Microsoft to Scale Custom AI Chips in Multi-Gigawatt Push to Cut Costs

3-line summary

• Microsoft confirmed Monday it plans to deploy its in-house-designed AI processors at multi-gigawatt scale across global data centers starting in 2027, aiming to reduce reliance on third-party hardware suppliers, cut operating costs, and ease the impact of GPU supply shortages, according to Investing.com.
• Azure Maia 300, the third generation of Microsoft's silicon program, will become the primary inference engine for internal AI applications including GitHub Copilot, Microsoft Foundry and Copilot enterprise tools. It uses a multi-chiplet architecture built on TSMC's N2P process node with eight HBM4 memory stacks, with Microsoft projecting 40% better energy efficiency per megawatt than competing hardware and 7x higher per-rack token throughput than the current generation.
• The current-generation Maia 200 is already handling production workloads in North America, Europe and Australia, delivering up to 1.7x the output-per-watt of standard commercial GPU setups on Teams and Copilot workloads; the report noted Google, Amazon and Meta are also accelerating their own custom chip programs, challenging Nvidia's dominance in data-center accelerators.

This summary is our own processing of the source article; see the source link below for full context.