As the computing power costs of generative artificial intelligence fall rapidly, there will be a significant mismatch between the rapid growth in AI demand and the slow growth of power infrastructure. Damo compared this mismatch to a “turtle and rabbit race,” suggesting that although AI is currently popular, like rabbits are leading the game, the increase in AI demand depends on electricity supply, and more data centers will pay higher electricity price premiums because they want to secure electricity faster. As a result, the slow-growing electricity infrastructure sector is likely to have better prospects as a turtle.
Morgan Stanley recently released a research report predicting that with the rapid decline in computational power costs of generative artificial intelligence, there will be a significant mismatch between the rapid growth in AI demand and the slow growth of power infrastructure. Damo compared this mismatch to a “turtle and rabbit race,” suggesting that although AI is currently popular, like rabbits are leading the game, the increase in AI demand depends on electricity supply, and more data centers will pay higher electricity price premiums because they want to secure electricity faster. As a result, the slow-growing power infrastructure is likely to have better prospects in the electricity stock sector, and has raised the target prices of many power companies.
Next-generation AI chips drastically reduce computing power costs or stimulate the need for upgrades
According to the research report, the computing power cost of generative artificial intelligence will drop rapidly, but the market situation is underestimated. Daimo's data center model shows that when switching from Nvidia's H100 Hopper GPU to B100's Blackwell GPU, the data center's capital cost of floating-point operations per second (TeraFlop) dropped by about 50%. Damo said that this figure is the ratio of the total capital cost of the data center to the number of floating point operations per second in the data center, and may help determine whether the generative artificial intelligence business model will generate a high return on investment.
The study found that in the data center economy model using Hopper GPUs, this figure was around $14 per TeraFlop, while for the Blackwell data center model, this figure dropped to $7 per TeraFlop. This rapid reduction in computing costs is made possible by the rapid increase in Nvidia GPU power efficiency, so it is expected that data center demand for new technology upgrades will increase.
AI electricity demand may double by 2027
In light of this, Morgan Stanley revised previous estimates and predicted that the global demand for electricity for generative AI could double. Damo has raised the GPU/custom chip utilization rate from 60% to 70%, and it is expected that the percentage of electricity generated by renewable energy in data centers will decrease, and traditional energy will play a greater role.
Damo estimates that under basic circumstances, in 2024 and 2027, global data center electricity demand will be around 430 and 748 terawatt-hours (TWh), respectively, which is equivalent to about 2% and 4% of the world's electricity demand in 2022, respectively. Meanwhile, the compound annual growth rate (CAGR) of generative AI power demand from 2023 to 2027 is expected to be around 105%, while the corresponding annual growth rate for global data center power demand (including generative AI) is about 20% during the same period.
In an optimistic situation (reflecting 90% chip utilization), Damo expects the global data center power demand to be around 446 and 820 TWh in 2024 and 2027, while the pessimistic scenario (reflecting 50% utilization) predicts data center power requirements of about 415 and 677 TWh in 2024 and 2027. Converting these numbers to gigawatts of power capacity (GW), the total power capacity of data centers in 2024 and 2027 will basically be around 70 and 122 GW.
Power infrastructure is growing slowly, data centers are willing to pay more electricity price premiums
At the same time that demand for AI power is surging, the growth of power infrastructure is facing challenges. Damo notes that one of the main concerns is the availability of grid connections needed to support the capacity of the new data center, key issues include limited power line capacity, delays in planning and licensing of new transmission and distribution projects, and supply chain bottlenecks.
For example, according to the Lawrence Berkeley National Laboratory in the US, due to regulatory hurdles, it may take three years or more to upgrade existing transmission lines, and the queuing time for new projects to connect to the grid has increased from less than 2 years in 2008 to 5 years in 2022. Therefore, increasing power generation and grid capacity in smaller secondary markets is an appropriate strategy.
Therefore, given the large amount of capital deployed in these new, very large data centers, and the rapid pace of AI chip innovation, data centers will want to connect to the grid as soon as possible with great value. Therefore, in order to make up for time costs, Damo believes that more data center developers may be willing to pay electricity price premiums.
As an example, the research report said that if a data center developer can secure a power source two years earlier than other data center developers, the developer is willing to pay about 101% of the electricity price premium assuming the economic life of the GPU is six years, or about 61% of the electricity price premium assuming a 10-year economic life span.
Ensuring electricity? The data center is best built within the fence of a nuclear power plant
Therefore, Damo believes that the most suitable one to build a data center is an American nuclear power plant, and points out that in the cooperation plan between Amazon and Talen Energy, it is planned to build a data center with a capacity of about 960 megawatts “within the fence” of a nuclear power plant in Pennsylvania.
According to the research report, the reason why nuclear power plants are most suitable is because the process of supplying power to data centers is faster, and nuclear power plant sites are usually very large, which can provide the space required to build large-scale data centers; at the same time, nuclear power plants already have a large amount of power infrastructure, which can reduce the cost of power infrastructure required for development. Furthermore, nuclear power plants can obtain large amounts of cooling water, which may be an advantage for data centers given that new, hotter data centers may require liquid cooling. Finally, dual-unit nuclear power plants provide redundancy benefits when any single unit has operational issues. Furthermore, the nuclear power industry has an excellent operating record, and unplanned outages are unlikely.
However, Damo also acknowledged in the research report that if the power infrastructure growth bottleneck can be solved, making it easier to develop new data centers, this may in turn reduce the value of existing data centers. However, few analyses show that even with the most promising solution such as nuclear power plants, the increase in power supply capacity is still limited compared to the growth of data centers.
At the same time, if Nvidia and other chip makers achieve rapid increases in chip computing power, existing data centers may need frequent upgrades to maintain competitiveness with newly built data centers. If data center owners are not fully compensated for these frequent modifications, data center owners may find themselves on a “capital expenditure treadmill.” Upgrading older data centers can present practical challenges, particularly when it comes to liquid cooling (water resources may be an issue — but some newer liquid cooling technologies use closed loop systems). If the generative AI business model fails to generate expected profit margins, data center customers may be at risk of default.
Editor/Somer