Samsung's 6.8 Trillion Won Bet on the Substrate War Behind AI

The chokepoint nobody talks about

The global AI chip shortage has been framed as a lithography problem — who can print the smallest transistors, who controls the most advanced EUV tools, who gets台积电 to prioritize their orders first. That framing was never complete. The real bottleneck is migrating upstream into packaging, and Samsung Electro-Mechanics’ announcement on September 28 makes that shift unmistakable.

The company revealed it will invest 6.78 trillion won — roughly $4.7 billion at current rates — to expand FC-BGA (flip-chip ball grid array) substrate capacity at its Sejong facility in South Korea and a new line in Vietnam. The Sejong expansion closes in May 2028; Vietnam follows a month earlier in April 2028. But the headline number is only part of the story. What actually matters is who is paying for it.

According to multiple industry sources cited by ZD Korea, Samsung Electro-Mechanics is building these production lines with advance payments from a global big-tech customer. The company has not named the customer — a standard practice in semiconductor supply agreements where pricing and volume are negotiated under mutual NDA. But the context is revealing. The same source confirmed that Samsung entered NVSwitch FC-BGA supply earlier this year, providing substrates for the multi-GPU interconnect chips that sit at the heart of AI cluster networking. Since Q2, Samsung has also been supplying FC-BGA substrates for Groq’s LPU-based inference accelerators.

This is not a diversified customer base. This is one customer, possibly two, funding a near-doubling of premium substrate capacity. The money trail points squarely at the players building the largest AI data center clusters — companies that cannot afford to leave their silicon on the table because they cannot package it fast enough.

Why substrates now

FC-BGA substrates are the interconnect layer between a bare die and the PCB. In AI accelerators, they do more than route signals — they manage thermal dissipation, maintain signal integrity at multi-gigahertz frequencies, and provide the mechanical stability needed for flip-chip bonding, which has largely replaced wire bonding in high-performance packages because of superior electrical and thermal characteristics.

As AI models scaled from training to inference at massive scale, the packaging requirements changed. Each GPU or LPU now requires tens of thousands of MLCCs (multilayer ceramic capacitors) for power delivery — up to 20,000 per chip, and potentially 600,000 per server rack. That is ten times the capacitor count of a traditional server, and it creates a compounding demand spiral: more capacitors mean larger, more complex substrates, which means fewer suppliers can make them, which means lead times stretch and prices rise.

Samsung Electro-Mechanics already has MLCC capacity in play. On September 1, the company disclosed a 1.1 trillion won MLCC supply contract with a “global large enterprise” — a label that in this context almost certainly refers to the same customer funding the FC-BGA expansion. The MLCC and substrate businesses reinforce each other: Samsung can offer a bundled supply agreement covering both the capacitors and the interconnect layer, a combination that is increasingly valuable to customers trying to de-risk their most constrained inputs.

The packaging bottleneck is a supply bottleneck

n
No major semiconductor fabless or integrated device manufacturer can ship a competitive AI accelerator without FC-BGA substrates in volume. TSMC, Intel, and Samsung Foundry all produce the dies. But none of them can turn those dies into shippable packages without substrate capacity — and that capacity is concentrated. Advantest, Samsung Electro-Mechanics, LG Innotek, and a handful of Japanese suppliers including Ibiden and Shin-Etsu Chemical account for the vast majority of premium AI-grade substrate output.

Ibiden, the Japanese specialist, has long dominated the highest-end AI substrate market, particularly for NVIDIA’s GPU interposers. Samsung’s push into FC-BGA at this scale is a direct challenge to that position, and the advance payments from big-tech customers are Samsung’s insurance policy — lock in demand before the capital even hits the ground.

The economics are favorable for Samsung in one critical way: substrates are a lower-margin business than foundry services, but they are also less capital-intensive per unit of revenue and less cyclical. Once a substrate line is qualified for a customer’s design, that customer rarely switches. The qualification cycle runs 12 to 18 months, and the switching cost is enormous. This means Samsung’s advance-payment model is not just about financing capex — it is about securing multi-year revenue visibility on a product that will remain in short supply through at least 2029.

What happens next

The immediate consequence is straightforward: Samsung’s Sejong and Vietnam lines will come online in 2028, adding significant FC-BGA capacity just as the next generation of AI accelerators demand even larger and more complex substrates. The question is whether Samsung can capture enough share from the Japanese incumbents to make this investment profitable, or whether it ends up as a second-source supplier with limited pricing power.

A more consequential scenario involves the vertical integration play. Samsung Electro-Mechanics sits inside the Samsung Group, which also operates Samsung Foundry and Samsung Data Memory. A customer that signs an FC-BGA and MLCC supply agreement with Samsung Electro-Mechanics is effectively strengthening its relationship with the broader Samsung semiconductor ecosystem. That is not a coincidence — it is a strategy, and one that other Korean suppliers will likely emulate.

For the AI infrastructure market, the packaging bottleneck is now the binding constraint. Chiplet architectures, co-packaged optics, and advanced substrate designs are all converging on the same problem: there simply are not enough premium substrates to go around. Samsung’s 6.78 trillion won bet is a signal that the industry expects that gap to persist for years, and that the companies controlling substrate capacity will extract rents accordingly.