GF Securities analyst Jeff Pu says NVIDIA may turn to Intel’s Foveros advanced packaging in the second half of 2028, while Taiwan Semiconductor Manufacturing Co. (TSMC) continues building the main GPU dies on its leading-edge process nodes.
If that happens, Intel’s role would be at the high end of packaging, not as the company making NVIDIA’s flagship compute dies. The likely job is advanced packaging, plus some non-core silicon around the processor.
For NVIDIA, the appeal is pretty clear. It could take some pressure off TSMC’s packaging capacity, spread supply across more than one partner, reduce bottlenecks, and help support future production ramps. For Intel, landing work like this would give real weight to its IDM 2.0 foundry push, which is all about winning business from outside chip designers.
Foveros matters here because it stacks chiplets vertically on top of a base die. That’s different from Intel’s EMIB approach, which links chips side by side. The upside is denser integration in a smaller footprint, along with faster communication between parts, better power efficiency, and stronger interconnect performance.
Anyone who follows AI chips should keep an eye on this. Earlier rumors in 2026 tied NVIDIA’s Feynman GPUs to Intel’s 14A or 18A process nodes, and other reports had already connected Intel to backend and foundry work for Google accelerators.
Right now, though, there’s no announced deal. Volumes, product names, revenue impact, and even the final structure are still up in the air, so this reads more like a credible industry signal than a finished contract. Still, if it does come together, it would be one of the clearest signs yet that the noise around Intel’s foundry business is starting to turn into actual business.