Apple is reportedly weighing Nvidia's NVLink for its own AI servers
The Information reports that Apple is designing AI servers around two or four M8 Ultra chips, with Nvidia's NVLink Fusion under evaluation as the link between them.
Apple is reportedly building AI servers around its own M-series chips and looking at Nvidia's NVLink Fusion to link them together. The Information reported on 16 September that the machines are expected in 2029 and would use M8 Ultra processors. Neither Apple nor Nvidia has confirmed any of it, and the report describes the interconnect choice as unsettled.
Two chips, or four
Two server designs are said to be in play: a smaller machine carrying two M8 Ultra processors and a larger one with four. Apple already joins two high-end chips with its own UltraFusion technology, but that does not stretch to the scale-up and scale-out connections a rack of servers needs. NVLink Fusion is not simply a faster wire; it bundles a protocol, switches, chiplets that add Nvidia connectivity to other companies' silicon and a software stack.
Not Apple's first rack
The company is not starting from nothing. It already runs custom servers for Private Cloud Compute, the system that handles AI requests too demanding to run on an iPhone or a Mac. The Information reports the server project began roughly a year ago and was backed by John Ternus while he was running Apple's hardware engineering organisation.
Why it would matter outside Cupertino
If Apple buys into NVLink, one of Nvidia's most determined competitors would use Nvidia's plumbing to connect its own chips. That would leave the interconnect, rather than the processor, as the part of an AI data centre nobody can design around, and it would mark a notable thaw between two companies that have spent years sniping at each other.
Our opinion
Apple's habit is to own the whole stack and say nothing until it is finished, so a report about a 2029 server is best read as a signal of intent rather than a product plan. The interesting question is not whether Apple can build servers, because it already does, but why it would concede the one layer it has never been able to design its way out of. A decade of custom silicon has given it a genuine processor argument and left it with no answer to scale-out. Renting Nvidia's is the pragmatic move and, for a company whose entire pitch is that the silicon is the difference, it is also a quiet admission that the wiring turned out to matter more.