Power Grab: Data Centers and the Grid Part 4 - AI & Wholesale Power Markets

What if the explosive demand for data centers could transform power grids across the nation? Join us as we unravel the latest developments in data center supply strategies and the energy market's evolving dynamics. From Vista’s power sales discussions with major companies to California's innovative district energy system marrying data centers with residential units, this episode is packed with groundbreaking insights. We explore the potential of modular nuclear startup Oklo’s partnerships and the intriguing collaboration between Sharon AI and New Era Helium Corporation for a new data center venture in New Mexico. Plus, Bloom Energy’s significant deal with AAP Ohio takes center stage as we discuss its implications.

As we navigate the intricate web of these developments, we turn our attention to their broader impact on electricity users and wholesale markets. The rapid growth in data center demand is reshaping the power grid landscape, particularly in the Mid-Atlantic and PJM regions. We delve into the potential consequences for bulk power grids, especially with hints that advancements in large language models may be slowing. Could this mean an overinvestment in capacity infrastructure? We tackle these provocative questions and more, offering a critical perspective on the challenges and opportunities in the ever-evolving energy and data center sectors. Tune in to stay ahead of the trends and anticipate the changes on the horizon.

Transcript

Speaker 1: 0:00

Last week we talked about data center supply strategies. Well, a few updates came to my attention in the ensuing week since we've spoken last. The space isn't slowing down anytime soon, it seems. First, vista announced it's in discussions with two large companies concerning sale of power from existing gas and nuclear facilities, with plants in ERCOT and PJM eliciting interest. One option may be co-location deals at multiple sites that might involve construction of new assets as well. Second, in a unique project, california-specific gas and electric and real estate developer West Bank planned to develop a district energy system combining three new data centers using up to 200 megawatts and as many as 4,000 residential units in San Jose. Excess heat from the data centers will yield heating and hot water to neighboring buildings. Further details have not yet been shared. Third, modular nuke startup Oklo says it has letters of intent to partner with two major data center providers and deliver up to 750 megawatts of power. Fourth, sharon AI and industrial gas company New Era Helium Corporation have announced a non-binding letter of intent to form a JV for the design, development and operation of a 90 megawatt data center at a site in New Mexico's Permian Basin. New Era already produces helium natural gas and dry natural gas and plans on supplying some of that gas for an on-site generator that would support the plant.

Speaker 1: 1:26

And fifth, last week we looked at the option of fuel cells and I commented in the face of tens of gigawatts being sought, these quantities are akin to a pimple on a fruit fly in terms of significance to anybody, but Bloom Energy. Well, bloom did score big in the intervening week, announcing a deal to sell up to one gigawatt of fuel cells to AAP Ohio. Still small in the scheme of things, but pretty significant for Bloom. Okay, back to business. We've looked at the deluge of data center interconnection requests totaling well over 100,000 megawatts and rising. We've looked at why this is happening, the AI arms race, and we've touched on possible supply strategies.

Speaker 1: 2:06

Now let's discuss the potential implications for electricity users. In this session we'll talk wholesale markets and next week we'll cover distribution utilities. But first one other quick update from last week. I mentioned that gas pipeline energy transfer was having discussions with data centers representing 3 billion cubic feet of usage per day. Well, it's just released November 12th numbers indicate it now has requests to connect to over 40 data centers in 10 states, total gas consumption Up to 10 BCF per day. To put that in perspective, last year's national usage was 89 BCF per day, and one more contextual number 2023 US LNG exports total 11.4 BCF per day. So this is big stuff just from one gas pipeline company. And one more thing just to cast some high-level doubt into the equation Very recent comments and indicators from the industry hint that just maybe the gains from these large language learning models are starting to slow down.

Speaker 1: 3:08

Indeed, one trade journal reported last week that GPT improvements are slowing down. This may be a result of a shortage of data for models, but the implications could be huge, because they imply that simply throwing stronger chips and more energy at the problem may not be enough to scale these models commensurately. We may end up building lots of capacity infrastructure with limited economic returns on the margin. The terms-granted assets should come to mind, but nobody knows just yet how the technology is going to play out. But let's ignore that for now and assume the tequila bottles will keep coming and the AI language learning model training party is going to go on full blast. If that's the case, what would be the likely impact to our bulk power grids and wholesale markets? The answer to that is nuanced, in that some markets will be much more effective than others the northeastern power markets, iso, new England, new York ISO they likely won't see that much of an effect. Land is limited there and the environmental regs are fairly tight, making it harder to site either data centers or new power plants to supply them.

Speaker 1: 4:14

The mid-Atlantic states and PGM are another story. In January PGM had tripled its forecasted growth numbers from last year. That was well prior to a lot of new utility announcements, so the next update in two months will likely get a lot higher. We all know of the pressure in Dominion Service Territory, the global data epicenter, but that load seems to be metastasizing to other regions. New Jersey's PSEG, for example, just filed an updated forecast to PGM in late October, projecting data center load to grow from today's 343 megawatts to 1196 by 2030. And Exelon, in late October, requested of PGM a large load forecast adjustment of 2600 megawatts through 2029. The mid-continental ISO, miso, is not seeing quite the same pressure, but its CEO noted this month that 2,500 megawatts of new data central load showed up and over 4,600 megawatts of data central load is noted in MISO's July 30th existing large load and new load additions update. So there's some pressure there too, but not compared to Texas, which sits in a league of its own, with a potential near doubling of existing demand.

Speaker 1: 5:27

Encore alone, as we noted previously, reports 59,000 megawatts of requests. In ERCOT's system. That peaked at 85,000 megawatts this past summer. And then what about SPP, the Southwest Power Pool in California? Well, spp is not a big data center market yet, but it is growing. Politicians in North Dakota, for example, noted that they may see massive growth from two undisclosed data center companies, with growth starting at 500 megawatts, going to a thousand megawatts and perhaps growing 10x from there, involving up to $250 billion of capital.

Speaker 1: 6:00

For its part, and despite Silicon Valley being nestled there, california's power prices may be just too high and land too expensive to attract much AI load. So, from a heat map perspective, the pressure is most likely going to be centered on competitive markets in Texas and PJM. But power prices will be affected everywhere for one principal reason All this new load and the supply to meet it will be facing a shortfall of equipment such as transformers and switchgear and the raw materials that go into them Copper, specialized grain-oriented electric steel for the transformers, those sorts of things. Add to that likely future tariffs on imports from China, from which we source billions of dollars of transformers, and demand has the great potential to outstrip supply. There's already a three-year wait or longer for some types of transformers. Add it all together and you have a general recipe for an inflationary energy environment across the entire country's grid, irrespective of where those data centers are going to be located. Then pile on top of that the fact that interconnection queues are stodgy and you have a very inelastic supply function. In other words, it doesn't matter what the buyer is willing to pay, the supplier is simply unavailable. Lawrence Berkeley Laboratory's latest interconnection report indicates that the median duration from transmission interconnection requests to delivery at first megawatt hour is now at five years.

Speaker 1: 7:24

So where are prices headed? Demand is exploding and supply is scarce and more expensive. Power prices are going up more in some markets than others. Consulting group Bain Company projects average annual cost increases of 1% annually. My conjecture is that these increases will be concentrated more in some areas of rapid growth than others where demand is more tepid. Some competitive power grids will see more marked price increases than others simply because of the ratcheting up of new demand, but nowhere will this dynamic be deflationary. In the next and final session we'll talk about the distribution utilities, the enormous uncertainties facing them in this build-out and the inherent risk of overbuilding in an extremely uncertain AI world. Well, thanks for watching and we'll see you again next week.