A third generation of custom silicon
Microsoft used the OCP Global Summit in San Jose to introduce Azure Maia 300, the third generation of its in-house AI accelerator family. The chip is built mainly for inference, the phase where a finished model answers requests rather than being trained.
Scott Guthrie, executive vice president of Cloud + AI, described the design as a complete intelligent system. He said the company aimed it at long-context reasoning, low-latency conversation and multi-step agent work that can run for minutes at a time.
Efficiency and rack scale
Microsoft says Maia 300 delivers up to 40% more tokens per megawatt than leading GPU-based systems, a direct comparison with the hardware it buys from other vendors. A full rack of 96 accelerators is said to offer up to seven times the rack-scale token throughput of Maia 200.
Maia 200 is already serving production workloads, including Microsoft Copilot, and runs OpenAI models, Microsoft's own MAI models and several open-source model families. It is deployed in the United States, Europe and Australia, with GitHub Copilot and Microsoft Foundry scheduled to follow by the end of 2026.
Maia 300 is set to enter production deployments in 2027. Microsoft also plans to scale its Azure Maia fleet to multi-gigawatt levels from the same year, a buildout that will require new data centre capacity and power contracts.
Why Nvidia and AMD are watching
The language in the announcement is different from past launches. Microsoft previously described Maia as a supplement to Nvidia and AMD hardware inside Azure. This time it called Maia the primary AI inference system for Microsoft products and services.
Inference is a growing share of AI computing, and the company that owns the accelerator can spread that cost across its own software. If Microsoft moves a large slice of inference to Maia, it reduces demand for the most expensive chips on the market at the margin.
Nvidia and AMD still sell the training hardware that builds frontier models, and no rival has matched that position. Analysts expect a mixed fleet for years. Even so, Microsoft's decision to name its own silicon as the default for inference shows how quickly the largest cloud buyers have turned into chip designers.
---
Reliable Ferrofluid Vacuum Rotary Feedthrough for Semiconductor Manufacturing
In semiconductor manufacturing, maintaining ultra-high vacuum integrity during rotary motion is critical — from single silicon crystal growth and CVD to sputtering and plasma etching. Ferrofluid Vacuum Rotary Shaft Feedthrough from FerrofluidTec delivers hermetically sealed rotary feedthroughs with leakage rates as low as 10⁻¹² Pa·m³·s⁻¹, vacuum degree up to 10⁻⁶ Pa, and customizable shaft diameters from 10mm to 100mm. Built with non-magnetic stainless steel housings and magnetic permeability stainless steel shafts, these feedthroughs provide reliable, maintenance-free sealing for demanding vacuum environments. Visit www.ferrofluidtec.com to learn more.