Apple Reportedly Explores NVIDIA NVLink Fusion for AI Servers

Apple is reportedly exploring a return to purpose-built server hardware with a system aimed at AI inference. Reuters reported on September 16, citing The Information, that Apple has discussed using NVIDIA NVLink Fusion networking technology in a planned server built around future Apple-designed chips.

The evidence status matters more than the headline. Apple and NVIDIA have not announced this server project, Reuters says it could not independently verify The Information’s reporting, and the reported design is far enough from a commercial product that its architecture, timing and even existence at launch remain uncertain. A reported internal project can change substantially or be canceled before customers ever see it.

NVLink Fusion is specifically designed for systems where the central compute silicon does not have to come entirely from NVIDIA. NVIDIA introduced the platform in 2025 as a way for hyperscalers and chip designers to build semi-custom AI infrastructure around their own CPUs or accelerators while using NVLink connectivity and other parts of NVIDIA’s rack ecosystem.

That makes the reported Apple discussion technically plausible without implying that Apple would be buying NVIDIA GPUs for the server. If Apple chose NVLink Fusion, the strategic value would be the interconnect and surrounding data-center ecosystem: Apple could keep its own processor architecture while adopting an established high-bandwidth scale-up fabric. That is an architectural interpretation of NVIDIA’s published platform model, not confirmation that Apple has selected it.

Reuters describes the proposed machine as an AI inference server. Inference is the phase where a trained model processes new inputs and produces outputs; it has different workload and deployment priorities from training large models from scratch. The distinction matters because a server optimized for inference does not automatically imply that Apple is trying to reproduce every capability of a large GPU training cluster.

The Information’s report, as summarized by Reuters, says Apple is considering versions using two or four future M8 Ultra processors. Those names and configurations should be treated as reported internal planning rather than final specifications. Apple has not published a product page, technical brief, performance target, power envelope, memory configuration or availability commitment for such a server.

The Apple report arrives as NVIDIA is expanding NVLink Fusion beyond its own GPU-centric systems. NVIDIA’s launch material names partners across custom silicon, CPUs, networking and design services, and d-Matrix announced this month that its Raptor inference accelerator will use NVLink Fusion in NVIDIA MGX rack-scale infrastructure.

That broader adoption is useful context: NVIDIA can remain part of an AI system even when another company supplies the main accelerator or processor. For a company designing its own silicon, the attraction is access to an established scale-up fabric and rack ecosystem. For NVIDIA, Fusion extends the reach of NVLink into semi-custom systems. None of that establishes that Apple will become a customer.

The Information reportedly places the project no earlier than 2029. At that distance, the date is better understood as a reported internal target than a launch window. Future Apple chip naming, manufacturing plans, system architecture and NVIDIA’s own data-center platform could all evolve before then.

The same caution applies to the reported M8 Ultra name. It is useful for describing what The Information says Apple is considering today, but it should not be converted into assumed core counts, memory capacity, process technology or performance. None of those specifications has been announced for the reported server.

If the report proves accurate, the notable part is not simply that Apple might build a server. It is that a company known for tightly integrated in-house silicon is reportedly evaluating an external interconnect platform for scaling that silicon in AI infrastructure. NVLink Fusion was created for exactly that kind of mixed-vendor architecture.

For now, the defensible conclusion remains narrow. NVIDIA has a real, shipping-to-partners NVLink Fusion strategy for semi-custom AI systems, and The Information reports that Apple has discussed using it in a future inference server. Whether that discussion becomes an Apple product, uses the reported chip configuration, keeps NVIDIA technology or reaches the reported timeframe remains unconfirmed.