Power Grab: AI and the Grid Part 3 – Supply Strategies
Previously, we examined projections for datacenter load growth and issues related to chips, power draw, availability of data, and future growth prospects.
Now, we’ll look at potential electricity supply options.
How is power used? Training and inference are estimated to consume about 70-80% of power. In training, most energy is used in initial run model – taking weeks or longer (20-30% of power use is for cooling).
The power used for “inference” – outputs or decisions from trained models – is growing rapidly. As one example, queries to Chat, Perplexity, or other platforms uses 10x the energy for a Google search. Training energy use is initially much larger, but the latter grows over time
So where to get power? Globally, there’s a preference for U.S. w/large grid, available space, stable economy, rule of law, and access to comms cables. Europe’s grid is old, and space is limited, so some datacenters will move to anywhere there’s power. China is its own case and growing rapidly.
AI is increasingly a national security issue – since ChatGPT in 2022, U.S. DOD has awarded $670 mn to over 300 companies for AI projects.
In the U.S., supply strategies are:
1) Grid.
2) Existing assets, like nuclear. Amazon Web Services accessed 300 MW of Talen’s nuke before FERC rejected plans for add’l 660 MW. Constellation and Microsoft plan on 20-year 835 MW deal to re-start TMI Unit 1. Nextera eyes restarting a 600 MW nuke in Iowa. But only so many existing and recently closed nukes exist.
3) Fuel cells. In 2024 Bloom Energy expanded a 6.5 MW agreement to an Intel datacenter and inked a 15-year 20MW deal w/AWS. But amounts will be small in the big picture.
4) Advanced/enhanced geothermal. This may help: Google has 3.5 MW deal w/Fervo Energy in NV and a 2nd deal for 115 MW of energy from Fervo through utility NV Energy. Sage Geosystems also has 150 MW deal w/Meta These first projects won’t come online for a few years, and geothermal isn’t going to see 10s of GWs anytime soon.
5) Modular nuclear? Small reactor company Oklo inked a 500 MW deal w/colo co Equinix. Google signed w/start-up Kairos Power for 500 MW from 2030 – 2035. AWS announced investment in reactor company X Energy, and agreement for 320 MW w/option to expand to 960 MW. Delivery date: early mid-2030s.
But w/exception of NuScale, no players even have design approval from the Nuclear Regulatory Commission.
A logical outcome is to bypass the grid entirely and go right to gas pipelines. Numerous gas cos report discussions w/datacenter operators for gas hook-ups to support BTM generation. Energy Transfer alone is in discussion w/datacenters for new demand over 3 Bcf/day.
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.
