CryptoSlate
Luxor Energy, a Texas Bitcoin mining infrastructure firm, paired with software maker Bentaus to throttle a single Nvidia B200 GPU running an inference workload down to roughly 25% of normal power in half a second. The chip returned to full speed when the restriction ended, with no jobs failing or work in progress lost, according to chief operating officer Ethan Vera. The experiment borrows directly from a playbook Bitcoin miners have run on the ERCOT grid for years: shut down when wholesale prices spike or a transmission-billing window is forming, then restart when conditions ease.
Why it matters
Bitcoin miners were the main price-sensitive participants in ERCOT's emergency demand-response programs, according to an April review by the grid operator. AI data centers, by contrast, are treated as firm loads: their tens of thousands of GPUs run continuously because every idle minute delays work customers are paying for. The Texas experiment, alongside a 256-GPU test at an Oracle cloud facility in Phoenix, suggests that line is starting to blur. Google said in March 2026 that it had put one gigawatt of data-center demand response under long-term utility contracts across several US regions, while Emerald AI, which led the Phoenix work, raised a $150 million round on August 25 valuing it above $1 billion. Flexible AI load is moving from research papers into commercial scale.
Market impact
Texas is where the pressure lands first. The Electric Reliability Council of Texas hit a preliminary July 22 record of 91,089 megawatts, while Governor Greg Abbott said in August that ERCOT was reviewing requests to connect more than 474 gigawatts of new load, roughly 90% from data centers. Senate Bill 6, passed in 2025, requires certain large new power users to cut consumption during severe grid emergencies, and a University of Chicago working paper estimated that an inference-focused facility could reliably commit to shedding 40% of its demand. The financial incentive is real: 2026 PUCT figures put ERCOT transmission costs near $6 billion against an average 4CP demand of 80,860 megawatts, or roughly $74.89 per kilowatt per year, meaning 100 megawatts of load during the four peak windows represents about $7.5 million in annual transmission exposure. Luxor's next test scales to Nvidia H100 GPUs.
Frequently asked questions
-
What did Luxor and Bentaus demonstrate with the Nvidia B200?
They showed that a B200 running inference could throttle its power draw to roughly 25% of normal within half a second, then return to full speed without failing any in-progress jobs, according to Luxor COO Ethan Vera.
-
How do Bitcoin miners use demand-response on the ERCOT grid?
Miners shut down when wholesale prices spike or a 4CP transmission-billing window is forming, get paid for cutting power during emergencies, and restart when conditions ease. ERCOT's April review called them the main price-sensitive participants in one emergency program.
-
Why does the ERCOT 4CP system matter for AI data centers?
4CP bases part of large users' transmission bills on power use during four 15-minute summer peaks. At 2026 rates, 100 MW of demand during those windows represents roughly $7.5 million in annual transmission costs, giving AI sites a strong reason to throttle.
-
How much data-center electricity demand is queued in Texas?
Governor Greg Abbott said in August that ERCOT was reviewing requests to connect 474 gigawatts of new load, with about 90% coming from data centers. Roughly 205 GW had enough supporting studies to qualify for ERCOT's first study batch.
-
What did Emerald AI show at the Oracle cloud facility in Phoenix?
A peer-reviewed Nature Energy paper described software cutting power used by a 256-GPU cluster by 25% for three hours while keeping priority jobs within promised performance. Emerald AI raised $150 million at over $1 billion valuation on August 25.
CoinTelegraph
TheBlock
CoinDesk
Crypto News