Highlights

• AI servers require thousands of multilayer ceramic capacitors (MLCCs) per board to stabilize power delivery, pushing these components into unprecedented regimes of sub-0.5 μm dielectric layers and extreme electric-field stress with near-zero failure tolerance.
• Manufacturers increasingly rely on rare-earth dopants (Dy, Y, Ho) to maintain MLCC reliability under DC bias and high temperature, quietly tying AI infrastructure performance to China's ~90% dominance in rare-earth separation and refining.
• Oxygen-vacancy accumulation in ultra-thin dielectrics has emerged as the dominant reliability threat for AI-grade MLCCs, requiring automotive-level validation rigor and elevating passive components from invisible parts to strategic supply-chain chokepoints.

---

Jung Rag Yoon of Samwha Capacitor’s R&D Center (Yongin, Korea), together with collaborators Seok No Seo (Samwha), Min-Woo Ha (Myongji University), and Moon-Taek Cho (Daewon University College), examine a critical but largely invisible constraint on modern AI infrastructure: multilayer ceramic capacitors (MLCCs)—the sand-grain-sized components that stabilize power delivery in GPUs and data-center power systems.

In their January 2026 review published in the Journal of Electrical and Electronic Materials, the authors argue that AI servers are forcing MLCC technology into an unprecedented regime of extreme miniaturization (sub-0.5 μm dielectric layers), elevated electric-field stress, and near-zero-tolerance failure requirements.

To maintain performance under these conditions, manufacturers increasingly rely on rare-earth-doped BaTiO₃ dielectrics (notably Dy, Y, and Ho) to stabilize capacitance under DC bias and high temperature. That materials solution, however, quietly ties next-generation AI reliability to a geopolitical reality: China’s dominant position in rare-earth separation and refining, particularly for mid- and heavy-rare-earth supply chains.

Google Data Center

Why MLCCs Matter to AI (A Lay Explanation That Investors Should Not Skip)

An MLCC functions as a local energy buffer and high-frequency noise suppressor on a circuit board. AI accelerators draw power in abrupt, microsecond-scale bursts. Without rapid local charge delivery, voltagessag, electrical noise spikes, and systems destabilize. This is why asingle AI accelerator board can incorporate thousands to tens of thousands of MLCCs, and why power-integrity engineering has become a first-order design constraint in data centers. The review underscores a clear trend across the supply chain: AI servers consume far more MLCCs than conventional servers, and demand is rising sharply as compute density and power draw continue to climb.

Multilayer Ceramic Capacitors (MLCCs)

Study Methods and What This Paper Actually Is

This publication is a technical review, not a single-laboratory experimental study. The authors synthesize peer-reviewed research, industry practice, and reliability frameworks across four domains:

• Materials engineering: BaTiO₃ particle size control, grain-boundary behavior, and core–shell dielectric microstructures
• Additives and dopants: rare-earth elements and multivalent oxides used to suppress defects and stabilize dielectric response
• Manufacturing processes: slurry dispersion → tapecasting → electrode printing → lamination → reducing-atmosphere sintering → controlled re-oxidation
• Reliability physics and testing: HALT, TSDC analysis, Weibull lifetime modeling, and the emerging “tipping-point” framework tied to oxygen-vacancy accumulation

As a review, its value lies in consolidating technical consensus and highlighting where failure modes are emerging as MLCCs shrink and AI duty cycles intensify.

Key Findings: MLCC Technology Has Entered a New Stress Regime

1. Ultra-thin dielectrics (<0.5 μm) raise the stakes

Shrinking dielectric layers increases volumetric capacitance but simultaneously amplifies electric-field intensity and defect sensitivity. At these scales, a single weak interface or vacancy cluster can become a catastrophic failure path.

2. DC bias “steals” capacitance—core–shell designs try to steal it back

Under sustained DC bias, BaTiO₃-based MLCCs can lose a substantial fraction of effective capacitance. The review highlights core–shell grain architectures that redistribute field stress and preserve dielectric response across temperature and voltage ranges.

3. Oxygen vacancies become the dominant reliability threat in base-metal electrode MLCC

The shift to Ni/Cu internal electrodes requires sintering in reducing atmospheres, which promotes oxygen-vacancy formation. Over time, these vacancies migrate, accumulate, and degrade insulation resistance, eventually forming conductive paths.

4. AI reliability expectations are converging on “ppm-level failure or else.”

AI data centers operate continuously. A single capacitor failure can disable a board, server, or rack. The paper argues that AI-grade MLCCs are approaching automotive-level documentation and validation rigor, but under a distinct stress profile dominated by electrical transients rather than mechanical shock.

Where Rare Earths Enter the “Passive Component” Story

The strategic takeaway is subtle but important: rare-earth dopants are becoming reliability enablers, not performance luxuries. Elements such as Dy, Ho, and Y suppress abnormal grain growth, regulate oxygen-vacancy behavior, reduce dielectric loss, and stabilize capacitance under extreme operating conditions. In short, rare earths are embedded in the power plumbing of AI, not just in motors and magnets.

The Controversial Intersection: AI Reliability Meets Processing Concentration

While the review itself is technical and Korea-centric, its implications intersect directly with geopolitics:

• The International Energy Agency and other bodies note that China’s dominance is more pronounced in rare-earth separation and refining than in upstream mining, with processing shares commonly cited around ~90% for several categories.
• Security and industrial-policy analysts describe this as structural leverage, particularly for mid- and heavy-rare-earth oxides used in advanced materials.
• Recentexport-control actions covering selected rare-earth categories reinforce that availability is shaped as much by policy as by geology.

REEx takeaway

Even when used in small dopant fractions, AI-scale deployment requires high-purity, process-qualified, and highly consistent inputs. That is precisely where processing concentration matters most. When refining is the bottleneck, small quantities can still be strategically decisive.

Limitations and What Readers Should Not Over-Interpret

• This is a review, not a new dataset. It consolidates knownmechanisms rather than establishing novel causal claims.
• Industry authorship matters. With the lead author based at a capacitor manufacturer, the paper naturally emphasizes manufacturable solutions and may underweight alternative architectures or substitution pathways.
• “China monopoly” is context-specific. Dominance is strongest in separation and refining; upstream mining and certain downstream manufacturing steps are more geographically distributed.
• Substitution is slow. Although alternatives are being explored (as acknowledged in funding disclosures), reliability qualification cycles, cost, and yield constraints limit near-term displacement.

Implications for Investors, Policymakers, and the AI Supply Chain

• AI scaling is a passive-component story. GPUs draw attention; MLCCs keep systems alive. Expect capacitor qualification, failure analytics, and supply assurance to become board-level procurement issues.
• Rare-earth strategy must extend beyond magnets. Ceramic dielectrics and passive components are an underappreciated demand vector.
• Processing chokepoints remain the center of gravity. Concentrated refining capacity means even the most advanced AI hardware inherits upstream vulnerabilities—even when rare earths appear only as dopants.

Yoon, J. R., Seo, S. N., Ha, M.-W., & Cho, M.-T. (2026). Multilayer ceramic capacitors for AI servers and data centers: Challenges, reliability issues, and future technology directions. Journal of Electrical and Electronic Materials, 39(1), 34–51. https://doi.org/10.4313/JEEM.2026.39.1.5 (opens in a new tab)