The GW Era of AIDC Begins: A New Normal for Data Center Construction

The exponential growth in AI computing demands is driving a paradigm shift in the data center industry. The era of gigawatt (GW)-scale power consumption for individual facilities has arrived, necessitating a complete reevaluation of traditional design philosophies and technological stacks. The “Gigawatt-Class Open AI Computing Center Infrastructure Technology Report,” presented at the Open Compute Project Global Summit in 2026, provides a detailed overview of this transformation. The focus has shifted from merely “stacking servers” to engineering data centers as integrated systems.

The Leap from MW to GW

Traditionally, large-scale data centers have operated with power capacities measured in tens of megawatts (MW). However, the explosive increase in AI computing demands has led to single facilities consuming up to 1 gigawatt (GW) of power—representing a scale that is tens of times larger than before.

An infrastructure manager at a leading internet company reportedly questioned their team upon reviewing the plans for a new AI computing center: “We used to build data centers measured in megawatts. Now the plan says gigawatts. That’s a thousand times more. Is this even the same concept?”

This question encapsulates the reality of the industry. Just five years ago, a single rack typically consumed 2.5 kilowatts, and scaling a data center to several megawatts was a challenge. However, with the rise of generative AI applications, the standard power consumption for a single AI rack is now over 30 kW. Future supernode racks are projected to consume 300 kW, 500 kW, or even megawatts.

According to Gartner’s latest forecast released in 2026, global data center power consumption is expected to reach 565 terawatt-hours in 2026, a 26% increase from 2025 levels. This is equivalent to the annual electricity consumption of approximately 230 million households in China. The power consumption of AI-optimized servers is predicted to jump from about 20% in 2025 to 31% by 2026.

Linglan Wang, Research Director at Gartner, noted, “AI computing capabilities are currently constrained by power supply. Ensuring a reliable power supply for data centers has become a new competitive frontier in the global AI race, enabling both scalability and profitability.” The competition for computing power has essentially evolved into a competition for electricity.

Demand for computing power in China is also expanding rapidly. According to industry estimates, the deployment scale of AI chips in China is expected to reach 3 to 4 million units by 2026. Based on the rule of thumb that AIDC power consumption is approximately twice that of AI chip consumption, China’s AIDC power demand in 2026 is projected to exceed 5 GW, with actual construction capacity potentially reaching 6 to 8 GW.

Speed as a Competitive Advantage

The construction of gigawatt-scale AIDCs is governed by entirely new rules compared to traditional data centers, with “speed” being one of the most significant changes.

Previously, it was common for data centers to take two to three years from planning to operation. However, AI companies now demand faster turnaround times. A notable example is the Colossus Data Center built by Elon Musk’s xAI team in Memphis, Tennessee, which achieved its first server deployment just 122 days after construction began—a groundbreaking speed.

Although the project revealed issues such as insufficient cooling system design, it sent a clear signal to the industry: the rules of the game for building computing capabilities have changed, and speed itself has become a core competitive advantage.

Similar pressures have spread to China. In May 2026, the National Data Bureau issued the “2026 Digital Economy Development Key Points,” emphasizing the rapid construction of a nationwide integrated computing network and the coordinated development of data, networks, computing power, and energy resources. For the first time, the coordination of computing and power was included in a government work report.

However, accelerating construction does not mean cutting corners. Building a gigawatt-scale AIDC can cost between $4 billion and $8 billion. It involves multiple emerging technologies, such as power supply upgrades, liquid cooling adoption, high-speed interconnectivity, and smart operations and maintenance. A single misstep in decision-making could result in catastrophic losses.

Innovations in Power Supply

The primary challenge in constructing gigawatt-scale AIDCs lies in the power distribution system. Traditional data centers typically use a 48V/54V direct current (DC) power supply architecture. However, as AI rack power consumption surges from tens of kilowatts to over 140 kW, the limitations of this architecture become apparent. High current levels lead to increased copper losses, larger copper bar volumes, and greater heat dissipation challenges, ultimately hindering improvements in computing density.

To address this, the industry is increasingly adopting 800V high-voltage DC power supply systems. These systems deliver high-voltage DC directly to the rack, with final voltage step-down occurring near the chips. Gao Xiaoqi, Senior Director of the Energy Innovation Division at Century Interconnection, explained that compared to traditional systems, 800V DC significantly reduces current, decreases the size of copper bars, reduces the number of conversion stages, and improves overall efficiency. “800V high-voltage DC power supply has become a necessity for current AIDC construction,” Gao stated.

He Yongzhan, Deputy General Manager of Baidu Intelligent Cloud’s Infrastructure System Department, also revealed that planning and research into 800V high-voltage DC projects are accelerating, with early implementation being a key objective. However, the pace of adoption depends not only on the intentions of major companies but also on the maturity of the entire high-voltage DC supply chain. Essential components, from solid-state transformers to high-voltage connectors and DC circuit breakers, require development to fully realize this technology.

Transition to Liquid Cooling

If power supply is the heart of AIDC, cooling is its respiratory system. The upper limit of traditional air cooling systems is around 20 kW per rack. Beyond this density, fan speeds, airflow design, noise control, and power consumption rapidly deteriorate. Meanwhile, the power density of AI supernode racks is quickly surpassing thresholds of 100 kW and 300 kW.

In June 2026, NVIDIA officially announced that its next-generation Vera Rubin platform would adopt 100% liquid cooling technology, with a cooling fluid operating at 45°C and no fans in the system. Mass production is slated for fall 2026. ByteDance’s AIDC technical specifications released in 2026 also mandated 100% liquid cooling for high-density racks exceeding 21 kW.

An AI server technology expert from Inspur stated, “The penetration rate of liquid cooling is gradually increasing, and many data centers are now being planned as liquid-cooling-native from the outset.” Baidu is also steadily exploring an upgrade path for cooling technologies, evolving its architecture from mixed air-liquid cooling to pure liquid cooling.

Liquid cooling serves purposes beyond temperature control. It aims to enhance energy efficiency and achieve lower power costs per token, a new metric for measuring data center competitiveness in the AI era.

The evolution of liquid cooling technologies is also accelerating. From mature cold plate systems to emerging immersion cooling and two-phase cold plate systems, the technology readiness curve is shifting rapidly. Two-phase cooling, which uses phase-change heat dissipation, can theoretically support chip power consumption exceeding 2,000 W and rack power consumption over 300 kW. The industry views it as a critical pathway for overcoming bottlenecks in next-generation high-density computing.

The market potential is enormous. Industry estimates suggest that if new AI data center racks in China grow to 5.0 GW in 2026 and 7.5 GW in 2027, with liquid cooling penetration rates of 38% and 54%, respectively, the liquid cooling market for data center racks in China will reach 9.8 billion yuan and 21.5 billion yuan. This excludes the aftermarket services market, such as system maintenance and coolant replacement.

Interconnectivity: The Greatest Challenge

While upgrades to power supply and cooling systems have been underway for years with measurable progress, the biggest bottleneck in gigawatt-scale AIDC construction in China now lies in interconnectivity.

Unlike traditional data centers that manage hundreds of chips, gigawatt-scale data centers operate tens or even hundreds of thousands of AI chips. Computing power is not an isolated entity; it functions as part of a network. The bandwidth, latency, and reliability of this network directly determine whether tens of thousands of AI chips can work cohesively.

Interconnectivity between chips is the most significant challenge. He Yongzhan remarked, “The core of breaking through supernode technology lies in interconnectivity. The related technology chain encompasses chips, modules, universal baseboards (UBBs), and high-density rack connectors.” Interconnectivity solutions include copper wire connections within racks, all-optical interconnectivity, and east-west network elements such as optical modules, fiber optics, and switches.

Gigawatt-scale AIDCs demand highly integrated system engineering that intertwines power supply, liquid cooling, high-speed interconnectivity, and intelligent operations and maintenance. The report emphasizes that the competition in gigawatt-scale AIDC construction revolves around comprehensive system capabilities rather than the peak performance of individual components.

Editorial Opinion

In the short term, the implementation speed of 800V high-voltage DC power supply and liquid cooling technologies will be the critical factors determining the success of AIDC construction. These technologies depend on the maturity of the entire industrial chain, and between the latter half of 2026 and 2027, the technical choices and supply capacity of component suppliers and system integrators will significantly impact market competitiveness. Particularly in the Chinese market, there is a risk of a gap between the government’s push for accelerated computing infrastructure development and the actual state of the industrial chain, which could create bottlenecks in construction.

In the long term, the race to construct AIDCs at gigawatt scale may become a decisive factor separating winners and losers in the AI industry. The Colossus project by xAI, which achieved a construction timeline of just 122 days, is setting a new industry standard. However, speed-focused designs may reveal issues such as inadequate cooling capacity. Over the next one to three years, the industry is expected to accelerate its efforts to find the optimal trade-off between speed and reliability in construction. Additionally, the high construction costs of gigawatt-scale AIDCs (ranging from $4 billion to $8 billion) will likely become a barrier to entry for smaller AI companies lacking funding capacity.

References

• “AIDC步入GW时代，未来数据中心怎么建？丨ToB产业观察”, by Leo张ToB杂谈 — 钛媒体, 2026-07-16T10:04:22.000Z (ARR)
• Source URL: https://www.tmtpost.com/8067651.html