• by Laurie Sullivan , Yesterday

Microsoft announced its "Maia 300" chip today. It will be available for delivery next year.

While custom silicon like Microsoft's "Maia 300" chip is built for cloud infrastructures and AI inference --for example, in running Copilot and large AI models -- it offers more efficiencies, such ass processing 40% more artificial intelligence (AI) data using the same amount of electricity.

For digital advertising, this means advertising platforms will be able to process massive amounts of data and generate AI-driven content at a lower cost, making ad delivery faster and far less expensive. 

Satya Nadella, Microsoft CEO, on Monday in a post on X introduced the plan to “scale our Maia family to multi-gigawatt deployment, making it the primary inference system for our own product and services.”

AI is increasingly being used to predict what a user will click on or buy.

By getting 40% more work out of every watt of power, Microsoft can cut its electricity bill -- as well as costs for brands using its products because the savings get passed down, making it less expensive for businesses to run AI-powered ad campaigns.

“Our goal is to build on the full-stack efficiency gains we’ve already seen, while continuing to bring the latest innovations from across the ecosystem into our fleet,” Nadella wrote in a post on X.

The BBC published an interview with Microsoft President Brad Smith on Sunday. He contends that AI faces fewer safety regulations than everyday consumer items like milk or washing machines, and that Britain should write the rules.

Smith suggests that companies should put a “safety brake on any AI system that’s controlling an autonomous device” and where a company may want to slow things down or slow things down -- for example, with emergency brakes on cars and trains.

“It’s a common feature,” he said. “Let’s build it into AI, not just in the models, but cloud infrastructure.” He also cited incidents such as privacy breach reporting.

Experts are arguing about how AI should roll out as the technology becomes more advanced.

AI-written ads or images produced today for each person across the internet can be prohibitively expensive, but chips like these that can process text and data for much less can give advertisers a way to use AI models that can generate real custom copy, localized offers, and unique graphics in millisecond as the webpage loads.

"Right now, your washing machine is subject to more safety rules than AI," Smith said, adding that milk cartons and passenger aircrafts are subject to inspection and certification.

Microsoft is not the only company looking at safeguards and advancing AI. Most advertising competitors are launching their own custom chip to reach or surpass similar numbers, from Google to Meta, and even Nvidia and others.

Google engineers built the company’s first Tensor Processing Unit (TPU) in 2015, and has advanced to TPU v7t and TPU 8i.

In Google’s naming ideology, the “t” is for training the models and “i” is for inference, the processes of training an AI model to make predictions.

Earlier this year at the Google Cloud Next event, this generation represents a shift in how Google builds hardware. Instead of making one general-purpose chip, they have split their TPU line into two specialized chips to handle different tasks.