CEO Huang confirmed x86 and ARM will co-exist in Nvidia rack-scale offerings; Intel foundry adoption remains uncertain

NVIDIA and Intel announced a blockbuster partnership, with NVIDIA's CEO Jensen Huang and Intel's CEO Lip-Bu Tan discussing the details during a webcast with reporters from multiple media outlets. As part of the collaboration focused on datacenter and consumer CPU segments, NVIDIA invested $5 billion in Intel's common stock to "celebrate" the partnership.

Jensen Huang disclosed the technical motivation for the deal: with current rack-scale products, NVIDIA is able to achieve NVL72 configurations with ARM-based CPUs, but with x86 datacenter CPUs such as Intel's Xeon, they are limited to NVL8 configurations due to the use of a PCIe-based interface. As Huang explained, "this architecture, the NVLink 72 rack scale architecture is only available for the Vera CPU that we build, the ARM CPU that we build. And for the x86 ecosystem, it's really unavailable except with server CPUs over PCI Express." He continued: "And that has limitations in how far you could scale these scale-up systems. And so the first opportunity is that we can now with Intel x86 CPU, integrate it directly into NVLink ecosystem and create these rack-scale AI supercomputers." NVIDIA claims that this is a $30 billion untapped market which they will address through the partnership.

The partnership involves mutual commitments from both firms. NVIDIA says it will be the primary consumer of Intel's datacenter CPUs and will serve as the seller of its RTX GPU chiplets for Intel, which will be used for a PC SoC, ensuring that the collaboration is not one-sided and demonstrating both companies' commitment to deep cooperation.

When asked whether NVIDIA would consider using Intel Foundry Services for new chips, Jensen Huang acknowledged that the firm has been working with the IFS, but TSMC's importance cannot be ignored. He stated: "I think Lip-Bu and I would both say that TSMC is a world-class foundry. And in fact, we're both very successful customers of TSMCs. And I can't -- you just can't overstate the magic that is TSMC. But today, our conversation today, our partnership today is completely focused." This suggests there is currently little room for IFS's node integration.

One potential avenue for Intel Foundry's involvement is through advanced packaging technology collaboration. Jensen mentioned Intel's "Foveros" packaging, which might be used for the PC chip to combine the RTX GPU with Intel's CPU chiplets. While concrete plans remain uncertain and could shift depending on how Intel's 18A and 14A nodes develop, for now both NVIDIA and Intel appear comfortable relying on TSMC.