[ChipTalk] "AI Server Components Also Booked for Five Years"… 'LTA' Model Spreading Beyond Semiconductors to Component Industry

Exploding AI Server Demand Drives Expansion of LTAs for Components
'NCNR' Clauses Take Hold in Supplier-Dominated Market
Production Ramps Up, but Supply Struggles to Meet Soaring Demand

With the explosive growth in demand for big tech (major information technology companies) driven by artificial intelligence (AI), the long-term supply agreement (LTA) model, once standard in the semiconductor industry, has now become established in the AI server component sector as well.

According to the semiconductor industry on July 27, LTAs are being signed for the core components required for AI servers, such as flip chip ball grid array (FC-BGA), high-performance multilayer ceramic capacitors (MLCC), and silicon capacitors. An LTA is a contract in which both the volume and pricing conditions are agreed upon for a period exceeding one year. As investments in AI surge and supply shortages intensify, the number of contracts specifying long-term supply has increased. An industry insider stated, "Recently, the contract periods for components like FC-BGA and MLCC are guaranteed to last at least three to five years."

Schematic diagram of the interior of an MLCC (Multilayer Ceramic Capacitor). Samsung Electro-Mechanics.

AI servers consume far more power than regular servers, with energy usage fluctuating sharply depending on computational demands. As a result, the importance of MLCCs and silicon capacitors—which stabilize voltage and reduce electrical noise—has increased. Package substrates that connect semiconductor chips such as graphics processing units (GPU) and central processing units (CPU) to the server mainboard are also seeing rising demand, as AI semiconductors evolve toward integrating multiple chiplets into single packages.

These components are essential for the stable delivery of power and signals in AI servers, but they require higher performance and reliability than general-purpose products, limiting the pool of manufacturers capable of producing them. Major suppliers of high-performance MLCCs include Samsung Electro-Mechanics, Murata of Japan, and Taiyo Yuden. For server FC-BGA, key manufacturers are Samsung Electro-Mechanics, Japan's Ibiden and Shinko Electric Industries, as well as Taiwan's Unimicron and Nanya PCB. As for silicon capacitors, only a handful of companies such as Murata and Samsung Electro-Mechanics maintain the supply chain.

As a supplier-driven market has formed, "non-cancellable, non-returnable (NCNR)" contract conditions are also becoming common. NCNR terms make it difficult for customers to cancel agreed orders or be refunded for any prepayments. Under such terms, suppliers can invest in equipment and materials based on confirmed orders, while customers are able to secure needed production capacity in advance. However, if future demand decreases, buyers may be forced to take on excess inventory and purchase obligations.

Global AI server component suppliers are recently entering a series of LTAs with major tech companies. In May, Samsung Electro-Mechanics signed a long-term supply agreement worth 1.5 trillion won for silicon capacitors with a large global company. In June and July, the company secured additional contracts for supplying MLCCs for AI servers, worth approximately 454 billion won and 295.1 billion won, respectively. The series of large-scale long-term MLCC supply contracts is considered unprecedented in the industry.

With long-term demand secured, Samsung Electro-Mechanics is also scaling up its production base. The company has invested about 1.9 trillion won in Busan and Vietnam to expand its high-performance FC-BGA production capacity for server and AI applications. It also recently announced a mid- to long-term investment plan to turn Busan into a core base for high-performance package substrates for AI data centers and high-value MLCCs by investing some 15 trillion won by 2040. In Sejong, the company additionally unveiled plans to invest 8 trillion won to establish it as a base focused on advanced package substrates, including FC-BGA.

Competitors are also ramping up their production capacity. Murata of Japan plans to invest an additional 80 billion yen (about 714.7 billion won) in the MLCC sector over the next two years until 2028, on top of its existing 680 billion yen (about 6 trillion won) investment plan. Taiwan’s Unimicron expanded its original facility investment plan for this year by 75%.

However, there is a consensus that these efforts remain insufficient to address the industry-wide supply shortage. According to TrendForce, global AI server shipments increased by 34.6% in 2023, 46.0% in 2024, and 24.2% last year. Shipments for this year are also projected to increase by 28.3% year-over-year, more than double the overall server shipment growth forecast of 12.8%. With the expansion of AI server shipments, demand for key components such as FC-BGA and high-performance MLCCs is also expected to rise.

In fact, new orders in the MLCC market are outpacing shipments. As of the end of last month, the order-to-shipment ratio (BB ratio) was 1.31 for Samsung Electro-Mechanics, 1.30 for Murata, and 1.25 for Taiyo Yuden. The industry average stood at 1.04, exceeding 1. A BB ratio above 1 means new orders outnumber shipments, indicating mounting supply pressures, primarily for AI server and automotive high-value-added MLCCs.

With the supply shortage intensifying, MLCC prices are also showing signs of an increase. It has been reported that Taiyo Yuden informed its customers of forthcoming price adjustments for main products such as MLCCs starting in May.

An industry insider noted, "In AI servers, even a single missing component renders it impossible to complete a product, so securing stable supplies is critical. This will strengthen the business model where suppliers demand long-term purchase commitments to justify large-scale investments, while customers will preemptively secure production capacity in anticipation of supply shortages."

This content was produced with the assistance of AI translation services.