Apple has discussed using Nvidia’s NVLink Fusion networking technology in a planned AI inference server powered by its own chips, a potential return to the server market it left in 2011, The Information reported on Wednesday. Reuters carried the report, and Apple shares reversed early losses to rise modestly in premarket trading on the news.
What Apple is building
According to The Information, which cited people familiar with the matter, Apple is developing an enterprise AI server aimed at AI developers, businesses and governments. The machine could come in versions with either two or four of Apple’s planned M8 Ultra chips, the company’s highest-performance processors. Apple has considered using NVLink Fusion, Nvidia’s chip interconnect technology, to link the M8 chips together so they can work on inference workloads as a single system.
The timeline is long. The server is not expected before 2029, and the report notes it could still be canceled or could launch without Nvidia technology. Apple and Nvidia did not respond to Reuters’ requests for comment, and Reuters could not independently verify the report.
A long break with Nvidia
The two companies have not worked closely together for nearly two decades. Apple cut ties with Nvidia after an alleged defect in Nvidia graphics chips caused widespread MacBook failures, and it never returned to Nvidia silicon for its computers, building its own Apple Silicon line with AMD as its discrete GPU supplier instead. The split dates to around 2008, when the Nvidia chips in MacBook and iMac models failed at rates that led to a class action and a rare public apology from Apple, which offered free repairs for affected machines.
NVLink Fusion is itself a recent concession from Nvidia. Announced in 2025, it opens Nvidia’s proprietary interconnect to third-party CPUs and accelerators, letting custom silicon participate in Nvidia’s rack-scale systems. Microsoft, Google and AWS have all explored it for their own chips. Apple using it would be the highest-profile adoption yet, and it would put Apple silicon in the same racks as Nvidia GPUs rather than against them.
Why Apple wants back in
Apple last sold dedicated server hardware under the Xserve line, a rack-mounted Mac it pitched to enterprise and education customers, which it discontinued in 2011 as it shifted enterprise focus to standard Mac hardware and consumer devices. Since then the company has watched cloud providers and chipmakers capture the AI infrastructure market, while its own AI features have lagged competitors and its Siri revamp has slipped repeatedly.
An inference server built on M8 Ultra would let Apple sell compute to enterprises running its AI models, and would give the company a hardware stake in the enterprise AI buildout it currently only participates in through services. It also fits the pattern of Apple bringing more of its stack in-house: its own modems replacing Qualcomm, its own displays replacing Samsung panels, and now possibly its own server systems competing with the hyperscalers it works with through iCloud and App Store infrastructure deals.
The market reaction was mild but positive, with shares recovering in premarket trading after the report. Investors have been looking for signs Apple has a coherent AI hardware strategy beyond consumer devices, and the stock has underperformed other megacaps this year on exactly that doubt. Any sign Apple intends to monetize its silicon investment outside phones and laptops reads as optionality the market had not priced in.
How it compares to the field
If the server ships as described, Apple would enter a market with established players at every level. Nvidia sells complete rack systems. Microsoft, Google and AWS design their own AI accelerators but rent them out by the hour through cloud contracts rather than selling hardware. Tesla, Groq and Cerebras have all tried selling dedicated inference systems with mixed results, and even well-funded entrants have struggled to build the software and support organizations enterprises expect.
Apple’s edge, if it has one, is vertical integration. It designs the silicon, the operating system and the model runtime, which few competitors can match end to end. Its weakness is the opposite of the cloud providers’: no sales force for enterprise data center deals, no history of multi-year support contracts, and no fleet of data centers to anchor them.
The obstacles
Several things could still kill the project. The 2029 timeline means the server would arrive against whatever Nvidia, AMD and the hyperscalers ship by then, on hardware cycles that turn over every two years. Apple’s chip roadmap would need the M8 Ultra to stay competitive for inference workloads against dedicated AI accelerators, which are optimized for exactly this use case and ship in volumes Apple has never approached in servers.
Depending on NVLink Fusion also means depending on Nvidia, the company Apple has avoided for two decades, for a core interconnect. The report explicitly says the server could launch without Nvidia technology, which suggests Apple has a fallback or is still weighing the supply-chain costs of the relationship. Nvidia controls the interconnect standard, the software stack around it and the pricing, which is a lot of leverage to hand a supplier.
There is also the question of demand. Enterprises buy AI inference capacity from cloud providers because it is elastic and billed by usage. A fixed Apple appliance would compete on control and data privacy, a real but narrow pitch that has never supported a large hardware business outside Apple’s own ecosystem.
Neither company has confirmed anything. For now this is a research-stage exploration reported by a single outlet, and Apple has canceled internal hardware projects far later in development than this one. But the fact that the discussions are happening at all says something about where Apple sees the AI market going, and about how far Nvidia will go to keep rivals inside its ecosystem rather than outside it.