The electricity problem for AI is not abstract. Grid operators in the US are fielding requests for new power connections that, if all approved, would exceed the country's entire current electricity output. Most will never be built, but even a fraction of them strain a system that was not designed to absorb this kind of growth this quickly.
The core bottleneck is time. Building a data center takes one to two years. Connecting it to a new power source takes closer to eight. PJM, the grid operator covering Virginia and 12 other eastern states, is the largest in the country and sits at the center of this crunch. Virginia alone is projected to see electricity demand rise 183% by 2040 if current data center plans go forward.
Emerald AI, a startup founded in Washington DC less than two years ago, is proposing a workaround. Its software, called Conductor, watches the state of the electricity grid in real time. When demand elsewhere spikes, it automatically reduces power to the less urgent computing jobs inside a data center, like background data processing, while keeping time-sensitive work, like live customer queries, running normally. The facility uses less power for a few hours, the grid stays balanced, and no one outside the building notices.
The numbers behind this idea are striking. Researchers at Duke University found that US grid capacity could absorb the projected growth of American data centers through 2030 if those facilities were willing to reduce their draw for just 22 hours a year. That is roughly 0.25% of the time. The Princeton research, funded by Google, found that a large facility willing to flex for less than 1% of the year could connect to the grid three to five years faster than one that will not.
Emerald AI has now raised about $68 million in total funding, with backing from Nvidia, GE Vernova, Siemens, Samsung, and others. Its first large commercial test, a 96-megawatt facility in Manassas, Virginia, is planned to open in the second half of 2026. That is roughly 750 times larger than the London simulation the company ran in December 2025.
The political environment is adding urgency. US residential electricity bills rose about 5% in 2025, with another 4% forecast for 2026. Community opposition groups fighting data center projects more than doubled between the end of 2025 and March 2026, reaching 833 active groups across 49 states. Over just the first three months of 2026, opponents blocked or delayed projects worth around $130 billion, roughly matching the total for all of 2025.
In the US Senate, a bipartisan bill called the GRID Act, introduced in February 2026 by senators from both parties, would require any new data center using 20 megawatts or more to generate all its own power and disconnect entirely from the public grid. Facilities already connected would have ten years to comply or face penalties of up to $1 million per day.
If that bill passed, it would make the flexibility approach irrelevant, pushing the problem entirely onto data center operators to solve on their own with private power plants. Most experts consider that outcome unlikely at scale, since there are not enough gas turbines or other generation assets available to meet the demand, and nuclear and other clean firm power takes even longer to build than grid connections do.
The more realistic near-term outcome is a split path. Larger operators with the money and land to build private generation will do so. Smaller and mid-sized facilities will need to negotiate flexible connection agreements with utilities. Neither path is fast or cheap, and both require decisions now about where to locate facilities and what power strategy to commit to.
For business operators who rely on cloud services, AI tools, or any software that runs in data centers, the practical consequence is likely to be higher costs passed through over time. The infrastructure bill is real and is being distributed across electricity customers, cloud pricing, and government incentive budgets. The flexibility approach, if it proves out in Virginia this year, would at least mean that more capacity comes online sooner, which puts downward pressure on the chronic undersupply. But it is not a solution to the underlying scale of the demand. It is a way to buy a few years while the larger buildout catches up.