For most of the internet era, a data center was a building that stored information and ran software. AI is turning it into something closer to heavy industry. The technology sector now calls these facilities AI factories, and the name fits. A factory turns raw materials into products. An AI factory turns electricity, chips and data into computing output, at a scale of power demand that looks more like a steel mill than an office park.
That changes the conversation. The AI race has mostly been described in terms of GPUs, semiconductors and model size. Electricity is now one of the main limits on where AI infrastructure can actually be built. Utilities are fielding requests for hundreds of megawatts, and some proposals approach several gigawatts. The generation, transmission and substations needed to serve them take years to build. The core problem is simple: AI can be deployed much faster than the grid can be expanded.
The rack is where the problem starts
The shift is easiest to see at the level of a single server rack. According to a Utility Dive discussion sponsored by Siemens Grid Software, which included speakers from NVIDIA and Dominion Energy, traditional racks drew roughly 15 to 25 kW. AI hardware has blown past that, and the roadmap keeps climbing.
| Rack type | Approximate power per rack |
|---|---|
| Traditional enterprise rack | 15 to 25 kW |
| Current AI rack (NVIDIA Blackwell generation) | About 120 kW |
| High-end AI deployments cited by Siemens and NVIDIA | About 230 kW |
| NVIDIA Rubin Ultra "Kyber" rack (planned 2H 2027) | Up to 600 kW |
| Future designs discussed by industry | Approaching 1 MW |
According to Tom's Hardware, NVIDIA plans to ship its Kyber racks, with 576 GPUs each, in the second half of 2027 at up to 600 kW per rack. At that density, a single row of racks can draw as much power as a small town. The industry is redesigning power delivery itself, moving toward 800-volt DC power inside the data center as early as 2027 to cut losses and copper.
The exact figure for any given site matters less than the trend: far more power in the same floor space, and far more pressure on everything upstream of the building. As I argued in Is Compute the New Oil?, an AI factory is essentially a machine for converting megawatts into model output.
Power, power, power
Commercial real estate has always been about location. For AI infrastructure, that rule is being replaced. "The old mantra of location, location and location has really given way to speed to power," Britt Burt of Industrial Info Resources told The Daily Reflector. ZON's Andrew Denis goes further, calling power "the defining constraint — not chips, not land, not capital."
This reverses the usual order of site selection. Developers used to find good land and then figure out how to power it. Now they increasingly start with where power exists, or where new generation can realistically be built, and work backward. Cheap electricity that needs seven years of transmission upgrades can lose to pricier power available in three.
A more useful way to score a site:
Time-to-Power + Cost-to-Power + Reliability-of-Power + Risk-to-Power
For a 500 MW or 1 GW campus, the gap between energizing in year three and year seven can outweigh modest differences in land cost or power rates. Time itself becomes a cost. And it's getting longer: high-voltage transformer lead times reached up to 160 weeks in 2026, part of the AI power wall now slowing projects nationwide.
The demand is concentrated, not just larger
The national numbers are large. In its October 2026 outlook, the U.S. Energy Information Administration projected record electricity use in both 2026 and 2027, driven largely by data centers. Data centers used about 4.4% of U.S. electricity in 2024, according to Penn State research cited by The Daily Reflector, and a Bloomberg analysis projects they could reach up to 20% by 2035. Forecasts rest on very different assumptions about AI adoption and efficiency, so no single projection should be treated as certain. Nearly all of them point the same way.
The regional picture matters more for utilities. Reporting cited by The Daily Reflector suggests data centers in Virginia could rise from about a quarter of the state's electricity use to 57% by 2030. Wyoming and Iowa could see data center use triple to around 30%, and Ohio, Texas and New Mexico could go from single digits to more than 20%.
Electricity doesn't move freely. Transmission lines have thermal limits, and substations and transformers have capacity ceilings. A region can produce plenty of energy over a year and still be unable to deliver 500 MW to one particular substation. That's why asking whether the country has "enough electricity" misses the real constraint: deliverability at a specific location on a specific timeline.
Energy is now part of the AI tech stack
One line from the Utility Dive discussion, attributed to NVIDIA, sums it up: energy has become part of the AI technology stack. Power used to be something a developer bought after the facility was designed. Now it shapes the design, location, schedule and economics from day one. GPUs that can't be energized earn nothing, and a finished building waiting on transmission upgrades is a stranded asset.
That creates a new dependency between utilities and tech companies, and neither side can plan well without better information from the other. Utilities need to know which projects are real, when load will arrive and how fast it will ramp. Developers need to know where capacity exists, what upgrades are required, and how long and how much those upgrades will cost.
Planning for many futures, not one
A 1 GW request doesn't mean 1 GW of load, now or ever. Projects get delayed, resized, phased, moved or canceled. If utilities build billions in infrastructure for demand that never shows up, existing customers can be left holding stranded costs. If they underestimate demand that's real, they choke off economic development.
The risk isn't theoretical. In PJM, the grid operator serving 13 states and Washington, D.C., utilities recently cut their near-term large-load forecasts by about 6.5 GW, and the emergency capacity procurement designed for data center growth could shrink sharply as a result.
The Siemens discussion recommends moving from single forecasts to scenario-based, probabilistic planning: testing many credible futures and prioritizing investments that hold up across most of them. That means separating requested load from contracted load and from credible load, and backing the difference with real commitments such as deposits, minimum-demand contracts, financial guarantees and phased interconnection schedules. As Parag Nathaney put it to The Daily Reflector, "Data center load growth will be limited by how much new grid supply can be added."
Flexibility: the fastest way through the bottleneck
One way through the bottleneck is to stop treating every AI data center as a load that can never move. Not all compute is equally time-sensitive. Real-time inference may need constant availability, but some training workloads can be scheduled, throttled or shifted to hours when the grid has headroom.
The potential is large. A 2025 study from Duke University's Nicholas Institute estimated the U.S. grid could absorb about 76 GW of new load, roughly 10% of national peak demand, without major new capacity, if those loads curtailed just 0.25% of the time.
At 1% curtailment, that rises to about 126 GW. The study found the typical curtailment event would last about two hours, the kind of window a data center can often cover with batteries, on-site generation or shifted workloads.
Bitcoin miners have shown for years that large loads can act this way, and Texas now requires newer large loads to be curtailable in grid emergencies. Battery energy storage makes flexibility easier for AI facilities too. Batteries don't create energy, but they change when a facility draws it, providing ride-through, peak shaving and backup.
Where the grid can't deliver in time, developers are turning to behind-the-meter generation, microgrids and dedicated generation paired with storage. At this scale, that's a power project in its own right, with fuel supply, permitting, emissions rules, redundancy and equipment lead times to manage. The next generation of data center developers may need the skill set of an independent power producer.
AI can help fix the problem it created
There's an irony here. Utility planning teams are being asked to study far more interconnection requests and scenarios without matching increases in engineering staff. AI-enabled planning tools can automate repetitive studies, integrate large datasets and run thousands of generation, transmission and storage combinations instead of a handful. Used well, they extend engineering judgment rather than replace it. AI may end up as both one of the biggest new sources of grid demand and one of the main tools for managing it.
Who pays still matters
Fast growth brings a familiar question: who pays for the new substations, lines and generation? According to The Daily Reflector, communities near major clusters in Virginia, Texas and Georgia have seen electricity rates rise 8% to 15%, and data centers have contributed more than $9 billion in added capacity costs in the Mid-Atlantic. States are responding with large-load tariffs, from Ohio's approved AEP tariff to North Carolina's proposed Duke Energy deal, which I covered in Who Should Pay for the AI Power Boom.
Developers don't have to see this purely as a headwind. Clear cost-allocation rules give them certainty, and knowing the price of admission up front usually beats years of regulatory limbo.
What this means for developers and utilities
For developers, treat power as the first siting requirement, not a procurement task after the land is bought:
- Confirm deliverable capacity at the specific substation, not just regional supply.
- Order long-lead equipment early, and plan for transformer lead times measured in years.
- Design for flexibility from day one, with storage, workload scheduling or on-site generation.
- Expect minimum bills, collateral and long contracts in most large-load tariffs.
For utilities, separate speculative requests from committed ones, plan across multiple scenarios, and reward large customers that can flex their demand.
The bottom line: the AI race is becoming a power race
For years the questions were who has the GPUs, who has the best models, and who can build the largest cluster. Those still matter, but a more basic question is moving to the center: who can actually power them?
AI factories are making electricity demand larger, denser and more geographically concentrated, while power plants, transmission lines and substations are still built on infrastructure timelines, not software timelines. Developers used to ask, "Where should we build?" The AI era is forcing a different question: "Where can we actually get the power, and how fast?" The answer will go a long way toward deciding where the next generation of AI infrastructure gets built.
Chris Kalowes
Founder, WattThe?!
Energy intelligence. Simplified.
