In Sizing Up Future AI Power Requirements, It PaysTo Watch For Zombies

A review of announced data center electricity supply requests over the past year suggests that over 125,000 MW of new capacity has been requested of utilities to meet data center demand. Most of this demand is AI-related (though some is of atypical data center operations, a small amount may be serving future crypto loads).

But as everybody (myself included) hyperventilates a bit about the potential impacts to our power grid, it’s important to step back and recognize where we are just now, which is probably in the midst of yet another classic hype cycle. We’ve seen these before, with the biggest being the Dot.com era in the early 90s (raise your hand if you were excited as I was about being able to buy dog food online from a sock puppet. BTW, he’s now residing at the Henry Ford Museum).

Yes, 25% of Virginia’s load already serves data centers. And yes, AI will have many uses, some of which we can only imagine today (many of which will make the grid far more efficient – a post for the near future) and it will need more electricity to do so. However, there may be some real limits as to how much power these AI-driven data centers will consume, which result in the actual supply number coming out well below some of the projections we hear today. Among other developments, there will likely be substantial gains in chips and cooling efficiencies (liquid cooling is making great strides) that conspire to drive energy requirements down from what they otherwise would have been. Some of this tech is so new, it's hard to know how it will eventually pan out.

Ask Jeeves

Right now, multiple AI companies are racing to accomplish pretty much the same thing: market dominance. There will undoubtedly be companies that fail to make it in this highly competitive global race that will eventually be dominated by a few deep-pocketed participants, with some of today's players n going bankrupt or being absorbed into larger competitors. This may be analogous to the race for search engine supremacy in the 90’s. Ultimately, we were left with only one or two – with very deep pockets - that mattered as the rest failed. In the end, the majority of AI companies will likely be consolidated, and if you don’t believe that statement, you can go Ask Jeeves.

The Process of Connecting

However, the point of this exercise here is to focus specifically on the data center connection capacity figures being announced by utilities across the country. There is a significant chance that these headline numbers vastly inflate what will actually be built, because of the way the process for connecting with the utility actually works.

To better understand this issue, let’s look at Dominion Energy’s five-step interconnection process for large loads. Since Dominion has lived in this world for a while (northern Virginia is the world’s leading data center hub) they’re further ahead on this than most utilities. These steps include:

1) The utility performs a high-level assessment of the applicant customer’s power needs.

2) Customers opting to proceed then sign an Engineering Letter of Authorization, committing to pay the utility for detailed studies analyzing optimal locations for substations and the required infrastructure that may need to be developed.

3) Customers wishing to continue then sign a Construction Letter of Authorization obligating the applicant to pay for all project-related expenditures regardless of whether the project eventually breaks ground.

4) Dominion then begins to develop the requisite infrastructure.

5) The entire undertaking is governed by an Electric Service Agreement (ESA) that sets the terms for how the customer will receive energy and the payments it will make in exchange for that service (the ESA is structured to ensure that the customer covers the utility’s whether or not it receives power at the specified levels).

AEP Ohio, which is attracting a great deal of load largely owing to its super-sized 765 kilovolt transmission lines, is also working hard to get out in front of this game. Last May, the utility served 600 MW of data centers, and at that time had an additional 4,400 MW of demand covered by ESAs or Letters of Agreement. As of that date, an additional 30,000 MW of potential new loads had also come knocking at AEP’s door (a figure likely to be even larger today).

In AEP Ohio’s case, the utility has stated that it simply cannot serve that much additional load without undertaking enormous new investments in transmission equipment. As a result, it has proposed a new tariff structure that would require data centers larger than 25 MW to pay for a minimum of 85% of their electricity requirements for up to 12 years, even if they end up using less energy than projected. Data companies must also demonstrate financial viability and fork out exit fees if projects are canceled or downsized. The utility’s goal here is to minimize the potential for orphaned infrastructure that other customers would subsequently have to pay for (in the industry jargon, these are called “stranded assets”).

Watch Out For Zombies

The reality is that only a fraction of the initial interconnection requests will get built. Although it’s impossible to gain insight as to what exactly is happening at any point in time (in this highly competitive global arena, these companies play their cards very close to their chests), many of the data companies are likely doing what you are I would do if we had a mandate to get as much electricity as we could as quickly as possible: We’d submit multiple applications to numerous utilities, with the hope that at least some of the applications would strike pay-dirt. Then, as soon as those projects looked like they had a chance of bearing fruit, we’d take the other chips off the table and scrap the “zombie requests” (a useful and colorful phrase used by Brian Janous – former VP of energy at Microsoft and co-founder of Cloverleaf Infrastructure).

Looking to Supply Interconnection Queues For Analogies

There’s an analogue on the supply side of the utility industry that may be instructive in this regard. Today in t4he U.S. there are over ten thousand generation projects in transmission interconnection queues, waiting to be connected to the electric grid before they can be commissioned and deliver power. In fact, the nameplate capacity of supply assets languishing in queues is roughly twice as large as that of the existing generation fleet. However, recent history shows that fewer than 20% of supply-side projects in the queue actually get built.

As utilities further tighten up their interconnection requirements on the demand side, implementing more formal procedures and instituting tariffs that require significant up-front financial commitments from the data center crowd, we should expect to see zombie applications decrease.

It’s clear that AI has real value to society, and we are beginning to see some use cases emerge. It’s also clear that we are still in the very early days of this dynamic, with rapidly evolving technologies and business models, and many unanswered questions. However, getting past the current hype cycle will take some time. We won’t know the full implications until we start to see some projects proceed, while others are canceled. If you don’t believe that statement, you can go ask Perplexity.AI. In response to the query, “How many of the AI data center utility interconnection requests will actually be built?” it responded, “several factors suggest that only a fraction of the proposed projects will likely be completed.” Taken right from the horse’s mouth...