There is an arms race between Microsoft, Alphabet, Amazon, OpenAI, Meta, Oracle, and others to be the largest platform for generative AI. These owners of large power-demanding data centers are known as hyperscalers. They are working with, and sometimes trying to work around, major utilities to support AI data centers that demand ten to one hundred times the power of current data centers.
Microsoft and Three Mile Island
Microsoft has signed a $16 billion (over 20 years) agreement to purchase energy from the Constellation Energy Three Mile Island Unit 1 nuclear plant to power AI data centers. The proposed 837 megawatt (MW) would be used in Microsoft AI centers in several states.
Constellation may only spend $1.6 billion and only four years to reopen Three Mile Island, a nuclear plant that is over 50 years old, but Kennedy Maise’s thoughtful analysis outlines a number of challenges. Neither Constellation, nor Microsoft, should expect ratepayers and taxpayers to foot the bill. The biggest issue is what sources of added generation are most cost-effective. Microsoft hopes the new AI data centers will be powered by Three Mile Island in 2028. Speed is of the essence.
The power demand for the next generation of AI goes far beyond the demand of Microsoft’s estimated 200 data centers globally that now serve Azure cloud customers.
An alliance between BlackRock, Microsoft, Global Infrastructure Partners and MGX was announced with a goal of building the backbone of future AI infrastructure. The new Global AI Infrastructure Investment Partnership plans to raise $80 billion to $100 billion to build data centers and the grid energy infrastructure to power them.
Speed and scale
Energy demand for data centers has been growing ten percent annually. AI accelerates this growth. Schneider Electric estimates that AI power consumption will grow at a rate of 25% to 33% annually through 2028, when it could reach as much as 18.7 gigawatts (GW), taking data center power consumption to 93 GW.
Microsoft, Alphabet, Amazon, and Meta have market capitalizations of over $1 trillion each. They see the value of being the platform leader in AI as worth trillions. They want speed and scale in getting massive new data centers built, powered and running.
Ideally these new data centers will be located near water for cooling tens of thousands of semiconductors, will be near high-voltage grid connections for 24/7 power, and on low-cost land. These demands leave developers out of data center space in many major markets. The average vacancy rate across North American data center markets was 2.7% in the third quarter of 2023, according to datacenterHawk. In some markets, like Northern Virginia, the vacancy rate is about one percent.
CBRE, a property management giant, states that data center construction completion timelines have been extended by 24 to 72 months due to utility electricity supply delays. CBRE US Real Estate Outlook 2024
In past years, the hyperscalers have spent billions on wind and solar power. They prefer to power the new AI data centers with renewables and nuclear power. They are also making venture capital investments in SMR, microreactors, nuclear fusion, enhanced geothermal, renewable and storage startups. Tech executives are optimistic about the kind of innovation and volume manufacturing that has resulted in dramatic cost reductions in semiconductors, computers, solar power, and energy storage. Such cost reductions may not occur with nuclear power, geothermal, and CCS.
It is not in the DNA of hyperscalers to wait for the pace of regulated utilities. They want to control their own power. They may be in for major surprises.
Amazon (AWS) and Talen Energy
Amazon Web Services (AWS) operates a vast network of data centers globally, strategically located in major population centers and business hubs across North America, South America, Europe, Asia Pacific, the Middle East, and Africa.
AWS is also investing heavily in clean energy, with a goal of powering its operations with 100% clean energy. In 2024, 100% of energy use is matched with renewables by investing in on-site generation with solar and wind, with power purchase agreements (PPAs), and by investing in new renewable energy projects globally.
Amazon recently bought a $650 million nuclear-powered data center from Talen Energy with a 2.5 GW capacity. In 2021, AWS contracted with Talen Energy to develop renewable energy projects in the United States. One notable project is the Montour Solar Energy Center in Pennsylvania, which will provide 65 MW of solar power to AWS. The new data center campus is in Pennsylvania, adjacent to the Susquehanna Steam Electric Station, a nuclear power plant. The campus has a potential capacity of 960 MW, with the nuclear plant itself boasting a 2.5 GW capacity. AWS has ambitious plans to expand the campus by building over a dozen new data centers on the property over the next decade.
Direct-Connect Power to Data Centers
Back-up power has long been used in data centers, where downtime can cost millions per minute for a data center powering financial transactions on Wall Street or AI for the Fortune 500. Diesel gen sets and batteries have long been popular. In 2020 Microsoft announced plans to eliminate its reliance on diesel fuel by the year 2030. Alternatives are being developed, including biofuels, hydrogen fuel cells and large batteries.
Technology giants have expanded from backup power to direct-connect power. Direct-connect power, also known as behind-the-meter power or onsite generation, refers to the supply of electricity directly to a data center from a dedicated power generation source. Most data centers using direct-connect are also grid-tied with significant onsite generation. Being grid-tied, hyperscalers need the added time of utility and regulatory approvals.
Vistra, NRG Energy, Constellation Energy and Talen have been aggressive in pursuing direct-connect power agreements with leading data center developers. Benefits of Direct-Connect Power:
- Reliability: Direct-connect power minimizes the risk of outages associated with grid disturbances. This is crucial for data centers, where even brief power interruptions can lead to significant data loss and financial consequences.
- Cost Savings can be achieved using owned generation. Additionally, some onsite generation technologies, like combined heat and power (CHP), can utilize waste heat for increased efficiency.
- Sustainability: Direct-connect power can incorporate renewable energy sources, such as solar or wind, reducing the data center's carbon footprint.
Many existing data centers use Direct-Connect Power:
- Apple Maiden, North Carolina Data Center, uses a 100-acre solar farm and Bloom Energy fuel cell using local biogas, and power from Duke Energy, including contracted wind and solar.
- Switch Citadel Campus in Nevada partnered with NV Energy to develop a dedicated solar farm that supplies renewable energy directly to its massive Citadel Campus.
- Google Data Centers in many locations are direct-connected to local wind farms and solar installations, to power its data centers.
- Equinix and Digital Realty are colocation providers that host direct-connect data centers.
While still relatively uncommon, off-grid data centers are gaining traction as a solution for increased reliability, sustainability, and potentially reduced costs. Off-grid data centers:
- EcoDataCenter (Falkenberg, Sweden) is 100% powered by renewable hydro and wind power, also using lithium-ion batteries and hydrogen fuel cells.
- Verne Global (Iceland) uses Iceland's abundant geothermal and hydroelectric resources for 100% renewable energy supply.
OpenAI and 5GW Data Centers
OpenAI is campaigning the US government for regulatory and financial assistance to add 5 GW in multiple regions for its planned AI data centers. If you have used ChatGPT, you have witnessed OpenAI’s rapid progress with generative AI.
President Biden has called for a tripling of U.S. nuclear power capacity to fuel energy demand that is accelerating in part due to expansion of AI and cloud computing.
Goldman Sachs Research estimates that data center power demand will grow 160% by 2030. Goldman Sachs 35-page report. “That kind of spike in power demand hasn’t been seen in the US since the early years of this century. It will be stoked partly by electrification and industrial reshoring, but also by AI,” states Goldman Sachs. The forecast is that 60 percent of new generation to support data centers will be gas; 40 percent, green.
The forecast includes a sharp increase in use of gas peakers, which is surprising given the strong growth of battery energy storage systems (BESS) with peak-shaving and demand response financial incentives, elimination of gensets, and renewables growth.
Expect the Unexpected
Each new release of generative AI uses a larger language model (LLM) that requires larger and more energy-intensive data centers requiring weeks of 24/7 LLM training and then years of 24/7 use by millions for complex inferences. Open AI CEO, Sam Altman, uses the analogy that GPT-2 to GPT-4 was a 10,000-fold improvement. Is there another 10,000-fold improvement by 2030? There is room for many data center energy growth forecasts and scenarios.
Billion dollar investments by Microsoft, AWS, Alphabet and other hyperscalers are being made in new data centers and new energy sources. The forecasted 160% data center energy demand growth by 2030 is creating opportunities for utilities, suppliers, and energy professionals.