This week, a quick look at how Halcyon uses AI to quickly arrive at ground truths in unstructured data.

In October 2024, the Virginia Corporation Commission initiated proceeding PUR-2024-00144. For industry novices, the state helpfully titled it “Ex Parte: Electric Utilities and Data Center Load Growth.” Despite its vintage, the proceeding remains active, and as of early this month, comments and filings are still ticking in on behalf of utilities, customers, service providers and (of course) data center operators.

At Halcyon, we revisit this proceeding frequently, both in the process of populating our Data Subscriptions (including our Large Load Tariff Tracker and Battery Energy Storage Systems Tracker) as well as a test for our Agentic Alerts and our summarization capabilities. Reviewing comment submissions is also a useful way to arrive at new ground truths about data center operations in the world’s biggest data center cluster: the so-called ‘Data Center Alley’ in Dominion’s service territory in Northern Virginia.

First, some of those queries. This one summarizes the comments from Sparkfund (developers that integrate DERs into utility systems), Mainspring Energy (developers of linear generators), and law firm Christian & Barton. Here is what I found useful: the two companies’ specific recommendations to manage data center load growth:

Sparkfund recommendations:

• Promote the Distributed Capacity Procurement (DCP) model: utilities identify distribution needs, then plan, site, own, and operate 1-5 MW BESS sited where capacity is needed, managed as a grid resource to benefit all customers
• Scale deployment (example: 100+ MW/year)
• Treat aggregated DCP resources like traditional power plants with full utility visibility/control

Mainspring recommends that the State Corporation Commission:

• Direct jurisdictional utilities to file non‑firm/interruptible large‑load tariff proposals (pilot + permanent)
• Establish utility reporting requirements to monitor tariff impacts
• Study cost, reliability, and emissions outcomes
• Convene stakeholders to align retail tariff design with PJM’s emerging “connect and manage” framework
• Promote fuel‑flexible, modular on‑site generation for resilience and emissions reductions

Here is another query that summarizes three separate filings from Dominion, Amazon, and Microsoft. This one bullet point alone tells us why Dominion is so important:

• Dominion serves a very large data center market: >50 data center companies, ~450 meter points, ~10 GW connected today and expects >20 GW additional by 2031 — experience grounds its views on practical flexibility limits

Dominion’s filing includes this as well:

The Company has not observed material volumes of Al training load in its service territory, and the data center industry’s historical load factor, approximately 91 percent over the past 12 years, further supports this conclusion. Because inference load provides continuous, real-time services (e.g., navigation, cloud computing, financial transactions), it does not lend itself to mandatory curtailment without significant downstream impacts to customers and society more broadly.

That’s a key finding! Data Center Alley is huge, but to date it is mostly a place for inference, not training. (‘Inference’ refers to running models to deliver results, whereas ‘training’ refers to developing and refining new models; training has historically attracted the most attention when it comes to AI's energy footprint, but as AI becomes more embedded in everyday life, inference is poised to account for a growing share of that demand.) That means that its energy flexibility potential is downstream of the criticality of the services it provides. To return to Dominion’s comments, this enormous cluster of data processing and storage is effectively always-on, and because it provides “continuous, real-time services…it does not lend itself to mandatory curtailment [emphasis added].”

In its filing, Amazon elaborates its own position that not every data center load is the same, nor is every data center operator. In some rather lawyerly language, it supports “the adoption of a suite of policies that accelerate the integration and energization of large load customers to the grid”. It adds that it supports:

The development of load flexibility programs through a tariff or suite of tariffs that are voluntary, aimed at solving a specificized [sic] need, and are technically and commercially feasible for both the utility and large load customer, including but not limited to data centers.

Microsoft, for its part, echoes Dominion’s position: reliability is its paramount concern in Northern Virginia, given its data centers enable “critical services like the 911 operations, hospital programming, and banking services provided by our customers.”

So what can we infer (pun intended) from these latest comments? Everyone wants flexibility, but the biggest data center operators in Virginia make a strong case that uptime of their critical services are paramount, not flexibility.

But, not every data center will be this way – certainly not those that are exclusively focused on training. Other operators, in other grids, with other compute priorities, will be able to do things differently. For more on what that might be, it’s worth querying the rest of this big, multi-year proceeding yourself. Jump in here, and check it out for yourself.