The silicon industry is moving into an increasingly common, yet uncomfortable territory: diversifying suppliers without breaking the “marriage” with TSMC. According to reports published by DigiTimes and gathered by industry analysts and leakers, NVIDIA is exploring a partial collaboration with Intel Foundry for its Feynman platform, scheduled for 2028. This involves a “dual foundry” scheme that, at least on paper, aims to reduce political and capacity risks without touching core performance.
The core idea of the rumor is quite straightforward: the main die of the GPU would still be fabricated at TSMC, while a I/O die might be produced in Intel using 18A or 14A (depending on maturity, performance, and yields at that time), then pass through EMIB for advanced packaging. In this arrangement, Intel could hold up to roughly 25% of the final mix, with TSMC handling about 75%. The implicit message is clear: it would be a low-volume, less critical, and non-core adoption to meet “U.S. manufacturing” pressures and minimize industrial risk.
Why does this pivot sound plausible (even though it’s still just a rumor)?
First, because the timeline aligns with what NVIDIA has publicly stated about its roadmap. Reuters summarized at GTC that, after Blackwell Ultra and Rubin/Rubin Ultra, Feynman appears as the architecture for 2028. If that’s the horizon, discussing supply chain moves now (evaluations, PDKs, packaging, master contracts) isn’t far-fetched: these decisions are “cooked” years in advance.
Second, because Intel has been emphasizing that its 14A node is aiming for high-volume production in 2028, matching the date in the leak. Intel also highlights that 14A relies on advances like RibbonFET/PowerVia and its foundry strategy to attract external clients. If NVIDIA wants to “test” a second foundry, doing so when the node promises volume and industrial maturity is the least risky option.
Third, because the real bottleneck is no longer just lithography: it’s advanced packaging and overall capacity to assemble complex chips. In the AI GPU ecosystem, packaging (and everything surrounding it) has become a strategic piece. In this context, opting for an alternative route like EMIB — even partially — makes operational sense as a safeguard, even if it’s a “limited” and carefully scoped one.
Geopolitical implications: diversifying to breathe
The rumor cannot be divorced from the political backdrop. In recent years, big tech companies have learned that supply chain optimization now depends not only on cost/performance but also on resilience, regulatory pressure, and national narratives. DigiTimes frames this trend as a shift from extreme concentration at TSMC towards multi-sourcing as a risk mitigation mechanism. From this perspective, both Apple and NVIDIA are choosing less critical components to “shift the pieces” without risking their core products.
Here emerges an interesting paradox: TSMC could lose some volume… but gain stability. DigiTimes offers a three-layer interpretation: easing regulatory focus due to its dominant market share, reducing political pressure in the U.S., and simultaneously maintaining premium orders where TSMC remains hard to replace. In other words: sacrificing “non-essential” parts to protect the essential ones.
What does Intel stand to gain (if this materializes)?
For Intel Foundry, even a “slice” of a NVIDIA design would be more than a contract: it would be a symbolic validation. The market looks not only at nodes but at who dares to manufacture what and with what guarantees. If the I/O component and/or advanced packaging come into play, Intel would bolster its credentials to attract other clients who hesitate due to technical risks, timelines, or recent histories.
Additionally, the leak mentions broader sector discussions around collaborations with Intel, often beginning with packaging (EMIB) before core dies. This pattern — packaging first, silicon later — aligns with how high-impact programs reduce risk.
The critical point: this isn’t a breakup with TSMC, it’s a “escape clause”
It’s important to highlight what the rumor suggests implicitly: this wouldn’t be “NVIDIA leaves TSMC”. Instead, it would be a design that maintains TSMC as the main backbone while using Intel as a controlled second option (for I/O, packaging, or both). In tight supply chains, such decisions serve as a lever: if capacity tightens, if politics shift, or if packaging becomes a limiting factor, the customer already has a backup plan — even if partial.
For now, though, this remains “supply chain talk.” DigiTimes mentions evaluations and strategy; NVIDIA and Apple, as usual in these cases, don’t usually comment on such leaks. Until we see confirmations through product documents, public contracts, official statements, or industry signals, it’s prudent to treat this as what it is: a plausible rumor, but not yet a fact.
Frequently Asked Questions (FAQ)
What does it mean that NVIDIA would use Intel for an “I/O die” and TSMC for the main die?
The chip would be split into parts: the critical compute die remains at TSMC, while the I/O part (interconnections, controllers, etc.) could be produced at Intel, reducing overall dependency without risking the core performance.
Why is 2028 a key date for Intel Foundry?
Because Intel aims for 14A to support high-volume manufacturing in 2028, aligning with NVIDIA’s public roadmap for Feynman.
What is EMIB and why does it matter for AI GPUs?
EMIB is Intel’s advanced packaging technology that enables high-speed interconnections between dies. In complex AI chips, packaging is as strategic as the manufacturing node itself.
Could the supplier distribution actually change if Intel “approves” NVIDIA?
In the short term, even if the rumor holds, the percentages might be limited (up to around 25%). But in the medium term, an validated alternative route could give NVIDIA more negotiation leverage, resilience, and options if bottlenecks occur.
Source: Jukan on X