Power Grab: Data Centers and the Grid Part 3 - Supply Strategies
"Previously, we examined the enormous projections for datacenter growth and the energy required to serve that potential demand. We then discussed issues related to chips, power draw, availability of data, and other issues affecting future growth prospects.
Today, we’ll look at potential supply options for datacenters.
Of all factors limiting datacenter growth, power is the big one. Today, most of that energy is used for training – the computers crunching through data to make connections and create intelligence.
Processing and inference are estimated to consume about 70-80% of the power used. In training, most of all the energy is used in the initial run model – which can take weeks or longer. 20-30% of that power use is dedicated to cooling.
The amount of energy used for “inference” - the outputs or decisions from the trained models – is growing rapidly. Inference usage adds up. Questiond to Chat or Perplexity or a host of other platforms burn through about 10x the energy in standard Google search. Training is initially much larger than inference in terms of electricity use, but the latter grows over time.
So where to get the power? Globally, there’s a preference for the U.S. with its large grid, stable economy, rule of law, and access to communications cables. Europe’s grid is old, and space is at a premium, so some datacenters will locate in Malysia, Singapore, Brazil, anywhere they can get power. China is its own case and growing rapidly.
AI increasingly becomes a national security issue - since the release of ChatGPT two years ago, the U.S. Department of Defense has awarded $670 mn to over 300 companies working on AI-related projects.
In the U.S., supply strategies are:
1) The grid.
2) Existing assets, such as nuclear. Amazon Web Services accessed 300 MW of Talen’s nuclear plant before the FERC rejected plans for an additional 660 MW. Constellation and Microsoft are planning on a 20-year 835 MW agreement to resuscitate Three Mile Island Unit 1. Nextera recently is in talks with the feds and conducting engineering assessments as they eye restarting a 600 MW nuke in Iowa. But only so many existing and recently closed nukes to go exist.
3) Some may opt for fuel cells. In 2024 Bloom Energy expanded a 6.5 MW supply agreement to an Intel datacenter and announced another 15-year 20MW deal with Amazon Web Services. But total quantities will be relatively small in the big picture.
4) Advanced/enhanced geothermal will help: Google has a 3.5 MW deal with developer Fervo Energy in Nevada and committed to buying 115 MW of energy from Fervo under a contract with utility NV Energy. Sage Geosystems announced a 150 MW agreement with MetaThese first large projects won’t come online for a few more years, and the industry isn’t going to see 10s of GWs anytime soon.
5) Modular nuclear? Small reactor company Oklo inked a 500 MW deal with colocation company Equinix. Google signed a deal with start-up Kairos Power for 500 MW between 2030 and 2035. Amazon Web Services announced an investment in modular reactor company X Energy, and an agreement for 320 MW with an option to expand to 960 MW. Delivery date: early mid-2030s.
W/exception of NuScale, no players even have design approval from the Nuclear Regulatory Commission.
The logical outcome is to bypass the grid entirely and go right to the gas pipeline. Numerous gas pipeline companies report discussions with datacenter operators for direct gas hook-ups to support behind-the-meter generation. Energy Transfer is in discussions with datacenters for new demand in excess of 3 Bcf/day.
The picture is still murky since it’s early days, but all things equal, new AI load whether directly supplied or grid-dependent will likely raise prices for everybody. In our next session, we’ll discuss why.