The power struggle behind the global AI competition: energy anxiety and the explosive growth of Lenovo (00992) liquid cooling technology.

"We have a huge stock of idle NVIDIA GPUs, but they can only lie in the warehouse because there is not enough power to light them up." Recently, in a podcast show, Microsoft CEO Nadella revealed a harsh truth to the market: the bottleneck of AI competition has shifted from chip supply to power. In the past 3 months, NVIDIA has jumped from a market value of 4 trillion to 5 trillion dollars. The trillion-dollar valuation growth is built on the continuous and exponential growth of AI infrastructure demand. The market expects that global investment in building data centers will reach an astonishing $3 trillion in the future. During the AI craze over the years, people were keen to discuss the nanometer battle of chip processes and the scale of trillion-level parameters in large models, but rarely paid attention to the "invisible cornerstone" that supports all of these- electricity. Now, what seemed like an AI competition dominated by algorithms and chips has quietly evolved into a global competition around power resources, cooling technology, and energy strategy. From the giant data centers being built in Tennessee, USA to the hydroelectric computing hub in the Yangtze River basin in China, the supply capacity and efficiency of electricity have become the core variables determining the competitiveness of countries and tech giants in AI. Energy Crisis under the computing power craze Why is AI becoming more and more energy-consuming? In the past, the energy consumption of AI was mainly concentrated in the model training stage- training a GPT-3 level model consumes the equivalent of the annual total electricity consumption of 300 households. But with the popularization of large language models (LLMs), energy consumption has shifted from "centralized training" to "distributed inference"- every question asked, every image generated by users, is an "inference request". Although a single LLM query consumes only 0.3 to 1 watt-hour of electricity, when daily request volumes reach millions or even billions, the cumulative energy consumption will increase exponentially. Data shows that the energy consumption of current LLM inference workloads is comparable to, or even exceeds, model training; it is estimated that by 2026, global AI inference demand will reach the exawatt-hour (1 exawatt-hour = 1 billion kilowatt-hours) level, a scale big enough to make many countries' power grids "overwhelmed". Former Google CEO Eric Schmidt had warned publicly: "The scale of power demand for AI data centers is something I have never seen in my career." His team's calculations show that by 2030, global AI data centers will need an additional 96 gigawatts of power capacity- equivalent to the annual power output of the entire Sweden or 100 standard US nuclear power plants (each with a capacity of 1 gigawatt). Even more shocking is the cost behind the power. If a data center's power demand is 1 gigawatt (1,000 megawatts), assuming it operates at full load continuously throughout the year (about 8,760 hours), its annual power consumption will reach as high as 8.66 billion kilowatt-hours. This means that its annual power cost could reach 700 million to 876 million dollars. In the future, there will be a large number of even larger data centers whose power requirements may reach 10 gigawatts, and the growth of this power cost is quite terrifying. It can be seen that for tech companies, energy is not only a "cost item" but also a "survival item". According to a report by Goldman Sachs, by 2030, the power demand of data centers will increase by 165%, and the energy consumption of AI inference in operating expenses (OPEX) has exceeded 20%. This means that for every 1% increase in energy efficiency, enterprises can save hundreds of millions of dollars in long-term costs; on the flip side, if power supply is interrupted or costs soar, revenue forecasts for companies will collapse instantly. Today, tech giants have had to play the role of "energy developers," viewing investments in energy infrastructure (CAPEX) as a "prerequisite" for AI deployment- without power, even the most advanced chips and algorithms are just "scrap metal". The Battle between China and the US behind AI Energy Strategies As the most feverishly invested country in global computing power, the US is currently facing an energy shortage dilemma. Currently, requests for the refurbishment of some aging power facilities in the US require waiting for more than 7 years. While the expansion of US data centers often only takes a few years, projects for natural gas power plants (from which most of the power supply for US data centers comes) without existing equipment contracts will have to wait until the next decade to come online. Faced with power constraints, global tech giants have no choice but to start adopting aggressive strategies- building their own power stations to have control over energy supply. The core of this model is the construction of an integrated park consisting of "data center + self-owned power station". Elon Musk's xAI company is one of the "aggressive practitioners" of this strategy. Its "Colossus 2" data center being built in Memphis, Tennessee plans to deploy 550,000 to 1 million AI chips, with a potential investment scale of up to several trillion dollars. To support this "computational behemoth" requiring 1 gigawatt of power, xAI is building a natural gas power plant in neighboring Mississippi, deploying multiple gas turbines for on-site generation and direct power supply. This "local power generation" mode not only avoids transmission losses in the grid but also allows for flexible adjustment of power generation according to computing needs, to deal with "load fluctuations". A larger plan comes from the "Stargate" project jointly promoted by OpenAI, SoftBank, and Oracle. The three parties plan to invest $500 billion in building multiple giant data centers worldwide over the next four years. Larry Ellison, Chairman of Oracle, revealed that the project has already started the construction of 10 data centers in Texas. This kind of energy independence is becoming a core competitive barrier for tech giants, and even countries in the AI industry. Differing from US companies achieving energy independence through the "self-built power plant" model, China is solving the energy needs of AI computing by relying on policy guidance and the advantages of clean energy. In July this year, the dormant Yarlung Tsangpo River was awakened by explosions, and the superpower project with a total investment of 1.2 trillion yuan officially began. With an installed capacity of 60 million kilowatts (equivalent to 3 Three Gorges dams) and an annual power generation of 300 billion kilowatt-hours, this century project is large enough to meet over 75% of the country's AI computing power growth needs, with electricity costs as low as 0.1-0.3 yuan/kWh, a 60% reduction compared to similar computing costs in the US. However, reducing energy consumption is clearly a more fundamental strategy. Liquid Cooling as the "Lifeline" of AI The other side of the energy crisis is the cooling crisis. Relevant data shows that the power consumption of an AI data center mainly comes from two parts: computing, accounting for about 40% of the data center's power demand, and cooling electricity to achieve stable processing efficiency, also accounting for about 40% of demand. The remaining 20% comes from other related IT equipment. Referring back to the data calculated earlier, a data center's electricity bill for cooling alone costs between 280 million to 350 million dollars per year, staggering expenses. Therefore, cooling capacity has become a "physical bottleneck" constraining the improvement of computing power- if the heat generated by chips cannot be dissipated in time, it will lead to performance degradation, shortened lifespan, and even direct burnout. These chips typically have a heat flux density exceeding 50W/cm, with local hotspots reaching up to 150W/cm- which is equivalent to generating 50-150 joules of heat per second on a surface the size of a fingernail, enough to boil a cup of water in a few minutes. If kept at high temperatures, it leads to performance degradation, reduced reliability, and even failures. For AI data centers that need to operate 24/7, these "throttle" effects and failure risks directly impact service stability and computing output efficiency. In such extreme cases, cooling AI chips with air is like trying to cool a boiler with a fan- completely futile, and unable to meet the demands of AI chips. Therefore, more advanced cooling technologies have become a necessary solution for lowering electricity costs in data centers, forcing the industry to switch to liquid cooling technologies. The core advantage of liquid cooling is its "efficient heat conduction" - liquid has a heat capacity of over 4 times that of air, so it can absorb heat more quickly. By directly contacting the chip or immersing the server, liquid cooling technology can raise the limit of heat flux density to over 300W/cm, easily meeting the cooling needs of current and future AI chips. Liquid cooling is not an "optional solution" but a "necessity" for deploying the latest AI chips. On a global scale, tech giants and professional manufacturers are accelerating the layout of liquid cooling ecosystems. Leading companies such as NVIDIA, Meta, Microsoft, Google, etc., have fully adopted liquid cooling in their AI infrastructure. Therefore, the explosion of AI computing power is driving the liquid cooling market into a period of rapid growth, forming a new hot spot for supply chain investments. The global data center cooling market is expected to grow from $18.78 billion in 2025 to $42.48 billion in 2032, with a compound annual growth rate (CAGR) of 12.4%. The domestic liquid cooling market also shows trends of technological diversification and accelerated growth. Therefore, we see in the capital market that the liquid cooling server segment has been continuously surging over a period of time, with related concept stocks on the rise. Inspur's cold plate liquid cooling solution has been applied to multiple large data center projects, especially in fields with high stability requirements such as finance and government affairs. Huawei mainly adopts cold plate cooling technology in its liquid cooling solutions, and its "air-liquid mixing" cooling technology has been applied in multiple AI data center projects. As the third-largest server manufacturer globally, Lenovo is one of the tech companies that have been early movers in the liquid cooling race, covering technical routes such as cold-plate, immersion, and spray cooling. Lenovo: The Underrated Player in Liquid Cooling As a "veteran" in the liquid cooling field for nearly 20 years, Lenovo has formed comprehensive competitive advantages in technology accumulation, scale application, and ecosystem construction, and the real value of its liquid cooling solution is gradually being recognized by the market, with a clear and solid long-term growth logic. According to the latest Q2 financial report for the 25/26 fiscal year, Lenovo's Neptune liquid cooling technology revenue increased by 154% year-on-year (compared to a 68% year-on-year growth rate in Q1), showing rapid growth. Meanwhile, this fiscal quarter, Lenovo ISG achieved revenue of nearly 30 billion RMB, a year-on-year increase of 24%, maintaining a growth trajectory for multiple quarters and improving its operating profit margin by 1.2 percentage points quarter-on-quarter, indicating that business profitability is stabilizing and rising. AI infrastructure-related businesses maintained rapid growth, with strong order reserves, and high double-digit growth in AI server revenue. With the growth rate leaping from 68% to 154%, it not only reflects the rapidly increasing market value of liquid cooling technology but also demonstrates Lenovo's strategic foresight and execution in the AI infrastructure domain. Data shows that by adopting a 100% full-cover cold plate liquid cooling design, Lenovo's "Neptune" liquid cooling solution reduces system power consumption by 40% compared to traditional air cooling, and PUE values can be reduced to below 1.1; the "double loop" phase change immersion cooling system released this year achieves precise temperature control of the phase change chamber and significantly improves the boiling heat transfer efficiency through innovative external single-phase heat exchanger design, doubling the heat dissipation capacity compared to traditional solutions, and the system's PUE can be lowered to 1.035. As of the third quarter of 2025, Lenovo's Neptune liquid cooling system has been deployed globally in over 80,000 sets, covering key areas such as artificial intelligence, supercomputing, government affairs, finance, automotive, etc. In terms of technical maturity, Lenovo's liquid cooling solutions have gained many construction experiences domestically and internationally, and have been involved in the formulation of industry standards such as "Technical requirements and testing methods for liquid cooling systems in data centers", positioning the company at the forefront of the industry in terms of technological discourse. Rich landing cases are the best evidence of Lenovo's strength in liquid cooling technology, and its solutions have withstood the test in many major projects, demonstrating strong scene adaptability. For example, to support the operation of Geely Xingrui Zhi AI Center, Lenovo customized the Neptune liquid cooling solution, optimized the heat dissipation path design for the Beijing Vastdata Technology processing requirements in automotive R&D processes, and controlled the annual average PUE to 1.1, reducing about 3,179 tons of carbon emissions per year. In overseas markets, liquid cooling solutions have been provided for projects such as the Barcelona supercomputing center, the Korean Meteorological Administration, and the Canadian Meteorological Administration. Goldman Sachs' latest research report points out that the global server cooling market is undergoing structural growth opportunities. With the expected increase in AI server shipments and the rapid rise in liquid cooling penetration, it is expected that the total global server cooling market size will achieve year-on-year growth rates of 111%, 77%, and 26% from 2025 to 2027, reaching $17.6 billion in 2027. In this market, Lenovo, with its threefold advantage of "leading technology, rich cases, and a complete ecosystem", is expected to continuously increase its market share. The global explosion of AI has turned energy infrastructure into a new, high-growth asset class. At the same time, it forces enterprises to internalize risk management and liquid cooling's efficiency becomes an important solution path. Operators who achieve an extremely low PUE through liquid cooling not only lower operating costs but also reduce their impact on the power grid. Therefore, the essence of this global AI competition will ultimately be a dual investment wave in power and heat management technology. And providers of liquid cooling technology represented by Lenovo will become important beneficiaries in the future.