World news 343K views since yesterday

Apple and Meta's local AI push comes with a hardware bill

BBC's AI Decoded examines the privacy appeal of running models on office computers, and why that would not end concerns about giant data centers.

Channel: BBC News ·

United KingdomStory of October 27, 2026Trari Newsroom

Why this story? One of BBC News's most watched videos since yesterday, with 343,072 views.

BBC News: The news channel of the BBC, Britain's public broadcaster, and one of the best-known news organisations in the world. 20M subscribers and 7.5B views.

The end of data centres? Apple and Meta’s bet against the cloud | AI Decoded

Things we can learn from the video

Press to jump

0:58Cieslak reports that Apple and Meta have collaborated to run an open-weight AI model locally on Mac Studio hardware.
2:01Hatton distinguishes closed, open-source and open-weight models, explaining what releasing model weights makes possible.
4:32Soares argues that frontier systems trained in massive facilities, rather than local models, present the main catastrophic risk.
7:36Hatton explains the substantial hardware costs of running a large model locally instead of paying a cloud provider.
9:34The discussion examines electricity-grid strain, community protests and the geopolitical vulnerability of large data centers.
11:45Soares links public opposition to data centers with broader fears about the direction of AI development.

Running powerful AI without sending a business's information to the cloud sounds attractive. Paying for the computers to do it is another matter. Consumer technology analyst Ben Hatton puts that trade-off at the center of a discussion about Apple and Meta's approach to locally hosted AI.

The United Kingdom's British Broadcasting Corporation, through BBC News, reports that the companies have collaborated to run an open-weight model on Apple's Mac Studio hardware. Presenter Marc Cieslak explores the development with Hatton, co-host Dr. Stephanie Hare and AI safety researcher Nate Soares.

Hatton distinguishes open-weight models from both closed systems and fully open-source models. Releasing weights means making the model's learned numerical parameters available so others can run it themselves; it does not necessarily mean releasing everything used to build it. Keeping the computing local can let businesses use AI without submitting their information to a cloud provider. But for the large model configuration discussed, Hatton says four computers are needed, putting the hardware bill at around 40,000.

The episode's title, "The end of data centres? Apple and Meta's bet against the cloud | AI Decoded", poses a bigger question than the panel settles. Cieslak describes pressure on electricity grids, demands for land and water, community opposition and the vulnerability of concentrated computing infrastructure as military targets. Hare also examines the companies' business motivations.

Soares draws a distinction between locally runnable models and the most advanced systems trained in enormous facilities. In his assessment, local open-weight models are almost never at the technological frontier, and they are not the main source of the extinction risk he fears from AI.

He interprets opposition to data centers as more than a dispute over nearby construction: it also reflects anxiety about where AI development is heading. The discussion ultimately separates two questions that can easily blur together: where AI runs, and how safely increasingly capable systems are developed.

Published on Trari