The AI infrastructure story has entered its second act. The first wave of investment chased silicon chips and the fabs that produce them. Now the market is confronting a different set of constraints: the physical ones. Heat must be removed. Power must be delivered. And increasingly, the companies solving these problems are headquartered in Taiwan.
For years, power supplies and cooling systems were treated as commodity inputs, sourced from multiple vendors with minimal differentiation. A standard air-cooled server could be designed with off-the-shelf fans and heat sinks. That era is ending fast. AI servers running Nvidia’s GB200, GB300, and forthcoming Rubin platforms require integrated thermal and power solutions engineered in close collaboration with GPU architects and system integrators. As rack power densities have surged from 20 kW to 120 kW and beyond, the companies building these systems have been elevated from component suppliers to strategic partners embedded in the design cycle.
The numbers tell the story. Taiwan’s power and cooling suppliers posted record revenues in 2025, riding the same Blackwell ramps that propelled the island’s AI server makers past US$455 billion in combined sales. The global data center cooling market reached approximately US$20 billion in 2024 and is projected to hit US$50 billion by 2030. The power market follows a similar trajectory, expected to expand from US$35 billion to over US$50 billion in the same period. But these aggregate figures understate the explosive growth occurring in the AI-specific segments where Taiwan’s suppliers concentrate.
What gives Taiwan its edge? The same ecosystem density that made it dominant in semiconductors. The island’s concentration of AI server ODMs, including Foxconn, Quanta, Wistron, Wiwynn, and Inventec, creates a natural gravity well for thermal and power suppliers. Cold plate manufacturers and power supply designers can iterate with customers in days rather than weeks. Liquid cooling components demand tolerances and leak-proof reliability that favor Taiwan’s mature precision manufacturing base, refined over decades of heat pipe and vapor chamber production. And unlike Western competitors that typically specialize in either power or thermal management, Taiwan’s leading suppliers offer converged solutions spanning both disciplines. Delta Electronics’ “Grid to Chip” architecture exemplifies this approach, delivering power and thermal management as unified rack-scale systems.
Two technology transitions are reshaping this landscape. The shift from air cooling to liquid cooling represents the most significant architectural change in data center thermal management since the industry’s inception. The physics are straightforward: air simply cannot keep pace with GPU power densities that now exceed 1 kW per accelerator. Direct-to-chip cooling, where cold plates mounted on GPU packages circulate liquid coolant, has become the baseline for AI servers, achieving up to 30% improvement in power usage effectiveness. Immersion cooling is transitioning from niche experimentation to mainstream deployment, with vendors at the 2025 OCP Global Summit showcasing racks exceeding 1 MW per bay. Nvidia’s Rubin platform, scheduled for mass production in 2026, will mandate liquid cooling entirely, eliminating hybrid air-cooled configurations and accelerating the transition timeline.
The second shift is equally fundamental. The migration from traditional 48–54V rack power to 800V high-voltage DC systems addresses the brute physics of delivering megawatt-scale power. At lower voltages, the current required creates severe resistive losses, heat buildup, and excessive copper usage. The jump to 800V cuts current by a factor of nearly 15. The architecture also enables easier integration of on-site renewables and battery storage, complements liquid cooling by reducing heat at the electrical level, and contributes to the power efficiency ratings that leading facilities now target at 1.05 or below. Industry analysts expect broad 800V adoption by 2027, but Taiwan’s ecosystem is already shipping systems for hyperscaler deployments.
Looking ahead to 2026, Taiwan’s suppliers face unprecedented demand. The Rubin ramp will drive step-function increases in cooling and power requirements. Hyperscalers have approximately US$500 billion in planned spending. Sovereign AI initiatives across Asia, Europe, and the Middle East are creating new demand for turnkey solutions. Based on projected Taiwan AI server revenues of US$700–800 billion and historical power and cooling content of 8–12% of system value, the addressable market for AI servers alone could reach US$60–90 billion.
