AI data centers are being pushed to behave more like Bitcoin miners, at least when it comes to electricity. A recent experiment by Luxor Energy and Bentaus showed that a single Nvidia B200 chip doing AI inference could cut its power draw to about 25% of normal in half a second. The work didn’t fail. It just slowed down.
The test was small, but it points at a bigger idea. Data centers could sort jobs by urgency and reduce load when the grid is strained. That would give utilities a new tool, and it could turn AI’s massive power appetite into something more flexible.
Why Texas matters
Texas is the clearest example of the tension. ERCOT set a preliminary record of 91,089 megawatts on July 22. Meanwhile, the connection queue includes requests for more than 474 gigawatts of new load, mostly data centers. That is over five times the record demand, and far more than the wires can handle.
Gov. Greg Abbott ordered an audit of those projects. Regulators are trying to figure out which ones are real. Even with the uncertainty, the numbers show how fast AI buildout is moving compared to power infrastructure.
What miners taught the grid
Bitcoin miners were the first big users to accept interruptions. When electricity prices spiked, they shut down. No customer waited on a response, and machines could restart almost instantly. That made them useful in demand response programs.
AI inference is different. Customers are waiting, even if some can tolerate delay. But not all work is equally urgent. Google has delayed video processing or moved it to other regions when grids were strained. Researchers have shown that software can cut power to a cluster by 25% without breaking performance promises.
A University of Chicago working paper estimated that an inference-focused facility could cut 40% of demand. A mix of inference and training could cut 24.6%. The numbers held up even at large scale.
Money and limits
The Luxor test was tied to Texas’s Four Coincident Peaks, a billing mechanism with four 15-minute windows in the summer. Large users can save millions by cutting usage during those windows. Luxor used live grid data to predict a peak and throttled the chip to lower transmission charges.
That makes the half-second response less about grid emergencies and more about billing. But it suggests faster controls could eventually participate in grid programs.
There are reasons for caution. One GPU is not a data center. Cooling, networking, and power conversion also consume electricity. A 75% chip cut does not mean a 75% building cut. And bringing thousands of GPUs back at once could create a new spike.
Texas is moving toward requiring flexibility. Senate Bill 6 asks large users to cut power in emergencies. Regulators are also considering replacing 4CP with 12CP, which would capture more winter and evening peaks.
The challenge is proving that data centers can deliver predictable, verifiable reductions. That means contract changes, software coordination, and careful scaling. But the direction is clear. AI computing has varying urgency, and the grid needs flexible demand. These two needs may finally meet.
