Apple's AI servers may need Nvidia after all

Apple spent years proving it could ditch Nvidia. Now, for the first serious return to dedicated servers since the Xserve era, Cupertino is quietly asking whether it needs Nvidia's interconnect after all. Reuters, amplifying a report from The Information, says Apple is building an enterprise AI server around planned M8 Ultra chips and has discussed using Nvidia's NVLink Fusion to wire those chips together. Versions with two or four M8 Ultras are under consideration. The target customers are AI developers, businesses, and governments hungry for inference capacity. The machine is not expected before about 2029, and it could still be canceled or launched without Nvidia inside.

Inference iron, not another Mac Pro

The project is a bet on inference: running trained models at scale rather than training them from scratch. That is where enterprise demand is shifting, and where Apple wants its custom silicon to matter beyond phones and Macs. NVLink Fusion is Nvidia's kit for letting other chipmakers plug custom accelerators into Nvidia-style data-center fabrics, with ready-made connectors and specialized memory. Some Apple staffers believe it is the best connectivity option available for multi-chip M8 Ultra boxes, according to the report. Apple and Nvidia did not immediately comment to Reuters, which could not independently verify the details.

The initiative reportedly began about a year ago with backing from John Ternus, then hardware chief and now Apple's CEO. Timing still looks soft. A 2029 ship date is a long runway in AI hardware, where roadmaps rearrange every funding cycle. The Information's sources also left open a clean exit: kill the program, or ship Apple silicon servers that talk over something other than NVLink.

A thaw with a long memory

Apple last sold a dedicated server line when it discontinued Xserve in 2011, after deciding Mac hardware and consumer devices were enough. A new rack product would reopen a market Apple abandoned for more than a decade. It would also test a relationship that froze nearly twenty years ago, after disputes that ended with Apple cutting Nvidia graphics from MacBooks following widespread failures. That chill is part of why an NVLink conversation feels strange: the company that built a marketing identity around "Apple silicon only" is shopping Nvidia's interconnect for the one place where bandwidth between chips decides whether a server is real infrastructure or a very expensive Mac in a rack.

The odd angle is dependency, not destiny. Apple can still walk away. Nvidia can still refuse to make the marriage easy. But if the M8 Ultra server ships with NVLink Fusion in the bill of materials, it will rewrite a decade of Apple-Nvidia folklore in a single SKU. The company that left the server market, and left Nvidia, may need both to look serious in the AI age.