Tech · 12 October 2026 · 4 min read
Microsoft Unveils Maia 300: Custom 2nm Silicon Aimed at Slashing AI Inference Costs
In brief: Microsoft has introduced Maia 300, the third generation of its custom AI inference accelerator designed for Azure data centers. Manufactured on TSMC’s 2nm N2P process using a chiplet architecture, the silicon promises a 40% increase in tokens per megawatt compared to leading commercial GPU platforms. Scheduled for production rollout in 2027, the chip is positioned to power Copilot workloads and complex agentic tasks at multi-gigawatt scale.
by Team Mocchi's
As artificial intelligence infrastructure scales at an unprecedented pace, cloud hyperscalers are running into two critical bottlenecks: power grid limitations and an expensive reliance on third-party silicon. To address both constraints simultaneously, Microsoft has introduced Maia 300, its third-generation custom AI accelerator tailored specifically for cloud inference across Azure.
As reported by The Verge, Microsoft intends to establish the new chip as its primary inference engine across internal products and customer-facing cloud offerings starting in 2027, backing the effort with planned multi-gigawatt data center deployments.
Chiplet modularity on TSMC's 2nm process
Engineering-wise, Maia 300 represents a substantial leap over Maia 200, which debuted earlier this year and already powers production workloads across Microsoft Copilot, GitHub Copilot, and Microsoft Foundry. Built on TSMC's advanced N2P process node, the new design embraces a modular chiplet architecture.
By segregating compute logic from memory controllers and input/output pathways, Azure's hardware teams have expanded the surface area allocated strictly to compute performance. In physical deployment, a complete rack configuration populated with 96 Maia 300 accelerators delivers up to seven times the rack-scale token throughput achieved by Maia 200.
Prioritizing tokens per megawatt for agentic workloads
The benchmark that matters most in current infrastructure planning is no longer raw theoretical compute, but energy efficiency. Scott Guthrie, executive vice president of Microsoft's Cloud and AI group, noted that Maia 300 is projected to deliver 40% more tokens per megawatt compared to current GPU-based systems from suppliers like Nvidia and AMD.
This efficiency gain directly addresses the changing nature of AI models. Workloads are rapidly shifting away from single-turn queries toward long-context reasoning, ultra-low-latency real-time voice, and multi-step agentic execution. Because autonomous agents generate sustained inferencing loops over extended periods, executing them on generalized GPU clusters creates unsustainable operational costs.
Mustafa Suleyman, CEO of Microsoft AI, highlighted that full-stack vertical integration—co-designing algorithmic model architectures down to custom physical silicon—is critical to matching the efficiencies of rivals like Google DeepMind. Owning the hardware platform allows Redmond to optimize memory bandwidth and execution timing specifically around its own model weights.
Mocchi's take
For digital agencies and European enterprises deploying generative AI solutions, Microsoft's push into custom silicon signals an essential macroeconomic shift: inference costs are primed to decline as cloud providers bypass third-party GPU premiums. As autonomous multi-step agents begin executing millions of background operations daily, token efficiency per watt will directly dictate whether enterprise automations remain profitable or prohibitive. Tailoring custom silicon to Azure's models turns inference from a resource bottleneck into a scalable, predictable utility for everyday production software.