Apple Reportedly Wants to Sell M8 Ultra AI Servers, and Nvidia May Be Wiring Them Together

Apple has not sold a server since the Xserve was discontinued in 2011. According to a report from The Information, it is thinking about going back, with a machine built for AI developers, businesses and governments. Apple has confirmed none of it, and the earliest anyone expects the thing to exist is 2029.

The report describes two configurations: a smaller one with two M8 Ultra chips clustered together, and a larger one with four. They would be built for inference, the running of trained models rather than the training of them. Apple has reportedly been at it for about a year, with backing from John Ternus while he was still running hardware engineering. He is CEO now. The report also says the project could be scrapped before it ships.

The Nvidia part

Apple has reportedly looked at Nvidia's NVLink Fusion to move data between those chips. NVLink Fusion is the licensable version of Nvidia's in-rack networking: the interconnect protocol, the switches, chiplets that bolt NVLink connectivity onto somebody else's silicon, and the software stack that runs the whole arrangement. Nvidia already licenses it to Qualcomm, Arm, Fujitsu and Marvell. Neither company has confirmed a thing, and the report says Apple could build the server without Nvidia at all.

Apple and Nvidia have history, and not the warm kind. They spent the late 2000s at odds, most memorably over the defective GPU packaging of the bumpgate era, and no Mac has shipped with Nvidia graphics since about 2014. Apple's pitch since the first Apple silicon Mac in 2020 has been that it designs the chip, the software and the machine around them. Licensing Nvidia's networking so its own chips can talk to each other would be a real concession, if it happens.

Apple already runs servers. It just does not sell them.

Private Cloud Compute has been handling the Apple Intelligence requests that are too heavy for a phone since 2024, on Apple silicon, in Apple's own data centers. Selling that hardware to other people is a different business: support contracts, firmware for customers who are not Apple, rack integration, a sales team that can talk to a government procurement office. Apple walked away from that business on purpose fifteen years ago.

Its own all-Apple-silicon story already has an asterisk, too. In June, Apple extended Private Cloud Compute beyond its own data centers onto Google Cloud, running on Nvidia GPUs.

Where the Europe angle comes from

The Next Web tied this to Europe. In June the European Commission proposed the Cloud and AI Development Act, which would sort public-sector cloud contracts into sovereignty tiers. The top tier calls for European cybersecurity certification and no third-country control over how a system is designed, built or maintained. A government shopping for AI hardware it can keep inside its own walls does not have many options. The Information's report does not say Apple is aiming at that market, so the sovereignty framing is other outlets' inference rather than a reported fact.

Three years is a long time

Apple has killed projects with more momentum than this one: the car after roughly a decade of work, AirPower after an on-stage demo. A server that currently exists as a roadmap inside hardware engineering, with the interconnect still undecided, is not a product.

Watch the Nvidia piece. Apple can design its own chips and has said so at every keynote for six years. Whether it wants to design its own data center fabric is a separate question, and the answer in this report is that it has not decided.