The Case for Turning Off Your Miners

Sometimes the best-paid hour in a Bitcoin mine is the hour the machines are switched off. That was the case Nate Givens, director of machine learning engineering at Revolve Labs, made at Mining Disrupt Miami 2024, held June 24 through 26, 2024. Revolve Labs does hosting for Bitcoin mining and HPC AI. On June 24, 2024, Bitcoin traded at $61,272. On October 16, 2026, it traded at $82,531.

Givens started with two problems. Power prices go up and, in his words, not down in the long run. And the efficiency of new ASIC miners is improving more slowly every generation. As he put it, we're kind of plateauing in how much we can expect to get out of increasing efficiency. With the halving adding pressure, he said, operators with higher production costs per coin will have a harder time, and that will apply again in 2028.

His answer is machine learning. He gave the audience a short history of the field, from a 1956 summer seminar at Dartmouth to the shift in the late 1990s, when, as he described it, we just give a ton of data to the machine and say, you figure the rules out.

Then he got back to mining. The first use is real-time curtailment. A large power buyer can lock in tomorrow's price in the day-ahead market. If the real-time price spikes, it can turn the miners off and get paid for not using power. As Givens said, you will actually get paid more for not using electricity than you would for mining. Whether to do it depends on the coin price and how long the spike will last, and a model can make that call at two in the morning without waking anyone.

The second use is bigger money. In the market where Revolve operates, part of each monthly bill is set by what a customer draws during the single hour of highest load on the network. If you're not using electricity during that one peak hour, then your fee is zero dollars, he said, and that fee can be 25 to 30 percent of the whole bill. The trick is knowing the hour ahead of time.

Givens wants one model doing both jobs, and he named the type: a reinforcement learner is the type of model that would be really good at that. Then he added maintenance. Collect temperatures, fan speeds and failure records, predict which machines will break, and fix them during hours the fleet is curtailed anyway. As he put it, these energy and operations optimization techniques, they're symbiotic.

The full story is on Intelligence Disrupt at intelligencedisrupt dot com, and many more like it play around the clock on Bitcoin Radio.