SK Hynix is weighing Intel as an additional manufacturer for HBM base dies starting with the HBM4E generation, a step that would loosen its reliance on TSMC. The headline tension is whether Intel can actually offer a cost advantage—and the early read is that it may not.
The strategically important shift here is not the supplier name but the part itself. The base die—the logic layer at the bottom of an HBM stack that handles I/O and increasingly integrates controller functions—is migrating onto advanced foundry processes as bandwidth demands climb. That quietly turns memory makers into leading-edge foundry customers and makes the base die a chokepoint sitting directly in the AI accelerator bill of materials. Whoever controls that layer gains pricing leverage over the entire HBM stack that feeds Nvidia, AMD, and the hyperscaler custom-silicon roadmaps.
For Intel Foundry, landing an anchor like SK Hynix would be validation it badly needs, but price is the wrong axis to win on. TSMC's advantage compounds through yield maturity, packaging proximity (CoWoS), and ecosystem gravity. If Intel can't undercut on cost, its pitch becomes supply assurance and geographic diversification—valuable in a capacity-constrained cycle, but a harder sell to a memory maker squeezed on margins. Expect SK Hynix to dual-source selectively rather than shift volume wholesale.
For Japan, the read-through runs through the supply chain rather than the buyers. Japanese suppliers sit deep in HBM economics regardless of which foundry wins: advanced packaging substrates, EUV-adjacent materials, deposition and etch equipment, and photoresist chemistry. A more fragmented base-die landscape means more process variants to qualify—more design wins and more revenue for Japan's materials and tooling vendors. It also sharpens the stakes for Rapidus, whose leading-edge ambitions target exactly the kind of logic work base dies are becoming; a diversifying memory sector is precisely the customer profile it must eventually court. Kioxia, by contrast, remains outside the HBM race and gains little direct benefit.
For Japanese enterprises and SIers, the practical signal is procurement risk. HBM supply concentration is a hidden dependency inside every AI infrastructure buildout. Multi-sourcing at the base-die level should, over time, ease the allocation bottlenecks that have throttled GPU availability—welcome relief for firms and integrators planning on-prem AI capacity, even if the pricing relief arrives slower than the supply relief.
