SAN FRANCISCO, the US (Oct 12): Microsoft said on Monday it will roll out its in-house Maia artificial intelligence (AI) chips on a massive scale starting in 2027, making them the main hardware for running AI across its products — a move that could reduce its reliance on Nvidia.

Chips like Maia are the technology inside the vast data centres powering the AI boom, and Microsoft plans to deploy enough of them to draw several gigawatts of electricity — roughly the output of several large nuclear reactors.

Nvidia has held a near-lock on the chips used to train AI models, but tech giants and rival chipmakers are now racing to stake out positions in inference — the work of running those models to answer questions and carry out tasks once they are up-and-running.

That market is expanding fast as companies roll out AI assistants and so-called agents that perform multi-step jobs on their own, with Nvidia, AMD, Amazon and Google all pushing out new products in the space.

"Today marks a new phase for Microsoft’s AI infrastructure," Microsoft executive vice president Scott Guthrie wrote in a blog post announcing the plans.

The company said the Maia 200 chip is already operational, including for the Microsoft 365 Copilot assistant, while a more powerful successor, the Maia 300, will launch in 2027.

Maia 200 runs models from OpenAI, Microsoft’s own in-house MAI models and leading open-source models.

The chip is deployed in the US, Europe and Australia, and delivers up to 1.7 times more output per watt than leading GPU-based systems for AI features built into Teams and Copilot, according to the company.

Microsoft said the Maia 300 will squeeze 40% more AI work out of the same amount of electricity than leading systems, and more than twice as much as its own previous chip.

Energy efficiency has become a central concern for tech companies racing to build data centres, as the electricity demands of AI strain power grids and drive up costs.

Each server rack, a cabinet-sized unit in a data centre, will hold 96 of the chips, twice as many as before, and handle up to seven times the workload, the company said.

The new chip’s computing components will be manufactured by Taiwan’s TSMC.

In a separate post, Microsoft engineers said a big focus was cutting the power wasted moving data back and forth inside the machines, a growing problem as AI tools handle more complex tasks.

The systems are mostly cooled with liquid, but Microsoft said they can also run in older data centres cooled by air.

Uploaded by Felyx Teoh