AI data centers adopt Bitcoin-miner power controls

Luxor Energy and software firm Bentaus ran a live test in Texas that reduced a single Nvidia B200 AI GPU’s power to about 25% of normal within half a second during an inference workload. The companies used software controls to limit the chip’s electricity use while it was producing model outputs.

Luxor reported that the GPU processed fewer requests while throttled and that no job failed and no work already in progress was lost. The team did not disclose which model was running, how long the GPU remained throttled, the exact drop in throughput, or the effect on user wait times. The test covered one GPU rather than a full server rack or a data-center campus.

Operators are testing these controls because AI workloads can draw large, sudden amounts of power. Large data-center campuses can use as much electricity as a small city. In Texas, the grid operator recorded a preliminary demand peak of 91,089 megawatts on July 22. State officials flagged about 474 gigawatts of new connection requests, roughly 90% tied to data-center projects; ERCOT’s initial review found about 205 gigawatts had reached sufficient study milestones and regulators ordered an audit of the list.

The experiment builds on practices used by Bitcoin miners, who routinely shut down equipment when wholesale prices spike or grid emergencies approach. Miners have participated in demand-response programs because their compute jobs can be interrupted and restarted without preserving in-progress customer tasks. Luxor applied a similar approach to an inference workload, using live grid signals to decide when to throttle a GPU. The company treated the action as a privately timed curtailment aimed at reducing transmission billing exposure tied to Texas’s four highest system peaks each summer, known as 4CP.

Larger-scale tests and commercial deployments have explored related ideas. A cloud experiment reduced power on a 256-GPU cluster by 25% for three hours while protecting high-priority jobs. One major cloud operator reported contracting roughly one gigawatt of demand-response capacity across multiple U.S. regions by March 2026. A startup that worked on the Phoenix experiment raised $150 million in financing and said its software is operating across data centers.

Academic modeling provides estimates of potential flexibility. A University of Chicago working paper estimated an inference-focused facility could reliably commit to cutting about 40% of demand during tight hours, while a mixed training-and-inference site might commit about 24.6%. University of Alberta modeling found that delaying or moving AI jobs could reduce needed power-plant capacity by more than 21% in some stressed scenarios and lower total system costs under certain congestion patterns.

Important limits remain. Throttling GPUs affects only a portion of a data center’s load; cooling, networking, storage and power conversion still draw electricity. Coordinating thousands of machines to slow down and then ramp back up without creating a simultaneous rebound is a technical and contractual challenge. Grid operators require verifiable meter data to count reductions, and customer service agreements will determine which tasks can be delayed. Texas passed Senate Bill 6 in 2025 requiring curtailment obligations for some large new connections and calling for programs to pay sites that can reduce demand; regulators are also proposing changes to transmission billing that would use a 12-month peak measure instead of four summer peaks.

Luxor plans follow-up tests with groups of Nvidia H100 GPUs in Texas. Next steps include developing software that classifies job urgency and coordinates throttles across many machines while honoring performance guarantees. The single-GPU test demonstrated rapid, software-driven power cuts; scaling that capability into predictable, meter-verifiable megawatts across whole facilities remains the central challenge for integrating AI demand with grid operations.