Microsoft to scale custom AI chips in multi-gigawatt push to cut costs

Microsoft plans to deploy its custom AI processors across its global datacenters at multi-gigawatt scale starting in 2027, aiming to reduce reliance on third-party hardware vendors and lower operating costs. The upcoming Azure Maia 300 chip is slated to be the main inference engine for internal AI applications such as GitHub Copilot, Microsoft Foundry and Copilot enterprise tools. Microsoft says the Maia 200, already running live workloads in North America, Europe and Australia, delivers up to 1.7 times the output per watt of standard commercial GPU setups for Teams and Copilot. The Maia 300, the third generation of the program, uses TSMC's N2P node and eight HBM4 memory stacks, and Microsoft projects a 40% gain in energy efficiency per megawatt and a sevenfold increase in rack-scale token throughput over its current system. The article notes that Google, Amazon and Meta are also pushing proprietary chips to challenge Nvidia's datacenter dominance.

Summary written by Upside AI from the original article. Read the source for full details.

Read the original on Yahoo Finance ↗

All news →