Data Centers Part II: Power Constraints – The Path Forward

The TD Cowen Insight

The accelerating U.S. electricity load growth tied to GenAI-driven data center demand is driving power constraints in major U.S. markets. While our proprietary channel checks and analysis point to growing power constraints, we identified five solutions central to the evolutionary response already underway. As such, we see a path forward for the broader ecosystem to navigate power constraints.

Our Thesis

The accelerating U.S. electricity load growth tied to Generative AI-driven data center demand is driving power constraints, as we estimate third-party U.S. data centers will represent 6.6% of 2028 U.S electricity consumption (versus. 1.5% in 2018) and 2.9% of 2028 global electricity consumption (versus. 0.8% in 2018). Based on our 11-analyst, multinational collaboration, including Washington Research Group, our proprietary channel checks indicate the lead time to procure utility power for new hyperscale data centers across 22 U.S. markets ranges from two and a half to seven years, with three to ten-plus year lead times in the major European FLAP (Frankfurt, London, Amsterdam and Paris) markets. Furthermore, our proprietary electricity runway analysis of future electricity demand relative to reliability-rated supply suggests:

• Northern Virginia will run out of reliability-rated power in winter 2027
• New Albany, Ohio will run out of reliability-rated power in summer 2028
• Silicon Valley will run out of reliability-rated power in summer 2034
• Dallas power demand already exceeds supply relative to levels that would ensure adequate system reliability

In addition, our proprietary, bottom-up global semiconductor shipment forecast aligns with our incremental global third-party data center delivery forecast.

We identified five solutions central to the evolutionary responses already underway including:

1. Building data centers where power is available
2. Repurposing stranded U.S. industrial assets
3. Investing in transmission lines
4. Behind the meter co-generation using natural gas and nuclear
5. Nuclear (Small Modular Reactor) exploration via partnerships

We believe non-latency sensitive workload migration, transmission upgrades and natural gas will play a critical role in servicing incremental data center demand in the near-to-medium term, with nuclear playing an important role in supporting long-term data center capacity growth. Considering the evolutionary response is already underway, and the hyperscalers facing these challenges are the among the best capitalized companies globally, we see a path forward for the broader ecosystem to navigate the growing power constraints.

What Is Proprietary

We conducted extensive proprietary channel checks to determine the lead time to procure power for new hyperscale data centers (+30 megawatts (MW)) across 22 U.S. data center markets. In addition, we engaged a consulting company to conduct a proprietary analysis to identify if and when the demand for electricity exceeds reliability-rated supply in four U.S. markets.

We also developed a proprietary, bottom-up, supply-driven power capacity and consumption estimate based on our view of total AI accelerator processors and Central Processing Units (CPUs), likely to be shipped each year to 2030E, factoring in the replacement of existing CPUs. Furthermore, we discuss in-depth the drivers of data center demand and the data center "clustering effect" (driven by the interconnection flywheel and network architecture), the impact of Electric Vehicles (EVs) on U.S. and global electricity demand, and the changes in semiconductor technology driving growth. In addition, we provide a comprehensive review of state and local, congressional, and federal power policy, and analyze the state of the natural gas and nuclear markets in the U.S.

Financial & Industry Model Implications

Based on our analysis, we estimate third-party data centers will represent 6.6% of 2028 U.S. electricity consumption (versus 1.5% in 2018) and 2.9% of 2028 global electricity consumption (versus 0.8% in 2018), representing a staggering 5.1% and 2.1% increase in third-party data centers share of U.S. and global data center electricity consumption, respectively. Furthermore, supporting our expectation for this data center power consumption growth is our bottom-up, supply-driven power capacity and consumption estimate based on our view of total AI accelerator processors (and CPUs) likely to be shipped each year that indicates a cumulative ~65 gigawatts (GW) of additional power capacity will be required to 2030E, with compute adding ~600 terawatts of global electricity consumption over the next six years. Importantly, our bottom-up semis forecast for 2024-2028 suggests 36.8GW of cumulative net power consumption from chips delivered versus our China-adjusted global data center delivery forecast of 41.0GW.

To meet the demand in the U.S., we believe every incremental 100MW of data center demand will require 19-25 million standard cubic feet per day of natural gas to satisfy power burn, resulting in an incremental 4.5-6.5 billion cubic feet per day (Bcf/D) of natural gas demand by 2028, or roughly 1.3 Bcf/D+ per year, which is not an insignificant ~1.5% growth driver per year of overall natural gas demand (historical reference ranges closer to 0%) but is also not an insurmountable obstacle. In addition, we estimate EVs will represent 2.7% of 2028 global electricity consumption (up from 0.07% in 2018) and 0.58% of 2028 U.S. electricity consumption (up from 0.03% in 2018).

What To Watch

On the data center front, we look for signs of additional large footprint data center builds (particularly hyperscale self-builds) in Tier 3 and Tier 4 markets. In addition, we look for messaging from the hyperscalers surrounding improved utilization of existing data center infrastructure leveraging software (i.e., modalities for power oversubscription in hyperscale self-builds). We also watch for signals of additional adoption of direct behind the meter natural gas co-generation, as well as further partnerships between data centers and hyperscalers and SMR companies to accelerate product development.

On the policy front, we are awaiting the results of the November 2024 election as a Biden re-election, with a Republican-led Senate and Democratic-led House of Representatives offers the highest probability of meaningful reform (environmental and siting). Furthermore, Republican-led efforts (state and federal) are likely to pursue actions designed to ease environmental permitting burdens but constrain transmission development (permitting and cost allocation), while Democratic-led efforts (state and federal) are likely to raise environmental permitting burdens but support transmission build.