Apple is developing an enterprise server built on its own silicon, aimed at AI developers, companies and government organizations, according to The Information, citing people familiar with the project. The machine would run already-trained models rather than train them, in configurations of either two or four M8 Ultra chips, and Apple is evaluating NVLink Fusion — Nvidia's data center interconnect, originally built for Nvidia's own chips and later opened to third-party hardware — to link them. The earliest expected launch is 2029, and Apple may still cancel the project or ship it without the Nvidia technology.
The internal politics here matter more than the specifications. John Ternus backed the project roughly a year ago, while he was still running hardware, on the view that Apple had an opening in the fast-growing AI compute market. Ternus is now Apple's chief executive. A speculative hardware bet whose sponsor becomes CEO has a substantially better survival rate than one whose sponsor does not, which is the strongest signal in the report that this is more than a study.
The demand side is already real, and it did not come from Apple. OpenAI and Anthropic are buying Mac mini and Mac Studio units in bulk to run AI workloads. Apple's Mac revenue rose almost 29% last quarter, to $10.4 billion.
Those two facts sit next to each other in the reporting without a causal claim between them, and it is worth resisting the temptation to draw one. A handful of labs buying desktop Macs by the pallet does not move a $10.4 billion line on its own. What the bulk buying establishes is narrower and more interesting: two of the most sophisticated buyers in the industry are willing to build inference infrastructure out of desktop hardware that was never designed for a rack. Apple's server is an attempt to productize a workaround its customers invented.
Which makes the NVLink Fusion detail the one to watch. Apple is the most vertically integrated hardware company in the industry, and the report has it considering Nvidia's interconnect for the part of the system it cannot supply itself. Building chips is a solved problem for Apple. Making many of them behave as one machine, at data center scale, is not — and the company appears to know it well enough to look outside. That the option is still open, three years out, suggests the decision is genuinely undecided rather than leaked as a negotiating position.
The gap in the reporting is software, and it is a large one. Nothing in the account describes what runs on this machine: what serving stack, what model formats, what the migration path looks like for a customer whose inference pipeline assumes Nvidia's tooling. Apple has no meaningful presence in data center software. Enterprises and government buyers do not purchase silicon, they purchase something that already works with what they have. A server announcement with no software story attached is a chip roadmap wearing a product's clothes.
And then there is the date. 2029 is three years away, for a product whose entire thesis is demand visible in the market right now. Apple is not committing to a launch; it is holding an option, cheap to keep open and cheap to close, on a market that will look nothing like it does today by the time the option matures. That is a defensible way to run a hardware portfolio. It is not a bid for the inference market.
Apple's Mac business became an AI compute business by accident, through customers who found a use for the product that Apple did not design for. The server would make that deliberate, if it ships at all — and not before 2029, by which point the labs currently stacking Mac Studios will long since have settled what their infrastructure looks like.