One of the semiconductor industry's long-standing "unwritten rules" is once again being rewritten.
Clear signs of a crack have emerged in TSMC's monopoly, which has until now handled the entirety of NVIDIA's AI accelerator manufacturing. According to the latest supply chain information, DigiTimes reports that NVIDIA has decided to partially adopt Intel Foundry for the production of its next-generation GPU architecture, "Feynman," slated for launch in 2028.
This comes on the heels of the $5 billion (roughly ¥750 billion) strategic investment NVIDIA made in Intel in September 2025, and represents an extremely political and technical structural shift, intersecting with geopolitical pressure to return manufacturing to the United States under the Trump administration.
The "Feynman" Architecture: A Hybrid Manufacturing Model Combining TSMC and Intel
As of 2026, the current flagship architecture is "Rubin," and its successor, slated to arrive in 2028, is "Feynman." Named after physicist Richard Feynman, this next-generation chip is expected to feature a hybrid structure that makes maximum use of "chiplet technology" in its manufacturing process.
Division of Labor in the Manufacturing Process
According to sources, NVIDIA is planning the following clear division of labor for the manufacturing of "Feynman":
- GPU Compute Die (Core): TSMC A16 Process
The most critical GPU core component, which determines computational performance, will continue to be handled by TSMC. TSMC's next-generation process, "A16 (equivalent to 1.6nm)," is expected to be used here. This TSMC-manufactured die is said to account for approximately 75% of the chip's overall value, making it clear that NVIDIA is making absolutely no compromise on performance. - I/O Die and Peripheral Circuits: Intel 18A / 14A Process
On the other hand, the I/O die, which handles data input/output, and other non-core silicon components will be outsourced to Intel Foundry. Specifically, depending on yield conditions as of 2028, Intel's "18A" or next-generation "14A" process is planned to be used.
A Shift in Packaging Technology: Diversifying from CoWoS to EMIB
Also noteworthy is the diversification of "packaging." Currently, the single biggest factor behind the AI chip supply shortage is the capacity constraints of TSMC's advanced packaging technology, "CoWoS (Chip-on-Wafer-on-Substrate)."
To resolve this bottleneck, NVIDIA will adopt Intel's proprietary packaging technology, "EMIB (Embedded Multi-die Interconnect Bridge)." According to reports, an allocation is being considered in which Intel would handle approximately 25% of the final packaging process for the Feynman generation, while TSMC would continue to handle the remaining 75%. This is aimed at diversifying the risk of over-reliance on a single company, TSMC, while boosting overall supply volume.
Why Intel? Three Structural Factors
For NVIDIA CEO Jensen Huang, diversifying away from TSMC—a long-time partner known for its overwhelming manufacturing quality—is a decision that carries technical risk. Nevertheless, three unavoidable structural factors lie behind the decision to move forward with adopting Intel.
1. Geopolitical Pressure and the Obligation to "Make in America"
As the Digitimes report points out, the Trump administration's "US manufacturing mandates" and tariff policies are a powerful driver.
TSMC is a Taiwanese company positioned on the front lines of US-China tensions. To avoid becoming a "single point of failure" in the event of a Taiwan contingency, the US government has continued to apply intense pressure on major tech companies to increase their proportion of domestic US manufacturing. NVIDIA's $5 billion investment in Intel, along with the manufacturing plans at Intel's Oregon facility and elsewhere reported this time, function as a "response" to these political demands.
2. TSMC's Capacity Limits and Pricing Leverage
"Concentration in TSMC" was itself a double-edged sword for NVIDIA. TSMC currently holds a de facto monopoly in the AI market, and rising manufacturing costs and difficulty securing production lines have become a constraint on client companies.
By beginning to outsource "low-volume, non-core products" to Intel, NVIDIA gains a bargaining chip in price negotiations with TSMC. There is also a rational calculation at work to concentrate TSMC's production lines on the highest-value-added, high-end dies, thereby improving overall throughput.
3. Technical Expectations for Intel 18A/14A
Intel, which once fell behind in the miniaturization race, is regaining its technical competitiveness through foundry business reforms under former CEO Pat Gelsinger and now under Lip-Bu Tan. In particular, the 18A process—which introduces new technologies such as "backside power delivery (PowerVia)"—and the subsequent 14A process are increasingly seen as having the potential to rival TSMC's leading-edge nodes. The fact that NVIDIA is starting its adoption with the relatively low-risk I/O die also carries significant meaning as a real-world test of Intel's manufacturing capabilities.
Apple Joins In Too: A Return to Intel for the M-Series
Remarkably, NVIDIA is not the only company considering outsourcing manufacturing to Intel. According to sources, Apple is also reportedly in discussions with Intel regarding an "entry-level M-series processor" for MacBooks.
Apple has a history of abandoning Intel CPUs in 2020 in favor of its own in-house designed Apple Silicon (manufactured by TSMC). The fact that Apple is now once again considering using Intel's fabs suggests that this movement is not simply a whim on NVIDIA's part, but rather a structural "swing back" across the entire US tech industry. Like NVIDIA, Apple appears to be positioning itself to test Intel's manufacturing capabilities starting with low-risk entry-level products, aiming to reduce its dependence on TSMC.
TSMC's Calculated Advantage: Avoiding Antitrust Risk
At first glance, this news might appear to be a negative development for TSMC—a case of "losing a customer." However, industry experts analyze this as a "strategic opportunity" for TSMC as well.
- Easing Monopoly Concerns: TSMC's share of AI chip manufacturing has grown so overwhelming that the risk of becoming subject to antitrust scrutiny by regulators in various countries had been rising. Allowing a portion of the share (particularly the lower-value-added portions) to flow to a competitor like Intel functions as a pressure release valve against regulatory scrutiny.
- Refining Toward a High-Margin Business Model: TSMC will continue to retain the "most technically demanding and highest-margin" portion—NVIDIA's GPU core. By offloading relatively lower-margin processes such as packaging and I/O dies to Intel, TSMC can concentrate its own resources on more cutting-edge R&D and higher-margin projects.
The Dawn of the Multi-Foundry Era
The 2028 "Feynman" chip will likely become a historic watershed moment, marking the transition of the semiconductor supply chain from "TSMC dominance" to a "multi-foundry (diversified sourcing)" model—going far beyond being simply a new AI accelerator product.
The strategy NVIDIA envisions is clear.
"Fiercely protect core performance with TSMC, while securing supply chain resilience and political compliance with Intel."
This is the hybrid strategy at play.
If Intel can establish reliable yields for its 14A process and dependable EMIB packaging by 2028, the balance of power in the semiconductor industry will stabilize dramatically. Conversely, if Intel once again runs into manufacturing trouble, NVIDIA's strategy will be forced to change course, carrying the risk of putting the brakes on the very pace of AI hardware evolution.
What happens at Intel's fabs in Oregon and Ohio over the next two years will determine the balance of AI supremacy in the 2030s.
Sources