Samsung Electro-Mechanics (SEMCO) and NVIDIA signed a multi-year strategic co-investment deal Wednesday that will convert dedicated production lines at two South Korean factories into ring-fenced FC-BGA substrate capacity exclusively for NVIDIA's next-generation GPU and AI networking chips — a structural supply security move that transforms a capacity-allocation problem the AI industry has quietly struggled with for two years into a long-term bilateral commitment.

Under the agreement, SEMCO will establish NVIDIA-exclusive flip-chip ball grid array (FC-BGA) package substrate lines at its Sejong facility in central South Korea and its Busan plant on the southeast coast. Volumes are set to grow at a double-digit annual rate through at least 2028, at which point NVIDIA is expected to represent a significant share of SEMCO's entire annual substrate production output.

NVIDIA and SEMCO are jointly funding the dedicated lines — a structure that converts the partnership from a conventional supply agreement into a capital co-commitment with legal and operational consequences for both sides. Neither company disclosed the dollar value of NVIDIA's capital contribution.

What the Co-Investment Structure Actually Does

The headline number is not a purchase order — it is a capital commitment. Both companies are co-investing to build out the new dedicated lines, a structure that moves some construction-phase financial risk off SEMCO's balance sheet while giving NVIDIA a stronger legal and operational claim on capacity that cannot be diverted to other customers on short notice.

This is a deliberate design, not a courtesy arrangement. When an AI chip company co-invests in a supplier's factory, it acquires an effective first-lien on that capacity — an option on output that is embedded in the production infrastructure itself rather than negotiated quarterly in the spot market. The deal mirrors co-investment arrangements that Broadcom and other AI chip companies have pursued with substrate suppliers as the FC-BGA layer of the semiconductor supply chain emerged as a genuine constraint on AI hardware delivery.

SEMCO CEO Chang Duck-hyun framed the partnership as a capability integration play, not merely a volume agreement.

"It is a great honor to join NVIDIA, a global leader driving AI innovation, as a key infrastructure partner," Chang said in a statement. "By combining our advanced FC-BGA manufacturing capabilities with our differentiated expertise in high-reliability MLCC technologies, Samsung Electro-Mechanics will fully leverage its strengths as a comprehensive solutions provider to ensure that NVIDIA's AI platforms are delivered to the market on time with the highest levels of performance and quality."

Debora Shoquist, NVIDIA's Executive Vice President of Operations, acknowledged the substrate supply chain as a systemic constraint the deal is designed to address.

"The pace of the AI infrastructure buildout requires us to expand manufacturing capacity on a global scale," Shoquist said. "Our work with Samsung Electro-Mechanics broadens our network of substrate partners and adds manufacturing capacity to help us scale production and deliver the accelerated computing platforms our customers need."

Why FC-BGA Became a Chokepoint — and Why SEMCO Has an Edge No Competitor Can Match

An FC-BGA substrate is the multilayer circuit board that physically and electrically bridges a GPU or AI accelerator die to the server motherboard. The die mounts face-down on the substrate via fine-pitch solder bumps — the "flip chip" arrangement — with solder balls on the substrate's opposite face connecting to the board below.

Building one requires alternating layers of copper conductor wiring and ABF (Ajinomoto Build-up Film), a dielectric insulating material produced by a single Japanese company, Ajinomoto Fine-Techno, which controls approximately 95 percent of the global CPU and GPU-grade ABF supply. Each additional copper-ABF layer pair multiplies the substrate's routing capacity — and its manufacturing difficulty. A consumer CPU package from 2015 required four to six internal layers in a roughly 40 mm footprint. An NVIDIA H100 AI training accelerator requires 12 or more layers in a 120×120 mm (4.7×4.7 inch) package. NVIDIA's B200 generation pushes the die footprint past 800 mm² (1.2 square inches).

The thermal expansion mismatch between ABF film (approximately 30 ppm per degree Celsius) and silicon (approximately 2.6 ppm per degree Celsius) means that at the package sizes AI accelerators now require, controlling warpage during manufacturing demands precision process control that only a handful of fabs globally have demonstrated at volume. Qualifying a new supplier for a 14-layer AI accelerator substrate takes 18 to 24 months under normal circumstances — there is no quick pivot available when a supplier runs out of capacity.

That structural lock-in is precisely why FC-BGA lead times have reached 48 to 56 weeks today, against a historical norm of roughly 12 weeks, according to Kim Min-kyung, an analyst at Hana Securities who tracks the Korean semiconductor supply chain. "The FC-BGA market has become supplier-led," Kim noted, with AI server buyers demonstrating high price tolerance that has pushed price increases above analyst expectations. SEMCO's own substrate utilization reportedly ran at 89 percent in the first half of 2026, with competitor LG Innotek reportedly operating at 94 percent over the same period.

SEMCO's competitive advantage over the other top-tier FC-BGA suppliers — Japan's Ibiden (holding roughly 35 percent of the high-end market), Shinko Electric (approximately 18 percent), and Taiwan's Unimicron (approximately 14 percent) — lies in a capability none of them possesses: SEMCO is simultaneously the world's leading manufacturer of MLCCs (multilayer ceramic capacitors) and a major FC-BGA substrate producer.

MLCCs are passive components that filter power-supply noise. At the operating frequencies of modern AI accelerators, power delivery path inductance — even across a few millimeters — degrades compute performance. The solution is to embed MLCCs and silicon capacitors directly inside the substrate, immediately beneath the chip die, creating a decoupling capacitor only microns from the power-hungry compute logic. No other major FC-BGA supplier manufactures MLCCs at scale. SEMCO makes both, and can integrate them in a single substrate package — the "total solution provider" claim in Chang's statement is technically specific, not marketing language.

SEMCO's ₩6.78 Trillion ($5.11 Billion) Construction Program Is Already Running

Wednesday's announcement did not arrive without context. In late September 2026, SEMCO disclosed a record capital expenditure program of ₩6.78 trillion (approximately $5.11 billion at current exchange rates) directed entirely at FC-BGA capacity expansion. The single largest commitment — ₩4.27 trillion (approximately $3.22 billion) — is earmarked for a new multi-story fab at the Sejong site, which SEMCO described as its largest-ever single-product investment. A further ₩2.51 trillion (approximately $1.89 billion) is allocated to its subsidiary in Hai Phong, Vietnam. Sejong construction runs through May 2028, with mass production targeted for September 2028.

When SEMCO disclosed that investment in September, it noted that "global customer funding support" and "mid- to long-term volume commitments" underpinned the decision. Wednesday's announcement identifies NVIDIA as one of the central customers behind that language — and extends the scope to include existing Busan production lines being reconfigured as dedicated NVIDIA capacity alongside the new Sejong build.

The financial trajectory already reflects SEMCO's strategic pivot toward AI packaging. SEMCO's Packaging Solutions division posted first-half 2026 sales of ₩1.497 trillion (approximately $1.13 billion), up 41 percent year over year. Operating profit surged 237 percent to ₩160.2 billion (approximately $120.8 million). Analysts expect margins to climb further as FC-BGA prices continued rising 10 to 20 percent in the third quarter of 2026, with one securities firm projecting quarterly operating margins approaching 30 percent and potentially rising toward 40 percent by 2027.

Chang Duck-hyun publicly stated that demand for server and data center FC-BGA substrates was running more than 50 percent above SEMCO's production capacity — a gap that drove both the capex program and the receptiveness to NVIDIA's co-investment proposal. He also anticipated that AI data centers — encompassing GPUs, networking chips, and power semiconductors — could represent 60 to 70 percent of total FC-BGA demand within a few years, a reorientation that has since accelerated faster than most analysts projected.

NVIDIA's Korean Supply Chain Strategy: Locking In Capacity Across the Stack

The SEMCO deal is the most significant in a broader pattern of NVIDIA systematically deepening Korean supply relationships to reduce dependence on any single geography or supplier. Earlier in 2026, NVIDIA confirmed that Samsung Electronics, SK Hynix, and Micron have all been certified as HBM4 memory suppliers for its Vera Rubin platform, deliberately distributing memory procurement across all three major DRAM makers rather than allowing concentration. NVIDIA CEO Jensen Huang, during his June 2026 visit to Korea, described the country as a central pillar of NVIDIA's AI ecosystem, citing collaboration across DRAM, robotics, AI factories, and advanced packaging.

For substrates specifically, the concentration of supply among a small number of qualified producers — Ibiden, Shinko, Unimicron, AT&S, and SEMCO together control the vast majority of advanced AI FC-BGA supply — has made it strategically urgent for NVIDIA to establish dedicated, long-term arrangements rather than competing for spot allocations. The ABF film underpinning all of these substrates remains tightly constrained: Ajinomoto's approximately 95 percent market position in CPU- and GPU-grade ABF means that no amount of manufacturing capacity expansion can outrun the film supply without Ajinomoto's own investment, and ABF prices rose approximately 26 percent between 2024 and 2025, with further increases expected. Taiwanese ABF substrate capacity is widely reported as sold out with order visibility stretching two to three years.

By co-investing in dedicated Korean lines, NVIDIA is pre-purchasing a ring-fenced slice of global substrate capacity at a moment when the alternative — competing for open-market allocations from a supplier-led market with 48-to-56-week lead times — carries unacceptable planning risk for a company with hundred-billion-dollar AI infrastructure programs outstanding.

The global FC-BGA market stood at approximately $10.9 billion in 2024 and is projected to reach $24.1 billion by 2031, according to TrendForce — a compound annual growth rate driven by substrate area expansion per chip, higher chip counts per server rack, and new AI-adjacent applications in automotive and humanoid robotics. SEMCO's deal with NVIDIA positions it to capture a disproportionate share of that growth at the margin most important to its business: the complex, high-layer-count AI accelerator substrate segment where MLCC integration creates a genuine technical moat.

What Comes Next for AI Infrastructure Supply

Deliveries to NVIDIA under the new agreement begin in 2026, drawing on existing Busan capacity while the dedicated reconfiguration and Sejong expansion proceed. The volume ramp is expected to be steep: double-digit annual growth in shipments is the publicly stated trajectory, with NVIDIA's share of SEMCO's total substrate output projected to become a cornerstone figure by 2028.

The larger implication — the one not stated in either company's announcement — is structural. The SEMCO-NVIDIA deal is a template. A market in which FC-BGA lead times have reached 48 to 56 weeks, in which new-supplier qualification requires 18 months or more, and in which open-market allocation cannot keep pace with hyperscaler demand growth will push other large AI chip buyers toward the same model: co-invest in dedicated lines at a preferred supplier, accept the reduced flexibility, and in exchange receive guaranteed output that cannot be redirected to a competitor. Google, AMD, and Broadcom are all positioned to face the same strategic calculus. The substrate bottleneck, long overshadowed by wafer and memory constraints in public discussion, has matured into a first-tier supply security problem — and Wednesday's announcement is the first major public evidence that AI chip companies are solving it the same way they eventually solved HBM: with long-term financial commitments to individual suppliers, not spot market bids.

Currency conversion note: all Korean won (₩) figures converted to U.S. dollars at the rate of ₩1,326 per USD as of October 21, 2026. Conversions are approximate.

Frequently Asked Questions

What is an FC-BGA substrate, and why does NVIDIA need them from Samsung Electro-Mechanics specifically?

An FC-BGA (flip-chip ball grid array) substrate is the multilayer circuit board that connects a GPU or AI accelerator chip to the server motherboard. It is built from alternating copper wiring and ABF dielectric film layers — 12 or more for modern AI chips — and the manufacturing difficulty grows with each additional layer and each increase in package size. SEMCO is specifically valuable to NVIDIA because it is simultaneously the world's leading MLCC (multilayer ceramic capacitor) producer and a major FC-BGA manufacturer, allowing it to embed power-filtering MLCCs directly inside the substrate, immediately beneath the chip. No other top-tier FC-BGA supplier (Ibiden, Shinko, Unimicron) makes MLCCs at scale. This integration reduces power noise at the extreme operating speeds of AI accelerators — a technical advantage that is difficult to replicate quickly.

How does the co-investment structure differ from a normal supply contract, and why does it matter?

In a standard supply contract, a chip company places purchase orders and competes for available capacity. In a co-investment deal, the chip company contributes capital to build the production infrastructure itself — giving it a contractual claim on dedicated output that cannot be diverted to other customers. For NVIDIA, this converts a spot-market allocation problem into a structural supply guarantee. For SEMCO, it offloads construction-phase financial risk while locking in a large, predictable customer. The consequence for the broader industry is that other AI chip companies — Google, AMD, Broadcom — now face competitive pressure to secure their own dedicated-lines arrangements before the available qualified substrate capacity is further ring-fenced by similar deals.

Why are AI chip substrate lead times at 48 to 56 weeks when the industry has been expanding capacity for two years?

Substrate capacity expansion takes time to qualify, not just to build. Qualifying a new FC-BGA manufacturing line for a 14-layer-plus AI accelerator substrate requires 18 to 24 months of testing and validation — a process that cannot be accelerated significantly regardless of capital spending. This means that capital committed in 2024 is only now beginning to translate into qualified output, while AI chip demand has continued to compound throughout that period. The ABF film supply adds a second constraint: Ajinomoto Fine-Techno controls approximately 95 percent of CPU- and GPU-grade ABF supply, and switching to an alternative requires its own re-qualification cycle. The 48-to-56-week lead times reflect both the qualification lag and the ABF supply tightness — and they are not expected to normalize until the current wave of capacity investments, including SEMCO's Sejong fab, reaches mass production in late 2028.

What does this deal mean for other FC-BGA suppliers like Ibiden and AT&S?

Every slice of SEMCO output ring-fenced for NVIDIA is capacity that does not enter the open market. For Ibiden, which holds approximately 35 percent of high-end AI FC-BGA supply, and AT&S, which has committed roughly €2.5 billion to a Malaysian expansion, the SEMCO-NVIDIA deal signals that the competitive landscape is shifting from allocation of available supply toward bilateral lock-in agreements with specific customers. Suppliers that establish long-term co-investment deals first are likely to see more stable demand; those relying on spot or short-term orders face a risk of being allocated the customers that leading AI chip companies did not commit to — a less attractive demand profile. The deal also underscores that SEMCO's MLCC integration capability is a meaningful technical differentiator that competing pure-substrate suppliers cannot quickly replicate.