Data Centers: The Computational Foundation and Systemic Transformation in the AI Era

Data centers—the physical infrastructure underpinning artificial intelligence, cloud computing, and the Internet of Things—are undergoing an unprecedented transformation. They are no longer merely server rooms housing IT equipment, but have evolved into complex systems engineering endeavors. Their scale, energy consumption, architecture, and strategic significance are being redefined. From traditional enterprise hosting facilities to the AI-driven intelligent computing centers of today, data centers stand at the intersection of technology, energy, and capital, and the logic governing their evolution is being fundamentally rewritten.

I. The Exponential Surge in Compute Demand: Scaling Up and the Electricity Bottleneck

The explosive growth of artificial intelligence has propelled the scale of global data center construction and capital expenditure to historic highs. Hyperscale cloud service providers are engaged in an intense "arms race" around compute infrastructure. However, this scaling up is encountering real-world bottlenecks.

First, the changing nature of workloads is effectively splitting the data center in two. The industry must simultaneously accommodate two distinct AI modes: large-scale training and distributed inference. Training workloads require tens of thousands of GPUs to be integrated into tightly coupled clusters, where latency and node spacing are critical. Inference workloads, by contrast, prioritize broader availability and responsiveness. These divergent requirements cascade down to the facility level, forcing a single data center to simultaneously support highly synchronized systems and user-facing distributed workloads, fundamentally increasing its underlying complexity.

Second, this complexity is compounded by a sharp rise in rack density. Engineers at Google have noted that the industry has broken through the density thresholds of the past decade—racks that once drew 30 to 40 kilowatts have now entered the hundreds-of-kilowatts range, with some designs approaching the megawatt scale. This shift has given rise to a "bimodal" environment: traditional compute and storage infrastructure follows a gradual density curve, while AI systems climb a much steeper trajectory. Under these high-density conditions, a data center is no longer about designing a single rack—it is about designing an integrated system.

Ultimately, power supply—rather than compute capacity itself—is emerging as the primary constraint on data center growth. As rack density approaches the megawatt level, traditional AC power distribution architectures are proving inefficient due to multi-stage conversion losses, high copper consumption, and cooling limitations. A Deutsche Bank report notes that the core debate has shifted from "whether power architecture needs to evolve" to "how quickly the industry will transition to 800V DC." The 800V DC architecture, through centralized rectification and reduced conversion stages, can compress the typical 5% to 10% power loss incurred in AC-to-DC conversion. Data from NVIDIA shows that, when transmitting equivalent power, 800V DC can save 45% in copper conductor usage compared to a conventional 415V AC system.

However, the transition to 800V DC will not happen overnight. In its initial phase, it is more likely to be deployed as rack-level power supplies layered atop existing AC infrastructure, creating a hybrid AC-DC architecture. At the same time, operators are actively introducing energy storage systems to cope with the increasingly volatile fluctuations of AI workloads—training clusters generate sharp dynamic load patterns that can propagate all the way to power plants, forcing real-time adjustments in generation output.

II. The Generational Revolution in Cooling: From Air Cooling to Full Liquid Cooling

As per-rack power density for AI training clusters has surged from a few kilowatts to tens or even hundreds of kilowatts, traditional air cooling has reached its physical limits. At the 2026 Data Center World conference, industry consensus had crystallized: liquid cooling is no longer an option—it has become a foundational requirement for high-density AI systems.

The underlying logic of liquid cooling is undergoing a fundamental shift. Traditional data centers had to keep ambient temperatures very low for fans and air conditioning to effectively remove heat, consuming enormous amounts of electricity. NVIDIA's Rubin platform takes a different approach: since the ultimate goal is to extract heat from inside the chip, the coolant does not need to be "ice water"—it can actually be warmer. Its coolant inlet temperature can reach 45°C, flowing through the chips and exiting at approximately 55°C. As long as the cold plate keeps the chip surface temperature within normal operating range, the chip runs at full speed. This architecture reduces energy consumption while nearly eliminating water usage, using a closed-loop system based on dry coolers that avoids the evaporative water loss of traditional cooling towers.

The benefits of full liquid cooling are substantial. NVIDIA estimates that a 50-megawatt hyperscale data center switching to liquid cooling infrastructure can save over $4 million annually in combined energy and water costs. Cooling systems once consumed 40% of a data center's electricity; full liquid cooling fundamentally changes that equation.

Liquid cooling technology is also advancing to deeper levels. Zhengzhou's National Supercomputing Internet Core Node has adopted domestically developed phase-change immersion cooling technology, submerging equipment in fluorinated liquid for cooling. The fluorinated liquid has a boiling point of around 50°C, while equipment operating temperatures range from 80°C to 90°C. As equipment temperature rises, the fluorinated liquid boils and vaporizes; after being cooled by a condenser, it returns to liquid form, achieving a closed-loop cycle. The team also overcame process challenges in applying diamond-copper composites to enhanced boiling components, achieving stable large-scale application. Intel, in collaboration with local ecosystem partners, has released a dual-channel cold-plate full-liquid-cooling server. Through "memory sleeper cold plate technology" and "SSD cold plate patent technology," it achieves high-percentage liquid cooling coverage of critical heat sources, reducing data center PUE to below 1.1.

It is worth noting that the proliferation of liquid cooling brings a new challenge—water consumption is becoming both a sustainability issue and an operational risk. The industry is calling for data centers to minimize water usage at the design stage wherever possible.

III. Compute-Energy Synergy: A Systemic Engineering Approach to the Energy Challenge

The energy consumption of data centers has become a significant challenge for the digital economy. Data shows that China's data center electricity consumption grew from 130 billion kWh in 2022 to 196 billion kWh in 2025, and is projected to exceed 700 billion kWh by 2030, accounting for over 5% of total national electricity consumption. In the energy mix, IT equipment is the dominant consumer (45%-60%), cooling systems are the second-largest segment (30%-40%), and power distribution systems account for approximately 5%-10%.

However, compute-energy synergy faces an "impossible triangle": balancing security and adequacy, green and low-carbon, and economic efficiency.

On the security and adequacy front, data centers demand extremely high power stability—a millisecond-level voltage interruption can disrupt large-model training, potentially costing millions of dollars in losses. Additionally, data center construction typically takes 12 months, while power infrastructure planning and construction requires 3-5 years. This timeline mismatch can leave data centers unable to connect to the grid in a timely manner.

On the green and low-carbon front, policy has explicitly mandated that green electricity consumption at newly built data centers in national computing hub nodes must reach 80%. However, green power trading carries a premium, and data center operators' primary goal is cost reduction—they are often unwilling to pay extra for green attributes. The structural contradiction between the intermittency of renewable energy output and the rigid load requirements of data centers remains acute.

On the economic efficiency front, some data centers, driven by scale rather than market demand, suffer from low capacity utilization—even after five years of operation, data centers in northwest China have achieved only 30% of planned capacity.

In response to these triple challenges, the industry is exploring a collaborative path between "compute scheduling" and "power regulation." Major tech companies and telecom operators have mature compute scheduling technologies that can allocate computing resources to different users on demand. The "regulation" desired by the power system, however, involves shifting compute tasks across time or space—from periods of low renewable energy output to high-output periods, or from electricity-constrained regions to relatively abundant ones. After the first large-scale compute-energy synergy green power direct-supply project in Zhongwei, Ningxia, reached full capacity, its annual power generation reached 4.3 billion kWh, equivalent to reducing carbon emissions by 3.65 million tons annually. Nevertheless, experts point out that most pilot projects of this nature remain in the experimental stage, with insufficient economic incentives being the core bottleneck—even when participating in virtual power plants or peak-valley pricing schemes, the benefits often fail to offset the risks of reduced compute reliability.

IV. Future Outlook: From Point Solutions to System Design

The data center industry is undergoing a profound shift from "point solutions" to "system design." HPE's 2026 network development forecast notes that "AI-native" will become the core concept defining the next-generation data center, following in the footsteps of "cloud-native." AI is no longer just an application workload running on top of data centers—it is becoming deeply embedded in the operational logic of the data center itself.

The data center of the future will gradually form a closed-loop system: capable of self-learning based on real-time telemetry data, proactively identifying risks and automatically adjusting operating states, and even negotiating energy prices with utilities. Networking will become the starting point for data center design in the AI era—the extreme demands of trillion-parameter model training on bandwidth, latency, and synchronization make the network a critical factor determining overall system performance. Security will also evolve from perimeter defense to an "inherent language" built into the network architecture itself, bringing the "zero-trust data center" from concept to reality.

In such an environment, the data center will no longer revolve around point technologies, but will operate as an integrated system with sensing, decision-making, and self-optimization capabilities. The real challenge for enterprises lies not in whether to adopt new technologies, but in how to integrate them into an efficient, reliable, and sustainable system.