Concerns Behind the AI Boom: By 2028, the U.S. Power Shortfall May Be Equivalent to 44 Nuclear Power Plants

Morgan Stanley recently released a report stating that the construction of artificial intelligence (AI) infrastructure in the United States is driving domestic electricity demand into a new phase, and the capacity of electricity supply may become a key limiting factor for the expansion of the AI industry.

The bank's strategist Stephen Byrd pointed out in a research report titled "Powering AI: Bitcoin Conversion, Business Models, a US Power Shortage and the Big Picture" that by 2028, the total electricity demand of data centers in the United States is expected to reach approximately 69 gigawatts (GW). Among them, about 10 GW comes from data centers under construction, another 15 GW can be accessed through the existing grid, but there is still an electricity gap of about 44 GW.

This latest data is an upward revision from Morgan Stanley's forecast last December (a gap of 36 GW). The report noted that, when converted to the output of nuclear power plants, 44 GW is equivalent to the scale of about 44 nuclear power stations.

The report mentioned that the Loan Programs Office under the U.S. Department of Energy recently stated that it is prepared to provide hundreds of billions of dollars in financing for nuclear power projects to promote the construction of clean energy capacity and alleviate potential electricity supply pressures.

Morgan Stanley believes that the electricity supply shortage may affect the implementation and pace of AI-related investments. According to estimates, the construction cost for each additional 1 GW of data center capacity is about $50 billion to $60 billion, and insufficient electricity access capacity may lead to an extended construction cycle for AI infrastructure.

Morgan Stanley emphasized that there are currently no new nuclear reactors under construction in the United States. Considering that the construction cycle for nuclear power typically takes more than ten years, if the U.S. does not enhance its power supply capacity in the short term through natural gas, fuel cells, and the renovation of existing facilities, it may not be able to support the rapid expansion of AI infrastructure demand.

Time to Power Solutions

To address this issue, Morgan Stanley proposed several "Time to Power" solutions, which are alternative measures that do not rely on traditional grid connection processes and can achieve power supply more quickly. If all these solutions are implemented, the electricity gap in the U.S. by 2028 could be reduced to about 20%, equivalent to 13 GW, still about the output of 13 nuclear power stations.

The report listed several potential solutions:

• Natural gas turbine projects could add about 15 to 20 GW of electricity;

• Fuel cell company Bloom Energy could contribute 5 to 8 GW (if its annual production capacity is increased to 3 GW, potential supply could be further expanded);

• The transaction of existing nuclear power plants providing direct electricity to data centers can bring about 5 to 15 gigawatts (excluding the indirect method of offsetting nuclear power usage with newly added natural gas generation, which has been included in the aforementioned 20 gigawatt gas turbine project);

• In addition, Morgan Stanley estimates that existing Bitcoin mining sites have large (over 100 megawatts) facilities with complete access agreements, totaling approximately 20 gigawatts of potential capacity, which can be converted into 10 to 15 gigawatts of actual supply.

Among these options, Morgan Stanley believes that the conversion of Bitcoin mining sites into AI data centers has significant advantages in execution speed and risk control, and may gain higher market recognition in the future. The report also points out that Bloom Energy's fuel cell system is a reliable "quick power supply" solution, expected to drive rapid growth in the company's shipments.

In addition to fuel cells and the conversion of Bitcoin mining sites, Morgan Stanley expects a diversified "Time to Power" trading model to emerge, involving multiple participants such as independent power producers, turbine manufacturers, and energy companies.

Attention on Bitcoin Mining Site Transformation to Data Centers

Against the backdrop of rapidly growing demand for AI computing power, Morgan Stanley is particularly focused on the trend of Bitcoin mining sites transforming into high-performance computing (HPC) data centers. The report notes that there are currently two main business models in the industry: one is the "New Neocloud" model, represented by IREN, where mining companies purchase GPUs or TPUs, build their own data centers, and then lease computing power facilities to hyperscale cloud service providers or enterprise customers for a short term. For example, IREN has signed a five-year lease with Microsoft.

The second is the "REIT Endgame" model, where mining companies are responsible for building the "energized shell" (i.e., infrastructure excluding chips and servers) and sign long-term leases with cloud computing companies. For example, APLD has signed a 15-year lease with an undisclosed cloud service provider.

Morgan Stanley believes that both models have considerable value creation potential and demonstrate the path for traditional cryptocurrency infrastructure to transition into the AI computing field.

The report also provides valuation reference data for the transformation of Bitcoin mining sites into data centers, showing that large mining sites with stable grid access and installed capacity exceeding 100 megawatts have significant differences in enterprise value/watt (EV/W) multiples. Morgan Stanley points out that the lower the valuation multiple, the more attractive the potential conversion opportunity.