 2.5 times the memory bandwidth, 33% more memory capacity and twice the HBM bandwidth per watt of Maia 200.

The plan also adds competitive pressure across the semiconductor supply chain. Foundry capacity and advanced packaging remain strategic constraints, a theme visible in GlobalFoundries’ $21 billion fab plan and other investments aimed at expanding specialized chip production.

What the 2027 rollout means

Microsoft says Maia 300 production deployments will start in 2027, with the broader Maia system moving toward multi-gigawatt capacity. The architecture uses liquid cooling to remove more than 80% of the thermal load, a necessity for dense 167-kilowatt racks.

Customers should treat the performance figures as Microsoft claims until real workloads provide comparable benchmarks. Token throughput can vary with model size, precision, context length, batching and latency targets, so the seven-times figure does not imply that every application will run seven times faster.

The rollout still matters for Azure’s economics. If Microsoft can increase useful inference per watt while reducing dependence on a single accelerator supplier, it can improve capacity planning for services ranging from enterprise agents to developer tools. The strategy also fits a wider market in which major cloud providers are pairing custom chips with external GPUs rather than choosing only one path, while AI companies continue expanding internationally, including through initiatives such as Anthropic’s Singapore office and AI agreement.

Elsy Kanana is a financial and cryptocurrency journalist at FusionMarketNews, covering digital assets, blockchain technology, financial markets, and emerging fintech trends. Her reporting focuses on market movements, regulatory developments, and on-chain analytics, delivering clear, data-driven insights to readers worldwide.