Updates since last week:
Vistra in discussions w/two large companies re: power from existing gas/nukes, w/focus on ERCOT and PJM.
PG&E & real estate developer Westbank are planning on district energy system combining three new data centers – using up to 200 MW – and 4,000 residential units in San Jose.
Modular nuke start-up Oklo has LOIs w/two major data center providers for up to 750 MW.
Sharon AI and industrial gas co New Era Helium Corp announce LOI for JV for to design, develop, and operate a 90MW data center in New Mexico’s Permian Basin.
Bloom announces deal to sell up to 1 GW of fuel cells to AEP Ohio.
Gas pipeline company Energy Transfer increases estimate of gas to datacenters from 3 bcf/day to 10 bcf/day.
An emerging question about future demand: Recent comments/indicators hint that gains from language learning models may be slowing. If true, it implies throwing stronger chips and more energy at LLMs may not be enough.
—–
But if we assume AI LLM training efforts continue at current pace, what are likely impacts to wholesale markets?
Some markets will be disproportionately affected. ISO-NE and NYISO likely won’t see much impact – it’s too hard to site many datacenters or related new power plants.
PJM is another story. In Jan, PJM tripled its forecasted growth numbers from 2023. That was before many new announcements. Dominion’s service territory is crowded, so new load is going elsewhere: New Jersey’s PSEG filed updated forecast to PJM in late October, w/datacenter load growing from 343 MW to 1,196 by 2030. Exelon in late October requested of PJM a Large Load forecast adjustment of 2,600 MW for 2029.
MISO is not seeing quite the same pressure, but its CEO noted this month an increase of 2.5 GW of new datacenter load, and MISO’s July 30th Existing Large Load and New Load Additions Update includes over 4,600 MW of data center load.
Texas (ERCOT) is its own case. Oncor by itself reports 59,000 MW of requests.
SPP and California? SPP is not a big datacenter market – yet. But it’s growing. North Dakota, may see growth from just two companies, starting at 500 to 1,000 MW and growing 10x from there, w/up to $250 bn in investment. Meanwhile, California’s power prices may be just too high, w/land too expensive to attract much AI load.
So, pressure will most likely be centered on ERCOT & PJM, but power prices will be affected everywhere because of equipment limitations, esp. transformers and switchgear, and related raw materials. Then add interconnection queue delays, and supply is simply unavailable.
Consulting group Bain & Company sees avg annual cost increases of 1%. These increases will probably be more concentrated in areas of rapid growth, but nowhere will this dynamic be deflationary.
In the next session, we’ll look at distribution utilities, the enormous uncertainties facing them, and the inherent risk of overbuilding in an extremely uncertain AI world.
