Apple may adopt Nvidia's NVLink to scale up its custom M8 Ultra AI servers
Apple is reportedly developing AI servers based on its M8 Ultra processors and evaluating Nvidia's NVLink Fusion for interconnects, according to The Information. The servers could arrive by 2029, with configurations using two or four M8 Ultra chips. Apple's own UltraFusion technology is considered insufficient for large-scale deployments, prompting the potential use of Nvidia's interconnect infrastructure.
Apple's current Private Cloud Compute servers rely on internally developed connectivity that the company reportedly finds too slow and costly for commercial-scale AI workloads. The M8 Ultra project, initiated roughly a year ago under hardware chief John Ternus, would pair Apple's system-in-package designs—which use TSMC's SoIC-mH stitching—with Nvidia's interconnect ecosystem.
NVLink functions as a scale-up fabric for tightly coupled accelerator domains, distinct from scale-out technologies like Ethernet or InfiniBand. Apple could expose the accelerator portion of M8 Ultra through an NVLink Fusion chiplet, though architectural questions remain about which SiP components would participate in the NVLink domain.
Apple's potential adoption of Nvidia's NVLink could reshape the AI infrastructure landscape, signaling that even vertically integrated hardware makers may need third-party interconnect technology for large-scale deployments. If realized by 2029, this partnership could influence pricing and availability of AI compute, affecting enterprises and researchers who depend on scalable server clusters. It may also pressure competitors to develop more robust interconnect solutions.