Apple's reported M8 Ultra server would pick up NVLink where UltraFusion stops

According to Tom's Hardware, the NVLink infrastructure Apple wants covers the interconnect protocol, switches, chiplets that add NVLink connectivity, and a software stack ^[6]. Nvidia built NVLink to scale up its own accelerators, so the fabric is tuned for accelerator-to-accelerator traffic and lets a rack of GPUs behave as one tightly coupled compute domain ^[9]. It has since been opened to third-party hardware ^[24].

That leaves an assembly question. An M Ultra is a system-in-package: a CPU chiplet and a GPU/neural engine chiplet joined with TSMC's SoIC-mH ^[11]. NVLink proper connects accelerators, and the coherent CPU-to-accelerator and CPU-to-CPU variant is a separate implementation called NVLink-C2C ^[10]. Tom's Hardware suggests Apple could expose the accelerator portion of an M8 Ultra through an NVLink Fusion chiplet, which would make the scale-up domain treat it as an accelerator ^[12]. The publication says there is currently no evidence for the alternative, moving the GPU/NPU chiplet onto its own substrate with its own memory ^[13].

Then there is the size of the thing. The largest configuration reported is four M8 Ultra sockets ^[1]. UALink, whose consortium Apple has joined, addresses up to 1,024 accelerators ^[14]. Four sockets is about 0.4 percent of that ceiling ^[25]. A four-socket box does not need a rack-scale fabric to talk to itself. Nvidia sells NVLink Fusion as one piece of a rack architecture that pairs scale-up NVLink with Spectrum-X Ethernet or Quantum-X InfiniBand scale-out, including switches with co-packaged optics ^[16].

Apple's own stitching stops at two dies. UltraFusion joins two high-end SoCs, and Tom's Hardware says the company does not have a proper solution for scale-up and scale-out connectivity between processors ^[7]. Its Private Cloud Compute machines already run mostly on Apple's internally developed connectivity, which The Information describes as too slow and too costly for large-scale commercial deployment ^[8].

The commercial pull shows up in the Mac line. OpenAI and Anthropic are buying Mac Minis and Mac Studios in bulk for AI workloads, and Apple's Mac revenue rose nearly 29 percent last quarter to $10.4 billion ^[20]. Apple last sold dedicated server hardware as Xserve, which it discontinued in 2011 ^[21]. Relations with Nvidia have been strained for nearly two decades after an alleged defect in Nvidia graphics chips caused widespread MacBook failures ^[22].

For the lock-in argument to hold, three things have to be true: the server ships in 2029, the scale-up decision is made while UALink switch choice is still limited ^[15], and an NVLink Fusion chiplet ends up on Apple's package ^[12]. None of it is settled. Tom's Hardware says the use of NVLink Fusion is not formalized ^[5], and the report allows for cancellation or a version without Nvidia technology ^[4]. The project was backed about a year ago by John Ternus while he headed Apple's hardware engineering organization ^[18], and the-decoder identifies Ternus as Apple's new CEO ^[19]. Tom's Hardware argues that if Apple picks NVLink Fusion over competing solutions, the implications run significantly wider than Apple buying Nvidia hardware ^[17].