In December 2024, the State Corporation Commission of Virginia held a conference on how the state might manage years (or decades) of data center expansion. Virginia already hosts more than 300 data centers, which together consume about as much electricity as Massachusetts, or Iowa, or Colorado. Data center developers and the world’s largest tech companies are planning potentially gigawatt-scale individual projects, and Virginia’s utilities have a pressing need to model data center power demand, understand where and how it might emerge, and how best to accommodate it technically and economically.
Part of that process is financial and power modeling, but an equally important part of that process is technical discussion. Thus, last month’s conference in Richmond, and with it a cache of documents which Halcyon has ingested and can now unpack, including a 273-page transcript of a seven-hour series of presentations and discussions.
This process is daunting, and even experts may not quickly grok the key information, which is how confident these state, electric power, and data center experts are in their view of the future. So, we had machines do the heavy lifting for us, with a clear purpose in mind: assess the individual and collective confidence in data center demand predictions among a group of deeply experienced subject-matter experts.
The process goes like this:
Then, we use Halcyon’s platform to determine how confident a speaker is about his or her assessment of Virginia’s data center growth. Here’s the prompt we used:
You are an expert economist. Your job is to evaluate how confident a forecast/projection is, based on spoken testimony. You will be given text from a speaker's transcript that may or may not contain a forecast/projection. If it does not, simply return:
{
"forecast": false,
"confidence": null,
"description": null,
"quotes": null
}
If it does contain a forecast/projection, then describe it in the "description" field and assign a confidence score from these choices: [null, "Vague", "Somewhat Confident", "Moderately Confident", "Very Confident"] where "Vague" is the least confident and "Very Confident" is the most confident. Provide at least one direct quote and a maximum of 5 quotes.
Here is the result:
The 273-page transcript features 18 speakers, two of whom were commissioners who did not express a view on Virginia’s future data center demand, and another two of whom did not provide enough information to rank their confidence. We therefore have 14 positions to rank:
Only one of the experts had what we would consider a very confident position on future data center demand in Virginia, and the modal expert was only somewhat confident. In other words, the future is uncertain to an uncertain degree.
We also pulled quotes directly from the testimony. A few are worth highlighting. The first is from Stan Blackwell, the director of Dominion Energy’s data center practice (this ranks as ‘very confident’):
"When you sign and you make a commitment, we’ll build the infrastructure to serve you, and it has requirements in there.
And so I will say again of this process…From July of ‘23 to July of ‘24, we signed over 6,000 megawatts of firm contracts in one year, again to support the forecast that we’re putting out."
Equally as interesting was a response that we ranked as ‘vague,’ from consultant and economist James Wilson: “Sitting here right now, I think our forecast of data center load for, say, 2032 is probably off by a factor of two. I just don’t know which way, two times too high or two times too low.”
That quote strikes us as both rhetorically elegant and strategically significant. The future is uncertain, but even experts are not confident in exactly how it will turn out. Today’s firm contracts give us reasonable clarity on the next five years, but beyond that, other factors will determine the shape of growth. Those factors will be some of today’s already-pressing needs such as land, water, transmission access. But they will also include technology, and, in particular, potential breakthroughs in data center processing capability and AI model efficiency.
Uncertainty makes thoughtful questions about the future quite important. It also makes the information necessary to answer those questions even more so. And when those questions — and their answers — expand to daunting levels, we can ask questions of questions, using data science and technology to make things a bit easier.
Comments or questions? We’d love to hear from you - sayhi@halcyon.eco, or find us on LinkedIn and Twitter