Why Samsung and Qualcomm's FC-BGA Deal Is a Bigger Deal Than It Seems

On June 22, 2026, Samsung Electro-Mechanics (SEMCO) began mass production of Flip-Chip Ball Grid Array (FC-BGA) substrates for Qualcomm's first data center AI accelerator, the AI200, at its Busan plant . The move is a significant inflection point for both companies — and for the AI hardware supply chain writ large.

FC-BGA substrates are the critical interconnect layer that electrically and thermally connects an AI accelerator die to the system board. Only a handful of global producers can manufacture them at the quality and scale required for data center chips . By supplying these substrates for the AI200, SEMCO is expanding its long-standing partnership with Qualcomm beyond mobile and PC into the hyperscale data center segment . It also validates SEMCO's position as a premier FC-BGA supplier, a status it has been building since it became the first South Korean company to mass-produce server FC-BGA substrates in October 2022 .

Qualcomm's AI Inference Bet

Qualcomm formally announced the AI200, launching in 2026, and the AI250, launching in 2027, in October 2025 — marking its official entry into the data center AI chip market to compete with Nvidia and AMD . The strategy is not to beat Nvidia at training. Analysts note Qualcomm has "absolutely no shot of creating anything that can take on Nvidia in AI training," where Nvidia is expected to generate roughly half of its ~$183.5 billion in data center revenue in fiscal 2026 . Instead, Qualcomm is targeting the rapidly growing AI inference segment — running models after they have been trained .

The AI200 is sold as a complete, liquid-cooled server rack (the "Qualcomm AI200 Rack") accommodating up to 72 accelerators operating as a single system — the same form factor Nvidia and AMD use, making it a drop-in competitor for hyperscalers . Its key differentiator is memory: the AI200 uses 768 GB of LPDDR5X per card, far more memory capacity than any HBM-based accelerator, and the rack delivers a total of 43 TB of memory . Qualcomm argues this LPDDR-based approach yields lower cost and higher capacity than the scarce, expensive HBM that Nvidia relies on .

The company claims the AI200 delivers a lower total cost of ownership (TCO) for inference workloads through better energy efficiency and a cheaper memory subsystem . HUMAIN has already partnered to deploy 200 MW of AI200-based racks starting in 2026 . Per-chip performance numbers (TOPS, TFLOPS) remain undisclosed, making a direct silicon-level comparison with Nvidia's B200/B300 or AMD's MI350 series incomplete .

Competitive Positioning vs. Nvidia and AMD

Nvidia dominates both training and inference with its GPU-based architecture, CUDA ecosystem, and HBM memory. Qualcomm's AI200 avoids direct GPU competition; instead it targets inference workloads where memory capacity — not just bandwidth — matters, such as serving very large models . However, Qualcomm lacks Nvidia's mature software ecosystem.

AMD's Instinct MI300X/MI350 series also targets inference, using HBM3 memory and a CDNA architecture. Qualcomm's pitch is similar: better efficiency, lower TCO, and differentiated memory capacity for inference-specific workloads .

Qualcomm is entering a market where Nvidia and AMD have years of data center relationships, software stacks, and deployment track records. The AI200's success hinges entirely on whether inference buyers value memory capacity and TCO over ecosystem maturity .

Samsung's Wider Substrate Strategy

SEMCO has aggressively pivoted its substrate business toward AI server and data center customers. Beyond the Qualcomm win, SEMCO has secured first-vendor status for FC-BGA substrates on Nvidia's Groq 3 Language Processing Unit (LPU), an inference accelerator integrated into Nvidia's forthcoming Vera Rubin platform, with mass production starting in Q2 2026 . The company is also confirmed to supply FC-BGA for Tesla's next-generation AI6 chips .

Demand is straining capacity. The FC-BGA utilization rate is expected to rise above 80% in 2026, up from roughly 60% currently . Customer demand exceeds current capacity by more than 50 percent, according to CEO Chang Duck-hyun .

To address this, SEMCO is investing heavily. The company has announced a $1.2 billion investment in Vietnam to build new FC-BGA production capacity . R&D spending surged 36% in 2026 as SEMCO reorients its substrate business toward higher-value AI server products . The company's goal is to increase the proportion of high-value FC-BGA (for servers, AI, automotive, and networking) to over 50% by 2026 .

CEO Chang Duck-hyun stated at CES 2026 that FC-BGA lines will run at full capacity in the second half of 2026, and further capacity expansion is under review . SEMCO is also competing with LG Innotek to attract Big Tech investment partnerships for substrate capacity expansion .

Diversifying further, SEMCO signed a 1.5 trillion KRW (~$1.1 billion) silicon capacitor supply contract with an unnamed global tech company for AI applications, running from 2027 to 2028 .

The Bottom Line

SEMCO has positioned itself as a critical, multi-customer supplier for the highest-value AI chip packaging substrates, serving Qualcomm, Nvidia, and Tesla simultaneously. Its capacity is strained by surging demand, and it is investing heavily — both organically (Vietnam, R&D) and through customer partnerships — to scale FC-BGA output. The Qualcomm AI200 win is the latest proof point that SEMCO is now a first-tier player in the AI substrate supply chain, not just a consumer-electronics component maker.