Executive Summary

With the explosive growth of artificial intelligence (AI) and large-scale models, the demand for computing power in data centers has surged sharply, and the corresponding power demand has also risen rapidly. Data from various countries show that the load growth of AI data centers far exceeds overall electricity consumption growth. One report predicts that global data center electricity consumption will double by 2030. In China, reports from the Ministry of Energy and the industry indicate that data center electricity consumption was approximately 77 TWh in 2022 and is expected to rise to 400 TWh by 2030. Some studies even suggest, based on a pyramid-shaped growth trend, that demand could reach 600 TWh by 2030. Currently, data center electricity consumption in China accounts for less than 3% of total national electricity consumption, but the growth rate is astonishing. For example, in Gui'an New District, Guizhou, data center electricity consumption grew by 452.7% year-on-year in the first five months of 2025. The U.S. market also anticipates a rapid increase in data center load by 2030, leading to power supply gaps. Power shortages have become a key bottleneck for AI computing power development. Morgan Stanley estimates that by 2028, U.S. data centers may face a power gap of up to 13–44 GW (approximately 20%). In this context, stakeholders are seeking diverse power supply and energy-saving solutions, ranging from grid upgrades and on-site power generation to renewable energy and energy storage technologies, as well as advanced cooling and energy consumption optimization. Each solution has its own cost, scalability, and deployment cycle, requiring comprehensive consideration.

This report systematically reviews the power and PUE requirements of data centers in the AI/large model phase, summarizes current power supply and energy-saving solutions along with major vendors, analyzes future potential technologies and their maturity, evaluates market size and investment opportunities in related sub-sectors, and finally proposes short-term/medium-term/long-term key investment recommendations. Using visualization tools such as tables and Gantt charts, it compares the characteristics of various solutions and key companies, providing investors with clear action guidance and risk warnings.

Problem Definition

• Surge in computing power and electricity demand: AI large model training and inference require sustained high computing power, driving IT equipment such as servers and GPUs to operate at high loads. Stanford research shows that training GPT-3 once consumes approximately 1.28 GWh of electricity. With increasing model scale and real-time applications, the power density of a single rack has risen from the traditional 10–30 kW to 120–132 kW, expected to reach 600 kW by 2027, and megawatt-level racks by 2030. High-density computing brings cooling challenges, increasing data center infrastructure energy consumption (PUE). China has set a target for large data centers to achieve a PUE below 1.25 by 2025, while countries like Germany have less stringent requirements (1.5 by 2027). Economically developed eastern regions have high demand for computing power, but power supply is becoming increasingly tight. Therefore, the "East Data West Computing" project promotes the construction of data centers in western new energy regions to alleviate pressure in the east.
• Current power shortage and regional distribution: Globally, data center capacity is highly concentrated in the U.S., EU, and China. The IEA report notes that the U.S., China, and the EU currently account for about 82% of global data center capacity, and over 85% of future新增 capacity will still be concentrated in these three regions. This puts immense pressure on local grids. For example, in the "Data Center Alley" area of Virginia, U.S., commercial electricity consumption grew by nearly 30,000 GWh from 2019 to 2025, second only to Texas. Morgan Stanley predicts that the U.S. may face a 13–44 GW gap by 2028. Data center electricity consumption in China accounts for about 0.9%–2.7% of total national electricity consumption, but growth has slowed in recent years, and there are significant discrepancies in various predictions. According to the National Energy Administration, data center electricity consumption was 77 TWh in 2022, expected to be 150–200 TWh in 2025, and possibly 400 TWh by 2030. Goldman Sachs has a more aggressive estimate, approaching 600 TWh by 2030. Regionally, energy-rich western regions such as Guizhou and Inner Mongolia encourage data centers to consume local wind and solar green electricity through supporting policies, while southeastern coastal areas already have high electricity loads, requiring incremental projects to rely on long-distance transmission and diversified energy sources.
• PUE and energy efficiency requirements: The energy efficiency of data centers is typically measured by PUE. Various regions in China have imposed stricter PUE requirements: cities like Beijing and Shenzhen have set PUE requirements for large data centers at 1.2–1.25; in the "East Data West Computing" plan, the PUE target for eastern nodes is 1.25 and for western nodes is 1.2. Under the "dual carbon" policy, the PUE threshold for new large data centers has been reduced to 1.3 or even lower. Efficient cooling and energy efficiency optimization have become important directions. For example, liquid cooling technology can reduce PUE to below 1.1. Overall, AI data centers have extremely high requirements for power supply stability (zero downtime), and urgently need to improve IT equipment utilization, reduce cooling and idle power consumption to improve PUE.

Existing Solutions

Grid-Side Solutions

• Expansion and dedicated lines: Building additional power plants (thermal, nuclear, hydro, etc.) and ultra-high voltage transmission lines is a traditional and fundamental means of expansion. Chinese power companies such as State Grid and Southern Grid have invested trillions of yuan in expanding the grid and advancing ultra-high voltage projects to support the transmission of western new energy to the east. Large-scale grid upgrades have long cycles (typically 5–10 years) and high costs, but once completed, they can comprehensively increase power carrying capacity. Advantages: Stable power supply, large scale, significant long-term effects; Disadvantages: Huge investment, long construction period, requires policy coordination.
• Demand response and time-of-use pricing: Incorporating large users like data centers into demand response mechanisms, using peak/valley electricity prices or real-time pricing to incentivize off-peak electricity use. China currently mainly uses peak/valley time-of-use pricing policies, with limited effectiveness. The U.S. and some EU regions are also piloting data center load control, such as temporarily limiting non-critical loads of data centers during grid stress. Advantages: Reduces peak demand through market incentives, alleviates short-term tension; Disadvantages: Significant impact on data center operations, requires fine scheduling and intelligent control, currently low operator participation willingness.
• Microgrids and virtual power plants: Deploying local microgrid systems that integrate generation, storage, and consumption, coordinating data centers with local photovoltaic, wind, and energy storage. For example, the Tencent Huailai Dongyuan Data Center "wind-solar-storage" demonstration project integrates 11MW rooftop PV, 150kW wind power, and 1.376MWh energy storage to achieve multi-energy complementation. Advantages: Can locally utilize renewable energy to reduce dependence on external grids, while energy storage can smooth fluctuations; Disadvantages: High site requirements, large upfront investment, complex construction and scheduling.

Figure: The rooftop and site of the Tencent Data Center in Huailai, Hebei, have installed 11MW photovoltaic and wind power generation facilities, providing clean electricity to the data center through a PV + wind + energy storage microgrid.

On-Site Power Generation

• Gas turbines and gas generators: Gas turbines offer high efficiency (simple cycle 40%, combined cycle higher), fast start-up (within minutes), and low emissions, and are widely used for grid peak shaving and backup power. In the AI data center field, gas turbines can serve as primary or backup power supplements, significantly improving system stability. It is estimated that demand for gas turbines in data centers in the U.S. and globally will grow at compound annual growth rates of 18% and 15% respectively in the coming years, with global new demand reaching about 40GW by 2030. Advantages and Disadvantages: Gas units respond quickly, have lower carbon emissions than diesel, and can utilize existing gas pipelines. Disadvantages: High initial investment and operation/maintenance costs, require stable gas supply.
• Diesel/oil generators: Traditional backup solution, mature technology, low initial investment, suitable for short-term emergencies or independent power supply in remote areas. Disadvantages: Low efficiency (≈30%), high fuel and maintenance costs, noise and emissions, generally used only for short-term supplementation or emergency backup (usually paired with UPS).
• Fuel cells: Technologies like solid oxide fuel cells (SOFC) can generate electricity on-site with fast response (seconds) and extremely low noise. Oracle's latest project plans to use Bloom Energy fuel cells to provide 2.45GW of full power for an AI data center campus. Fuel cells have low carbon emissions (≈500g/kWh when using natural gas) and are expected to replace some gas turbines and diesel applications. Advantages and Disadvantages: High reliability, low maintenance; Disadvantages: High technology cost, limited lifespan, requires advance purchase or self-production of hydrogen for future scalability, short-term mainly natural gas.

Renewable Energy and Energy Storage

• Photovoltaic, wind power: Directly supplying renewable energy to data centers can significantly reduce carbon emissions. Foreign cloud providers have signed long-term green electricity purchase agreements: Google signed a 15-year 1.5TWh solar/wind PPA with Total Energies in France; Meta has signed multiple nuclear power purchase agreements to ensure clean electricity supply. In China, eastern data centers are gradually exploring near-site PV layout, while newly built data centers in western regions with abundant solar and wind resources typically incorporate wind and solar fields. Advantages and Disadvantages: Zero fuel cost, strong policy support; Disadvantages: High volatility, requires large-scale energy storage or flexible scheduling for reliable use.
• Energy storage systems: Mainly lithium-ion battery storage, pumped hydro storage, and hydrogen storage. Lithium batteries (e.g., Tesla Powerpack, CATL products) can provide short-term peak shaving and UPS redundancy for data centers; data center energy storage is expected to grow significantly in the next five years. Pumped hydro projects are geographically limited but are widely used for smoothing renewable generation at the grid level. Hydrogen storage (generating hydrogen and then using fuel cells or gas turbines to generate electricity) has long-duration storage potential as hydrogen costs decline. Advantages and Disadvantages: Battery systems respond quickly and can be located nearby; hydrogen storage enables large-scale long-term storage; Disadvantages: Battery life/decay, requires cooling; pumped hydro/hydrogen technology has large investment and low efficiency (hydrogen ~30% cycle efficiency).

Thermal Management and Energy Efficiency Optimization

• Liquid cooling and immersion cooling: Traditional air cooling consumes significant energy at ultra-high power densities. Liquid cooling (cold plate, immersion) has much higher thermal conductivity than air cooling, significantly reducing PUE. According to statistics, the current penetration rate of liquid cooling in data centers is only 13%, expected to reach 33% by 2030, with a market CAGR of 41% from 2023 to 2028. Liquid cooling can achieve rack PUE < 1.13 and support cooling above 160kW per rack. Domestic and international immersion cooling suppliers (e.g., Submer, 3M, Iceotope) and deployment cases exist, and governments and manufacturers have listed it as an important measure to reduce PUE. Advantages and Disadvantages: High energy efficiency, supports high density; Disadvantages: Requires server modification for compatibility, high infrastructure requirements (liquid medium management).
• Heat recovery and utilization: Recovering waste heat from data center cooling for building heating or industrial heat sources can further improve overall energy efficiency. Some northern cities have piloted using low-grade waste heat from data centers for heating. Advantages and Disadvantages: Saves heat energy, reduces heating fuel consumption; Disadvantages: Limited by geography and pipeline networks, low recovery rate, often implemented alongside high-efficiency room renovations.
• Energy consumption optimization scheduling software: Using AI/algorithms to optimize load scheduling, scheduling non-real-time computing tasks during periods of low grid stress or high renewable energy availability, or intelligently distributing loads across racks. Some IDC operators and tech companies have developed internal energy management platforms for QoS-aware load scheduling to reduce peak demand. Advantages and Disadvantages: Low software investment, flexible adjustment; Disadvantages: Requires hardware compatibility, effectiveness limited by load type and business requirements.
• Migration and edge computing: Regional migration: Moving some computing demand to areas with abundant and cheap electricity, such as the "East Data West Computing" policy encouraging offline computing centers in the west; foreign cloud providers also have cases of building in states like Texas and Indiana with rich renewable resources. Time-based migration: Scheduling deferrable tasks like training to nighttime or low-demand periods to reduce peak loads. Advantages and Disadvantages: Effectively smooths overall load; Disadvantages: Migration requires consideration of network latency and business continuity, scheduling requires sophisticated system support.

(Note: The above table is just an example; each solution category corresponds to other suppliers; deployment cycles and cost ranges vary significantly by project scale.)

Future Potential Solutions and Research Directions

• More efficient cooling technologies: Continuing to advance liquid cooling innovations, such as two-phase flow cooling, microchannel cooling, adaptive cooling based on phase change materials, etc. Research directions include vortex liquid cooling systems, liquid-cooled server design, and new immersion cooling fluids. In the short term (1–3 years), liquid cooling applications will further expand, technology maturity gradually improves; medium to long term (3–7 years), higher temperature difference working fluids and automated control may appear; long term (7–15 years), megawatt-level cooling units and more efficient heat recovery systems may be developed.
• Carbon-neutral electricity procurement: Accelerating the development of green electricity purchase and carbon trading mechanisms, such as companies signing more long-term PPAs, investing in virtual power plants, and purchasing green certificates. Technologically, blockchain and other methods can ensure transparency in green electricity trading. With the improvement of market mechanisms, short-term effects are possible (most actions already underway); medium to long term, a stable carbon-neutral energy supply chain can be formed.
• Hydrogen fuel power generation: Gas turbines or fuel cells using hydrogen as fuel to eliminate fossil carbon emissions. Over the next 10–15 years, with declining green hydrogen costs, hydrogen-based backup and supplementary power sources will be established. Japan, Germany, and other countries have already demonstrated hydrogen fuel cell power supply projects. Technology maturity is low, expected to commercialize gradually in the medium to long term (7–15 years).
• Microgrids and distributed energy systems: Intelligent microgrid integration for data center campuses, including wind-solar-storage DC networking, virtual power plants, etc. It can flexibly schedule power sources, loads, and storage resources to support local autonomy and peak/valley regulation. Technologically, efficient inverters, energy storage management, and microgrid scheduling systems are required. Short term (1–3 years) to promote renewable + energy storage microgrid models, medium to long term to form replicable business models and products.
• Superconducting transmission: High-temperature superconducting cables can significantly reduce transmission losses, addressing long-distance transmission bottlenecks. China has studied superconducting demonstration lines in ultra-high voltage and power transmission/distribution. Due to material and cost constraints, commercial application is still in the outlook stage (medium term 3–7 years requires breakthroughs in material costs; 7–15 years may begin large-scale deployment).
• Energy recovery and thermoelectric conversion: Exploring the use of electromagnetic radiation or temperature differences from servers (thermoelectric materials, thermoacoustic power generation, etc.) to recover more energy from the system level. Currently mainly in the laboratory research stage, with high long-term potential, possibly seeing commercial prototypes in 7–15 years.
• AI adaptive energy management: Using artificial intelligence for self-learning optimization of power and cooling demands, such as on-chip power management, overall thermal flow optimization, real-time predictive scheduling, etc. Google DeepMind has already applied this to data center PUE optimization. With algorithm advancement and 5G/IoT support, short-term gradual deployment is possible, and medium to long term it will become a standard feature in data center operations.

The above technologies are sorted by maturity. Short term (1–3 years) can focus on "energy + computing synergy" (source-grid-load-storage), more efficient liquid cooling, distributed energy storage, algorithm optimization scheduling, etc.; Medium term (3–7 years) focus on hydrogen applications, microgrid commercialization, solid-state energy storage, superconducting technology verification, etc.; Long term (7–15 years) focus on the feasibility and commercial promotion of disruptive technologies (advanced cooling materials, thermoelectric recovery, all-hydrogen grids, etc.).

• Cloud Service Providers/Data Center Operators: Global AWS, Microsoft Azure, Google Cloud, Meta, Oracle, etc., and Chinese Alibaba Cloud, Tencent Cloud, Baidu Cloud, Huawei Cloud, etc. These companies are both huge consumers of computing power and globally deploy data centers and invest in supporting power facilities (e.g., self-built combined heat and power or direct PPAs). For example, Microsoft and Chevron are building a 4GW gas-energy storage plant in Texas; Google signed a long-term PPA with TotalEnergies in France; Chinese operators cooperate with power companies through "computing-electricity synergy" strategies to plan computing layout.
• Power Companies: State Grid, China Southern Power Grid, SPIC, Huaneng Group, Huadian Group, China Three Gorges Corporation and other large Chinese power central enterprises, as well as global equipment suppliers and generation operators like GE Vernova, Siemens Energy, Mitsubishi Heavy Industries. State Grid plans to invest 4 trillion yuan in smart grids and ultra-high voltage, and many companies are accelerating renewable energy generation, pumped hydro storage, and hydrogen power projects. Competitive advantage/risk: Central enterprises can gain policy support and scale advantages, but investment returns are long-term and require regional balance; Western energy companies have mature technology but face Chinese localization competition and international trade friction risks.
• UPS and Generator Manufacturers: Schneider Electric, Emerson (Vertiv), Huawei Digital Power provide UPS and precision power systems; Caterpillar (CAT), Kohler (Kohler-SDMO), Weichai Power, Cummins produce backup fuel/gas generators. Chinese companies like Weichai have entered the fuel cell field. Advantages/Risks: These manufacturers have mature products and stable market share; but relatively high prices, affected by global supply chain and raw material fluctuations.
• Energy Storage and Battery Manufacturers: Tesla (Powerpack/Megapack), CATL, BYD, Guoxuan High-Tech and other lithium battery and complete energy storage system suppliers; Ningbo Shenli, Nandu Power focus on data center UPS and batteries; large-scale renewable energy storage projects have China Three Gorges, China Hydropower, etc. Future new technologies like iron-air batteries, sodium-ion batteries from companies like Energus or Envision Energy are also worth attention. Advantages/Risks: Battery costs continue to decline, can be deployed quickly, strong global demand; but life degradation, thermal management, and supply chain (lithium, cobalt) are major risks.
• Liquid Cooling/Immersion Cooling Manufacturers: Internationally 3M, Submer, Asperitas, GRC, Iceotope offer single-phase and two-phase immersion cooling solutions; cold plate solutions from Nutanix (CORE), Huawei, etc. Domestic Haida Zhileng, JD Digital have deployments in data center liquid cooling. Advantages/Risks: Liquid cooling manufacturers have high technical thresholds, generally focus on high-performance applications, large market growth space; but require cooperation with server manufacturers, educate customers, limited early adoption and operational experience.
• Energy Management and Software: Emerson (GE Digital), Schneider EcoStruxure, Carbon Satellite offer intelligent energy management platforms; OPAL-RT, National Instruments provide simulation and control systems; Alibaba, Tencent and other cloud providers develop their own scheduling systems. Advantages/Risks: Software and AI solutions are flexible and easy to deploy, can iterate quickly; but effectiveness depends on data and algorithm quality, require high expertise and modification of existing networks, standardization level is currently low.
• Startups and Innovators: For example, U.S. Kalray (high-performance AI chips + networking), Chinese Qiyuan Bo, Jingjia Micro (AI chips) are not traditional energy companies but their improvement of computing efficiency indirectly affects energy demand; Deep Blue Electric, Green Ridge (liquid cooling technology) follow trends; Weichai, Envision cross-industry explore new technologies. Assess these companies by technology feasibility, patent barriers, and financing capabilities.

Investment Opportunities and Risk Analysis

Sub-sector opportunities: High-potential areas include efficient cooling equipment, energy storage systems, intelligent microgrids, new power generation equipment (fuel cells/hydrogen), green electricity purchase agreements, etc. The global data center green electricity and energy storage market is expected to be in the tens of billions of dollars, with annual growth rates of tens of percentage points. For example, the liquid cooling market CAGR from 2023–28 is expected at 41%; the gas turbine market CAGR from 2023–30 is 3.6%, with data center demand growing 15% annually; global data center renewable energy supporting investments also show double-digit growth.

Market size estimation: Based on industry reports and calculations. According to IEA predictions, global data center electricity consumption in 2030 ≈ 945 TWh, assuming each kWh corresponds to about $0.5 in energy and related infrastructure spending, the market space over the next decade exceeds hundreds of billions of dollars. For China, the government target of 400 TWh of data center electricity by 2030 corresponds to power and energy-saving renovation needs that account for a significant share of the global market (about one-third). Additionally, related supporting markets such as UPS, batteries, power distribution equipment, cooling equipment, etc., globally total hundreds of billions of dollars.

Growth rate and driving factors: According to different institutions, data center power density, existing and new capacity are in a rapid growth phase (see references). Key drivers include the explosion of AI computing demand, government "carbon neutrality" policies, digital economy growth, etc. Conservatively, data center power demand is expected to grow at a compound rate of over 10% in the next five years, with corresponding equipment market growth also at 10–20%+. Investment entry points: Participation can be through various means—direct investment in related listed companies (e.g., power equipment manufacturers, energy storage companies), bonds (grid and new energy projects), project financing (participation in large-scale energy storage/new energy power stations), M&A or equity investment (green technology startups), industry funds, etc. Hedge funds, green energy funds, and private equity funds focused on AI infrastructure are also options.

Time window and exit: Considering technology and policy evolution, short term (1–3 years) is suitable for sub-sectors with existing business models, such as high-power UPS, liquid cooling equipment, microgrid projects; medium term (3–7 years) focus on technologies still in growth stage but with clear prospects, such as hydrogen fuel cells, iron-air energy storage, intelligent control platforms; long term (7–15 years) requires risk tolerance, such as cutting-edge technologies like new materials, high-temperature superconductors. Exit paths include project revenue, equity transfer, public market exit (IPO), etc.

Policy and technology risks: Potential risks include new government regulations on grids and real estate (e.g., power rationing policies, energy use reviews), subsidy phase-outs, technology substitution (e.g., hydrogen replacing gas turbines), supply chain bottlenecks (chips, battery raw materials), etc. Be cautious of slow progress in electricity market reform, incomplete green electricity trading mechanisms leading to uncertain investment returns. At the technology level, new technologies failing to meet expected performance or high costs also constitute risks.

Recommended List

Based on the above analysis, list 10 key areas/companies in order of investment priority (short/medium/long term) (examples only, not investment advice):

1. Data center liquid cooling equipment manufacturers (e.g., Huawei Digital Power, Submer, Chinese Gaolan Shares, etc.): Short-term benefit from PUE requirements and high-density racks, rapid market penetration, expected stable returns, low technology risk.
2. Energy storage companies (e.g., Tesla, CATL, BYD): Lithium battery energy storage costs continue to decline, can be flexibly deployed in data centers and grids. Large market space over 10 years, good growth potential, but need to guard against raw material price fluctuation risks.
3. Gas turbine and fuel cell companies (e.g., GE Vernova, Mitsubishi, Bloom Energy, Weichai Power): Strong demand for backup/peaking units in the AI era, promising order outlook. Need to monitor gas price and carbon emission policy impacts on costs.
4. Distributed new energy integrators (e.g., JinkoSolar, Goldwind, SPIC, etc.): Encourage wind-solar + storage microgrid models (see Huawei's "Computing Pujiang" solutions), many short-term projects with strong policy support, high replicability in medium to long term.
5. Grid upgrade and intelligent power distribution (State Grid, Southern Grid, Huawei NARI, etc.): As a national strategic focus area, backed by government budget and policy. Long investment cycle but stable fundamentals, relatively certain returns.
6. Microgrid and virtual power plant operators (e.g., State Grid virtual power plant projects, TBEA, etc.): Support data center side collaborative scheduling, can improve green electricity consumption, large potential as market mechanisms mature.
7. New energy and carbon trading service providers: Companies seizing carbon neutrality opportunities in consulting and trading platforms (e.g., Carbon Satellite, Energy Flow Technology), short-term service demand grows steadily but affected by policy changes.
8. Intelligent energy management software companies (e.g., Schneider Electric, Emerson, domestic AIoT companies): Can achieve energy savings and consumption reduction through software upgrades in the short term, asset-light model, suitable for moderate risk appetite investors.
9. Hydrogen technology companies (e.g., Toyota, Mitsubishi Heavy Industries (Hydrogen), CGN Hydrogen, etc.): Focus on medium to long term development prospects, despite few short-term projects, long-term potential is huge, suitable for long-term allocation.
10. Wind-solar PPA platforms and exchanges: As industrialization progresses, more professional PPA platforms and renewable energy trading markets are expected. Participating in low-risk long-term power contract sharing or trading can yield stable returns.

The above targets cover key links in AI computing and energy convergence. Investors should consider their own capital scale and risk appetite, diversify: e.g., short-term focus on equipment manufacturer and operator equity, medium-term layout infrastructure project financing, long-term allocation to emerging technology funds or white-label stocks. Also closely monitor government subsidy policies, technology roadmap maturity, and market demand changes to adjust strategies timely to control risks.