Apple is reportedly evaluating Nvidia’s NVLink interconnect technology to accelerate its custom M8 Ultra AI servers, aiming for a coordinated 2029 data center launch despite historical rivalry.

Apple is reportedly developing AI servers built around its own M-series processors and is evaluating Nvidia’s NVLink Fusion technology for system interconnects, according to The Information. The machines are expected to feature M8 Ultra processors and reach the market in 2029. While the potential use of NVLink Fusion remains unconfirmed by either Apple or Nvidia, adoption over competing solutions could carry market implications far beyond a straightforward hardware procurement decision.

Apple is reportedly weighing at least two server configurations: a compact model equipped with two M8 Ultra processors and a higher-end variant featuring four M8 Ultra system-on-chips. Although Apple utilizes its proprietary UltraFusion technology to seamlessly link two high-end SoCs, the company currently lacks a robust solution for scale-up and scale-out processor connectivity. This gap is where Nvidia’s NVLink ecosystem becomes relevant. Reports indicate Apple intends to leverage the full NVLink infrastructure—including the interconnection protocol, switches, dedicated chiplets, and associated software stack—for its server architecture. The initiative was reportedly launched approximately a year ago under the direction of John Ternus, who then led Apple’s hardware engineering organization.

Apple already manufactures custom servers for its Private Cloud Compute platform, which handles AI workloads too intensive for local execution on iPhones and Macs. These systems currently rely on Apple’s internally developed connectivity technologies, which reports suggest are too slow and costly for large-scale commercial deployment, prompting the company to explore external alternatives.

The Information highlights Apple’s requirement for connectivity technology suited to large-scale deployments, though it does not specify the exact architectural implementation. If the goal is linking multiple servers into larger clusters, this typically falls to scale-out protocols like Ethernet or InfiniBand rather than a scale-up fabric such as NVLink. Nvidia originally designed NVLink to scale performance across its accelerators, optimizing it for accelerator-to-accelerator communication that allows a rack of GPUs to operate as a tightly coupled compute domain. A separate implementation, NVLink-C2C, provides a coherent chip-to-chip interface for connecting CPUs to accelerators and CPUs to CPUs.

Modern Apple M Pro and M Ultra processors utilize a system-in-package (SiP) architecture, combining a CPU chiplet and a GPU/neural engine chiplet bonded via TSMC’s SoIC-mH technology. If Apple maintains this design for the M8 Ultra—which is highly likely—the processor functions as both a CPU and an integrated accelerator. This raises questions regarding how Apple plans to integrate M8 Ultra chips into an NVLink domain and which SiP components would participate. One possibility is that Apple could expose the accelerator portion of the M8 Ultra to NVLink through an NVLink Fusion chiplet, effectively treating it as an accelerator within a scale-up cluster. Alternatively, Apple may be developing a distinct accelerator architecture for its servers, such as placing the GPU/NPU chiplet on a separate substrate or interposer with dedicated memory. However, there is currently no evidence to support a design intended for a product years away from release.

Another factor to consider is Apple’s membership in the UALink Consortium, which oversees the development of UALink, an industry-standard accelerator-to-accelerator interconnect supporting up to 1,024 accelerators. While switch options remain limited today, industry-standard UALink switches offering varied performance tiers are expected to be available by 2029. This timeline makes Apple’s apparent preference for Nvidia’s NVLink as a scale-up fabric somewhat unexpected.

A plausible explanation is that Apple is interested in more than just the NVLink protocol itself. NVLink Fusion serves as a component of Nvidia’s broader rack-scale and data center infrastructure architecture, capable of integrating NVLink scale-up connectivity with Nvidia’s Spectrum-X Ethernet or Quantum-X InfiniBand scale-out networks, including switches featuring co-packaged optics. By adopting Nvidia’s ecosystem, Apple could secure unified scale-up and scale-out connectivity without developing an entire data center networking stack in-house. While speculative, such a strategy would mean Apple is constructing AI servers around substantial elements of Nvidia’s data center architecture while retaining its own processors and avoiding Nvidia accelerators. If realized, this arrangement would underscore Nvidia’s growing role in establishing de facto standards for AI data center infrastructure, regardless of the underlying CPU or accelerator vendors.

Given Nvidia’s position as a leading supplier of data center hardware, a collaboration appears logical if the reports hold true.

Historically, however, Apple and Nvidia have maintained a strained relationship. The friction dates back to the early 2000s, when Steve Jobs accused Nvidia of infringing on Pixar’s patents. Nvidia responded that it owned more graphics intellectual property than Pixar and could therefore pursue legal action. Subsequent disputes arose over GPU design decisions for Apple-supplied components. The relationship deteriorated further during “Bumpgate” (2007–2008), when Nvidia supplied defective GPUs to Apple and other PC manufacturers. Nvidia initially refused to acknowledge the flaw and resisted providing full compensation for repair costs, severely damaging corporate ties. Apple continued using Nvidia GPUs until 2014 or 2015, after which it transitioned to AMD’s Radeon lineup before eventually abandoning discrete third-party GPUs entirely.

In recent years, Apple has resumed utilizing Nvidia hardware. Siri’s latest AI capabilities are primarily driven by Apple Foundation Models developed in collaboration with Google using Gemini technology. Server-side inference for these models runs through Apple’s Private Cloud Compute architecture, with many workloads hosted on Nvidia Blackwell GPUs within Google Cloud. Nevertheless, relying on Nvidia hardware in third-party cloud environments differs significantly from formally adopting Nvidia’s proprietary interconnect and infrastructure technologies for Apple’s own server platforms.