From 259 worldwide in 2015 to more than 1,136 today, hyperscale data centers have grown into a $167 billion core industry and are projected to reach $600 billion by 2030. However, the driving force behind this transformation is not merely scale expansion; it is the “cooling crisis” spurred by AI workloads. The demand for power densities exceeding 50 kW per rack is compelling a comprehensive overhaul of data center cooling technologies and overall architectural design, with liquid cooling emerging as the pivotal breakthrough in this evolution.

Scale Leap: Operational Transformation from Megawatt to Hundreds of Megawatts

The gap between hyperscale data centers and traditional enterprise data centers has long transcended mere scale; in the AI era, this divide has widened even further. Traditional enterprise data centers typically deploy only a few hundred servers, with power consumption ranging from 1 to 5 megawatts. By contrast, hyperscale data center campuses often house millions of servers, with individual campuses consuming 20 to over 100 megawatts of power. A single hyperscale data center can span more than 10,000 square feet and accommodate no fewer than 5,000 servers, while modern hyperscale campuses increasingly adopt a clustered architecture featuring multiple buildings interconnected and operating in concert.

Behind this leap in scale lies the unwavering commitment of leading companies to AI infrastructure. At the 2025Next conference, Google CEO Sundar Pichai made a clear statement: the company will invest $75 billion in building AI infrastructure. His assessment aligns precisely with industry trends: “At this critical inflection point in the technology adoption curve, the risk of underinvestment far outweighs that of overinvestment, and such infrastructure will serve as the core enabler of enterprises’ long-term competitive advantage.” Crucial to realizing this massive investment, however, is overcoming the thermal-management bottleneck posed by the high computational workloads inherent in AI.

Commodity servers are shifting toward hardware systems custom-optimized for specific AI workloads, and this customization is deeply synergized with innovations in cooling technology.

Amazon’s foray into custom hardware is highly representative: its Trainium chip is specifically designed for machine-learning training workloads, while the Inferentia processor focuses on inference workloads. The next-generation Trainium 2 chip not only delivers 30%–40% better price-performance than commonly used GPU-powered compute instances, but also features a built-in cooling-plate interface at the chip-level, enabling direct connection to liquid-cooling systems and thereby reducing heat buildup within the chip.

In his 2024 Letter to Shareholders, Amazon President and CEO Andy Jassy emphasized: “Custom silicon enables us to reduce our reliance on traditional semiconductor suppliers while optimizing power consumption for specific use cases; moreover, co-design with cooling systems further amplifies both performance and cost advantages.”

Google, meanwhile, continues to lead the pack with its Tensor Processing Units (TPUs): the seventh-generation TPU, codenamed “Ironwood,” is custom-built for Google’s AI workloads, featuring a more compact transistor layout and optimized thermal distribution. Paired with an immersion-cooling system, it delivers a 25% performance boost at the same power level. This synergistic hardware–cooling design fundamentally breaks away from the traditional approach of “building hardware first and then addressing cooling,” emerging as the core strategy for reducing costs and boosting efficiency in hyperscale data centers.

The shift toward AI workloads is not only driving upgrades in cooling technologies but also spurring a disruptive rearchitecting of data center network infrastructures—wherein network innovation must be deeply aligned with the layout of cooling systems.

The traditional three-tier network architecture is designed specifically for north–south traffic between clients and servers, which is fundamentally at odds with the east–west communication patterns that dominate AI training. Today, the spine–leaf topology has become the mainstream architecture for hyperscale data centers: this two-tier design fully interconnects every leaf switch (which connects to servers) with every spine switch, ensuring that any inter-server communication requires no more than three network hops. This not only guarantees predictable latency but also enables horizontal scalability by adding more spine switches. However, implementing this architecture depends on a modular liquid-cooling system that places servers and switches in close proximity, thereby preventing reduced cooling efficiency caused by dispersed equipment.

The explosive growth in bandwidth demand is also driving technological upgrades: 400G networks, once considered state-of-the-art, are rapidly being replaced by 800G and 1.6T connections; meanwhile, the Super Ethernet Alliance—comprising Arista, Broadcom, Intel, Meta, and Microsoft—is developing network standards specifically tailored for AI workloads, which could challenge InfiniBand’s dominant position in high-performance computing. At the same time, software-defined networking (SDN) has become essential for managing complex network topologies: hyperscale operators use SDN controllers to automatically configure network paths, balance traffic loads, and respond to failures—all without human intervention—making it indispensable for hyperscale facilities with hundreds of thousands of network ports and providing the network infrastructure needed to support centralized management of cooling systems.
The rapid expansion of hyperscale data centers has given rise to unprecedented power challenges, making the enhancement of cooling system energy efficiency a critical lever for alleviating this pressure.

In the third quarter of 2024, U.S. data center electricity demand reached 46,000 megawatts, with projections indicating a further increase of 35 gigawatts by 2030—equivalent to adding 35 new nuclear power plants. Gartner forecasts that by 2027, 40% of existing AI data centers will face power constraints, as grid limitations in traditional core markets such as Northern Virginia compel operators to seek alternative locations. Meanwhile, the persistently high power consumption of AI workloads—driven by thousands of GPUs running in parallel for weeks or even months—stands in stark contrast to the more flexible power demands of conventional web servers, thereby exacerbating overall power supply pressures.
Liquid cooling technology delivers a 30–40% improvement in energy efficiency, directly reducing power consumption in the cooling subsystem and providing critical support for alleviating energy pressures. At the same time, hyperscale operators are addressing these challenges through innovative electricity procurement models: an increasing number of companies are partnering directly with utilities to develop new generation capacity, often with a strong emphasis on renewable energy to fulfill sustainability commitments—Amazon has already entered into agreements for more than 15 gigawatts of renewable energy capacity, while Google has pledged to achieve net-zero emissions across all its operations by 2030. In short, the energy efficiency of cooling systems has become a key determinant of whether hyperscale data centers can integrate renewable energy and achieve sustainable development.