Nvidia is seemingly planning to build a portion of its upcoming Feynman AI chips at Intel fabs, in order to hedge against supply constraints, according to supply chain reports seen by Taiwanese tech site Digitimes. This new strategy could allow Nvidia to boost production and reduce costs by forcing TSMC and Intel to compete with each other, while also satisfying the USA’s new chip sovereignty endeavour.
As TSMC’s customers face mounting supply chain and political pressures, it wouldn’t be surprising to see some of them looking to other fabrication options. Morgan Stanley estimates that Nvidia could gobble up about 60% of TSMC’s CoWoS (Chip-on-Wafer-on-Substrate) production in 2026, creating a stress point in its ability to satisfy the demands of AI datacentres. This packaging technology enables the fabrication of multi‑chip stacks with high‑bandwidth memory (HBM), making it ideal for building high‑performance AI chips, but there’s a question mark over TSMC’s ability to meet demand.
Looking ahead, the majority of Nvidia’s Feynman AI GPU production looks set to stay at TSMC, with the core parts built using the 1.6mm node. This is expected to represent about 75% of the total chip value, with the remaining 25% moving to Intel’s 18A or 14A nodes. Furthermore, the final packaging is also expected to be produced on US soil by Intel, perhaps using EMIB (Embedded Multi‑die Interconnect Bridge) chiplet bonding. If correct, this should allow the implementation of the latest memory standards, such as HBM4E or HBM5, significantly boosting memory bandwidth for AI algorithms.
This production split indicates that TSMC will be responsible for the high-value GPU die, leaving Intel fabs to handle the I/O die or packaging. However, due to current limitations in 18A technology, meaningful cooperation is likely to begin with 14A production in 2028. Now, whether Intel can match TSMC’s renowned execution standards remains to be seen.
According to Digitimes Asia, both Apple and Nvidia are reportedly experimenting with low-volume production of non-core products at Intel fabs, in order to align with Trump’s administration directives, all while maintaining leading-edge manufacturing at TSMC. Several others, such as AMD and Qualcomm, are also reportedly pursuing a similar dual-foundry strategy, this time pairing Samsung with TSMC.
That said, TSMC isn’t staying put and letting Intel steal its dinner, as it plans to raise its US investments to about $165 billion. These funds are targeted towards building three fabs, two advanced packaging facilities, plus a major R&D centre. In other words, TSMC is also coming to Intel’s home turf to directly compete for the made-in-US title.
