Data centers are not the villain of the energy story. They might turn out to be one of the more useful characters in it.
That sounds wrong, I know. The dominant narrative right now is simple and grim: AI eats power, power grids are strained, therefore AI is a problem to be contained. It’s a tidy story. It’s also missing something important, which a pilot program in Santa Clara is now trying to demonstrate in public.
In April 2026, Silicon Valley Power, the municipally owned utility serving the City of Santa Clara, announced a pilot with a startup called Emerald AI. The goal is to show how flexible data centers can support grid reliability while unlocking power capacity for AI. NVIDIA has also collaborated with Emerald AI on this effort, and Emerald AI’s software has been integrated with NVIDIA DSX Flex.
Let me translate what that actually means, because the press release language hides a genuinely interesting idea.
The assumption everyone makes about data centers
When people picture a data center, they picture a constant. A building that draws a fixed, enormous amount of electricity, twenty-four hours a day, indifferent to everything happening around it. Under that assumption, the only question a utility can ask is whether there’s enough capacity to say yes. If there isn’t, the answer is no, or wait several years.
That assumption is what Emerald AI’s software is designed to challenge. The software adjusts a data center’s power consumption in response to grid conditions. So instead of a flat, immovable block of demand, you get something that can flex — easing off when the grid is under pressure, running harder when there’s room.
If you’ve ever set your dishwasher to run overnight because electricity is cheaper then, you already understand the principle. The work still gets done. It just gets done at a moment that’s easier on the system.
Why this matters for anyone who uses AI agents
Here’s the part that connects to what we usually talk about on this site. AI agents — the tools that research things for you, write drafts, handle multi-step tasks — all run on someone else’s hardware in someone else’s building. Every request you make turns into electricity somewhere.
Which means the availability of AI tools is quietly tied to the availability of power. When a utility can’t connect a new data center, that’s not just an infrastructure story. It eventually shows up as slower services, higher prices, or capacity limits on the products people want to use.
The pilot’s stated aim is to unlock power capacity for AI while maintaining performance. Those last three words are doing a lot of work. A flexible data center is only useful if flexibility doesn’t mean your AI assistant starts timing out during a heat wave. The entire premise rests on the idea that not every computing task is equally urgent.
And that’s true in ways most people don’t think about. Training a large model is a long, patient process. It doesn’t care much whether a particular hour of computation happens now or in ninety minutes. Answering your question right now, though, is urgent. A system that can tell the difference between those two kinds of work has room to maneuver that a naive system doesn’t.
Why a city utility is the right place to test this
Silicon Valley Power being municipally owned is a detail I find genuinely interesting. This isn’t a private experiment happening behind a fence. It’s a city-owned utility working with a startup and a chip company in one of the most data-center-dense parts of the country.
Santa Clara is roughly the hardest possible test case. If flexible demand can be shown to work there, the argument travels. If it can’t, that’s worth knowing early, before more places bet on the concept.
What I’d watch for
I want to be careful not to oversell a pilot. A pilot is a pilot. It’s designed to produce evidence, not conclusions, and the evidence isn’t in yet.
The questions I’d keep in mind:
- How much flexibility is actually available before performance suffers in a way users notice
- Whether the approach works across different operators, not just ones already partnered with NVIDIA
- Whether utilities elsewhere find the results persuasive enough to change how they handle connection requests
What I like about this pilot is that it treats the problem as a coordination challenge rather than a moral one. The argument isn’t that AI’s energy appetite is fine. It’s that a data center willing to cooperate with the grid is a different kind of customer than one that isn’t — and that difference might be worth more than it sounds.
For the rest of us, the takeaway is smaller but still useful. The AI tools you use are physical things drawing real power from real wires. Teaching them to pay attention to that is not a constraint on progress. It’s part of what makes the next round of progress possible.
🕒 Published: