AI Data Center Power Demand Surges, 47GW Shortfall Expected by 2028 – Latest Analysis by Morgan Stanley
Editor
As the generative artificial intelligence (AI) revolution accelerates, the global data center industry is facing an unprecedented surge in power demand. Morgan Stanley’s latest ‘Powering Gen AI’ analysis report presents this phenomenon with concrete figures, projecting that U.S. data centers will require approximately 72 gigawatts (GW) of power from 2025 to 2028. This is roughly three times the current power consumption of all U.S. data centers, a result of the commercialization of large language models (LLMs) like ChatGPT, Claude, and Gemini, and the competitive race among companies to build AI infrastructure.
Notably, the analysis indicates that the existing power supply system cannot meet this demand. While approximately 10GW can be secured from data centers currently under construction and about 15GW from idle grid capacity, a shortfall of 47GW is still expected. This figure is 3GW higher than Morgan Stanley’s previous forecast of 44GW, suggesting that the pace of AI data center construction is advancing faster than initially anticipated. In fact, major tech companies like Microsoft, Google, and Amazon have allocated over $60 billion for AI infrastructure investment in the third quarter of 2024 alone, with a significant portion focused on data center expansion.
The background to this surge in power demand lies in the increasing complexity of AI models and the expansion of training data scales. Latest models like GPT-4 require more than ten times the computing power compared to previous generations, and high-performance GPU clusters, operating 24/7, are essential for real-time inference services. A single NVIDIA H100 GPU consumes about 700 watts of power, and large AI data centers house tens of thousands of GPUs, exponentially increasing total power consumption. OpenAI’s ChatGPT service alone is estimated to incur monthly power costs of approximately $500,000, which is more than ten times higher than traditional web services.
## Emergence of Rapid Power Supply Solutions
Morgan Stanley analyzed that ‘rapid power supply’ solutions will play a key role in addressing this power supply gap, rather than waiting for traditional grid connection procedures. Traditional grid expansion takes a lengthy 5-7 years from permitting to construction, while AI data centers need to commence operations within 2-3 years. This temporal mismatch is creating demand for new power supply methods, and several innovative solutions are gaining attention.
Gas turbine-based power supply is expected to account for the largest share. Morgan Stanley projected that 15-20GW of power could be secured through gas turbines, due to their relatively quick installation and high reliability. General Electric’s (GE) latest H-class gas turbines achieve over 60% efficiency and can be operational within 18-24 months from installation. In particular, several gas turbine power plants dedicated to data centers are already under construction in Texas and Georgia, and Amazon Web Services (AWS) has signed a contract to build a 300MW dedicated gas turbine power plant in Virginia.
Bloom Energy (BE), based in California, is also emerging as a significant alternative with its fuel cell technology. Morgan Stanley analyzed that Bloom Energy’s fuel cells could provide 5-8GW of power. Bloom Energy’s solid oxide fuel cells (SOFC) convert natural gas directly into electricity, achieving over 60% efficiency, and their modular design offers excellent scalability. In the third quarter of 2024, Bloom Energy recorded a 47% year-over-year increase in revenue to $320 million, with over 70% of orders related to data centers. In particular, they have accelerated their entry into the Asian market by signing a long-term supply contract with Korea’s SK Group.
Utilizing existing power plant sites is also one of the noteworthy solutions. Morgan Stanley projected that this method could secure 5-15GW of power, but noted political risks. While utilizing nuclear power plant sites is considered the most efficient option, the public opinion and regulatory environment surrounding nuclear power in the U.S. are complex. The plan to restart Pennsylvania’s Three Mile Island nuclear plant is a representative case, where Microsoft has signed a 20-year power supply contract, but delays are occurring due to opposition from local residents and the federal regulatory approval process.
## Trend of Converting Bitcoin Mining Facilities to Data Centers
One of the most innovative approaches is the conversion of Bitcoin mining facilities into data centers. Morgan Stanley analyzed that this method could secure 10-15GW of power. Bitcoin mining facilities already have large-scale grid connections, cooling facilities, and power management infrastructure, making them ideal for conversion into AI data centers. Especially after the Bitcoin halving in April 2024, as mining profitability significantly decreases, many mining companies are seeking alternative business models.
Texas-based Riot Platforms and Marathon Digital are leading this conversion. Riot Platforms announced the construction of a 200MW AI computing center at its Rockdale facility in Texas, expecting to reduce construction costs by over 50% by utilizing existing mining infrastructure. Marathon Digital is pursuing similar conversion plans at its Nebraska and North Dakota facilities, aiming to build a cluster of 100,000 NVIDIA H100 GPUs. These conversion projects can bypass the traditional grid connection approval process, enabling operation commencement within 18-24 months.
Similar movements are also appearing in South Korea. Korbit, the largest Bitcoin mining company in the country, announced plans to convert part of its facility in Paju, Gyeonggi Province, into an AI computing center, with major companies like Samsung SDS and LG CNS showing interest in partnerships. Particularly, South Korea’s data center market is growing at an annual rate of over 15%, and power demand is surging due to the expansion of AI services by domestic big tech companies like Naver and Kakao.
However, there are technical challenges in converting Bitcoin mining facilities into data centers. Mining equipment like ASICs and AI training GPUs have different power consumption patterns and cooling requirements, necessitating infrastructure modifications. Additionally, while AI workloads require continuous high performance 24/7, mining operations involve relatively simple and repetitive calculations, requiring the accumulation of operational know-how. Despite these challenges, many companies are actively pursuing conversion due to the advantage of utilizing existing power infrastructure.
According to Morgan Stanley’s probability-weighted analysis, these various rapid power supply solutions are expected to secure an additional 31-50GW of power by 2028. However, this still leaves a shortfall of 6-16GW compared to the anticipated AI data center construction demand, indicating it is not a complete solution. This supply shortage is likely to lead to increased data center rental rates and AI service costs, potentially impacting the growth rate of the AI industry.
## Global Power Infrastructure Investment and Policy Response
This surge in power demand is spreading globally. In China, AI data center power consumption increased by 80% year-over-year in 2024, with major companies like Baidu, Alibaba, and Tencent announcing data center investments totaling 120 billion yuan (approximately $16.5 billion). The European Union (EU) has also established a digital infrastructure investment plan of 100 billion euros by 2030 to secure digital sovereignty, with over 40% allocated to power infrastructure related to AI data centers.
The South Korean government is also pursuing a 3 trillion won data center investment plan by 2027 through the ‘K-Cloud Belt’ initiative. Large-scale data center complexes will be established, focusing on southern Gyeonggi Province and the Chungcheong region, with plans to build a power supply system based on renewable energy. Korea Electric Power Corporation is considering introducing a dedicated power tariff for AI data centers and preparing an incentive system to distribute power usage during peak hours.
Significant changes are also occurring on the investment front. Global asset management firms like BlackRock and Vanguard are classifying data center power infrastructure as a new alternative investment asset class, with approximately $50 billion flowing into the sector in 2024 alone. Particularly, companies related to fuel cells, gas turbines, and energy storage systems (ESS) have seen their stock prices rise significantly alongside the AI boom. Bloom Energy’s stock price has increased by over 300% compared to early 2024, and Westinghouse Electric is gaining attention for its development of small modular reactors (SMRs) utilizing nuclear technology.
However, this rapid growth also carries risks. If power supply shortages persist, a decline in AI service quality and rising costs are inevitable, which could directly impact the survival of AI startups. In fact, some AI startups are already considering service reductions or fee increases due to rising computing costs. Additionally, rapid power supply solutions based on fossil fuels could lead to increased carbon emissions, potentially not aligning with ESG (Environmental, Social, and Governance) investment criteria.
The implications of Morgan Stanley’s analysis are clear. For the continued advancement of the AI revolution, innovative approaches to power infrastructure are essential, as existing methods cannot meet the surging demand. While various rapid power supply solutions like gas turbines, fuel cells, existing power plant utilization, and Bitcoin mining facility conversion can provide short-term solutions, a more fundamental change in the power system is needed in the long term. This transformation process is expected to create new investment opportunities and business models, significantly reshaping the market positions of related companies.
This analysis is provided for informational purposes and is not intended as investment advice or a recommendation for specific stocks. All investment decisions should be made at the discretion and responsibility of the individual.
