Learnings from VT Data Center Summit

I recently had the opportunity to attend a Data Center Summit hosted by Virginia Tech (VT). The event brought together a diverse group of academic researchers, policy officials, and industry executives to discuss trends and structural shifts occurring in digital and physical infrastructure driven by the rise of artificial intelligence and machine learning.

Northern Virginia remains the undisputed heart of global data storage, currently accounting for 13% of all global data center capacity (25% of capacity in the Americas). As the industry faces unprecedented power demands and space constraints, speakers reflected on collaboration opportunities and shared key insights on topics ranging from power grid architecture to AI performance metrics.

Here are the core learnings and insights from across the summit’s panel sessions and presentations.

The Power Solution: Capacity, Reliability & the Energy Mix

The influx of data centers and general increasing electricity demand puts a strain on existing power grids. To date, Dominion Energy has connected over 10 gigawatts (GW) of capacity dedicated solely to data centers in the region. For comparison, the entire Richmond metro area requires about 1 GW. By the end of 2025, Dominion had around 48.5 GW of contracted data center capacity in its pipeline.

The pace of growth is staggering. Strikingly, 8 of Virginia’s top 10 historical peak demand periods occurred in 2026 alone, with the highest peak hitting 25 GW in February 2026.  Dominion anticipates adding 25 GW of capacity with set projects over the next five to six years, essentially doubling its capacity over that timeframe. Challenges to building necessary infrastructure include community acceptance, which in some cases is escalating from NIMBY (Not In My Back Yard) to BANANA (Build Absolutely Nothing Anywhere Near Anything).

Data centers energy needs that are driven by the transition from traditional CPUs to power-hungry GPUs to handle heavy AI training workloads. A standard server rack historically drew about 3 kW while now it is more common to see racks drawing 50 kW or more for AI and ML clusters.  NVIDIA is exploring 800V DC power architecture design that would enable the support of 1 MW racks starting in 2027.

A final topic of discussion was around ensuring that data centers pay their fair share so residential ratepayers aren’t left on the hook for massive infrastructure costs. The Virginia State Corporation Commission approved a new GS-5 rate class for Dominion Energy to use for any customer consuming over 25 MW. Under these updated contracts, data center operators must pay a minimum of 85% of contracted distribution/transmission demand and 60% of generation demand regardless of actual consumption. Furthermore, operators must sign a 14-year contract backed by collateral to protect against “sunk cost” risks if a facility shuts down prematurely.

Building at Scale: How Data Centers Get Designed, Built & Deployed

Building a single data center is a multi-year endeavor that typically takes 3 to 5 years. The construction industry has mastered how to build an individual facility; the challenge now is building them at scale.

One speaker stated that the two greatest threats to data center deployment timelines are power availability and labor efficiency. Small changes like opening up a second entryway to the build site can provide significant productivity gains during construction to alleviate a singular bottleneck at the start or end of the day. A multifaceted approach includes engaging the construction team with design early and shifting as much production offsite as possible through modularization. Equipment can be pre-procured with confidence that it will work. Rather than assembling complex electrical and mechanical systems in the field, they are built in controlled factory environments that also enable testing whole systems, or components, to identify failure points.

Additional points of flexibility include designing for scale where day one load is more modest and infrastructure is adapted with no throwaway for a ramp up in load. Another approach involves starting with behind-the-meter generation and more intentional timing of connecting to the grid, assisting with “use it or lose it” power contracts.

Grab Bag Lunch Breakout Group

A smattering of takeaways from lunch sessions:

• Offshore data centers could be a possible solution to cooling needs.
• VT BRICCS lab is researching a variety of concepts including AI chips designed to mimic the brain – with parallel compute and much lower power consumption.
• TerraPower is constructing its first Natrium plant, which features two separate “islands” in a decoupled architecture.

Inside the Machine: Compute, Cloud & Complex Workloads

Data center design involves understanding the workloads that will be supported by that data center.

• Training (like a construction project): A concentrated engineering effort where teams care about throughput and floating point operations per second (FLOPS). Training requires vast quantities of interconnected GPUs computing across racks, consuming data, and testing accuracy over prolonged cycles.
• Inference (like a public utility): Once a model is trained, running it to respond to user prompts is more like serving customers on a utility grid. Here, the primary unit of throughput is measured in tokens. The engineering challenges shift to include managing latency and ensuring there is enough high-bandwidth memory to store conversational contexts.
• Agentic: an emerging challenge as these workloads require longer running times and require long-term memory architectures.

Panelists also engaged in a brief discussion around workload efficiency. The gold standard metric for data center efficiency has been PUE (Power Usage Effectiveness), which measures how much power enters a facility compared to how much actually reaches the computing equipment. One speaker shared his favorite emerging metric is tokens per watt-hour, measuring the actual useful computational output of an AI model relative to the energy consumed.

Closing Thoughts

The 2026 Virginia Tech Data Center Summit highlighted how technology and the physical infrastructure supporting it is evolving and progressing.  The progression from standard CPU racks toward 1-megawatt GPU clusters illustrates the demands data centers are placing on  power grids. These stressors are requiring innovation and collaboration between academia and industry entities such as data center operators, constructors, and utilities.