How Tesla and Nvidia Made Korea the Hinge of the AI Chip Supply Chain
Samsung Electro-Mechanics is building AI chip substrate lines with $2.8 billion in co-investment from Tesla and Nvidia — a move that makes South Korea the critical link between America's two biggest AI champions.
The Quiet Realignment
Two American technology giants just made the same bet on a South Korean factory they don’t own. That bet changes everything about how you should think about the AI supply chain.
Tesla is committing 1.1 trillion won and NVIDIA is putting in 2.6 trillion won — roughly $2.8 billion combined — into Samsung Electro-Mechanics’ Sejong facility. The money is going toward FC-BGA substrate production lines, the critical circuit boards that sit between an AI chip and the rest of a system. Mass production starts late 2028. Supply to Tesla begins in 2029 and runs through 2033.
This isn’t a vendor relationship. It’s infrastructure lock-in.
Why This Matters Beyond Korea
The conventional story treats this as another Korean supplier landing a big American customer. That story misses the architecture underneath it.
FC-BGA substrates are the bottleneck nobody is talking about. As AI chips get more powerful, they generate more heat and require denser circuit routing. The substrate holding the chip has to handle both. Japan’s Ibiden dominates this space with over 70 percent market share. Taiwan’s United Microelectronics and Samsung Electro-Mechanics are the only other players producing high-end versions at scale.
The global FC-BGA market was $7.1 billion in 2026 and is projected to reach $13 billion by 2035, according to Business Research Insights. The growth isn’t speculative — it’s mechanical. Every autonomous vehicle, every humanoid robot, every data center GPU needs these substrates in increasing quantities.
Samsung Electro-Mechanics now has two of the world’s most aggressive AI hardware companies financially committed to its capacity expansion. That is an unusual position for a component supplier to hold.
The Tesla Deal
Tesla is not buying substrates for its existing Dojo supercomputing chips. It is building for two products that don’t have commercial volume today: the Cybercab’s autonomous driving computer and the Optimus humanoid robot. Both require complex on-board AI inference. Both need the kind of substrate Samsung Electro-Mechanics will be producing.
The 5-year supply agreement starting in 2029 gives Tesla a guaranteed runway through the period when Optimus is supposed to move from prototype to meaningful production. It also aligns with Tesla’s A15 chip program — the next-generation AI processor that Elon Musk confirmed last September will be manufactured by Samsung Foundry. The A15 will almost certainly use Samsung Electro-Mechanics’ substrates.
Tesla is hedging against the same bottleneck it created. Autonomous driving requires enormous compute density. The substrate supply didn’t scale at the same speed. This deal removes one variable from Tesla’s hardware risk equation.
The NVIDIA Deal
NVIDIA announced its participation at GTC 2026 in Germany on February 21st. The company described the investment as supporting a dedicated production line for AI computing chipset and data center network substrates. The exact amount — 2.6 trillion won — emerged from Maeil Business News’s reporting rather than NVIDIA’s own materials.
NVIDIA’s motivation is structural. Its Blackwell and Rubin architectures are moving to larger packages that require larger, more complex substrates. The company’s current supply chain depends heavily on Japanese and Taiwanese manufacturers. Adding a Korean source with financial backing from the buyer itself is a supply chain diversification play that also deepens NVIDIA’s influence over Samsung Electro-Mechanics’ capacity planning.
The Hinge Position
South Korea occupies a position in the AI hardware stack that is becoming increasingly strategic. Samsung Electronics fabricates chips. Samsung Electro-Mechanics builds the substrates those chips mount on. SK Hynix and Samsung Electronics produce the memory feeding the GPUs. LG Energy Solution supplies the batteries for the vehicles running the autonomous driving stack.
No single country controls this entire chain. The United States designs the chips. Japan supplies the most advanced substrates. Taiwan fabricates the most sophisticated semiconductors. Korea sits in the middle — fabrication, substrate production, and assembly — and is now receiving direct capital commitments from the two American companies most dependent on this hardware.
That positioning creates leverage. It also creates vulnerability. If trade policies shift, if export controls expand, if the US decides that Korean semiconductor infrastructure is too entangled with Chinese demand to remain unrestricted, this entire arrangement faces friction.
What Happens Next
Samsung Electro-Mechanics announced in September that the total Sejong expansion will cost 4.27 trillion won. Tesla and NVIDIA are covering 3.7 trillion won of that through their combined investment. The remaining funding and operational risk fall to Samsung Electro-Mechanics and its parent, Samsung SDI.
The timeline is tight. Construction is already underway. The facility should be completed by end of next year. Equipment installation and trial runs run through mid-2028. Full production starts in the second half of 2028, with Tesla deliveries beginning in 2029.
If Samsung Electro-Mechanics can hit those targets and maintain yield rates on high-end FC-BGA — a significant if, given that Ibiden holds the process lead — the company’s revenue mix will shift dramatically toward AI substrate production. That would be the first major repositioning of a Korean electronics components business toward the AI infrastructure stack.
The broader implication is that the AI hardware arms race is creating supplier alliances that look nothing like the traditional buyer-vendor model. When two competitors in the chip design space co-invest in the same Korean factory’s capacity, they aren’t just securing supply. They’re sharing ownership of a critical node in the infrastructure that will power the next decade of computing.