The most profitable thing a Bitcoin miner can do is sometimes nothing at all. That was the argument of a session by Nate Givens, Director of Machine Learning Engineering at Revolve Labs, at Mining Disrupt Miami 2024 on June 24–26, 2024. Givens, who has a background in math, systems engineering and economics and has worked in finance, healthcare, logistics and gaming, argued that hardware efficiency gains are plateauing, and that machine learning models should decide when a fleet runs, when it curtails and when it gets repaired. Revolve Labs, he told the room, does hosting for Bitcoin mining and HPC AI, runs a carbon negative model and offers operations and ASIC repair.
On June 24, 2024, Bitcoin traded at $61,272; on October 16, 2026, it traded at $82,531. The two problems he opened with, rising power prices and slowing gains in ASIC efficiency, and the 2028 halving he pointed to, are still ahead of every operator.
The End of the Efficiency Gold Rush
Givens opened his presentation by addressing the two primary headwinds facing modern Bitcoin mining: rising energy costs and diminishing returns on hardware efficiency. He displayed a chart tracking the monthly average cost of electricity from November 1978 to May 2024, a line that, as he put it, doesn't go down in the long run. Against this backdrop, he showed a second chart detailing the efficiency of ASIC miners, measured in joules per terahash, by the year of their release. While early generations of hardware saw dramatic leaps in efficiency, Givens said those gains are “kind of slowing down, at least in nominal terms,” and the chart shows the frontier of new machines, not the fleet in the field.
we're kind of plateauing in how much we can expect to get out of increasing efficiency
Nate Givens · 2:05
The halving adds pressure, he said. He cited a pre-2024 industry quote that operators with higher production costs per Bitcoin would have a harder time surviving the halving, and said it will apply again in 2028 and 2032. His answer was to look at energy and operations, and to use machine learning on both.
Why Your Credit Card Knows Your Risk
Before diving into mining-specific applications, Givens took the audience on a brief historical tour of artificial intelligence and machine learning. He traced the field’s origins to a 1956 summer seminar at Dartmouth, where researchers aimed to create machines that were not just as smart as humans, but smart like humans. This initial philosophy led to the development of neural networks designed to mimic human cognition, but these early efforts failed, leading to the first AI winter between 1974 and 1980.
The narrative of AI’s revival is one of trial and error. In the 1980s, researchers turned to expert systems, which involved encoding the rules from human experts into computers. Early research found the pooled rules beat asking the experts directly, but it did not pan out, leading to a second AI winter from 1987 to 1993. Expert systems worked on the second try in the 1990s, he said, because there was far more compute power and memory storage. Givens highlighted Deep Blue, the expert system that defeated the reigning chess champion,, the first time an AI beat the reigning chess champion. Machine learning took over in the late 1990s, when researchers stopped asking human experts for the rules and gave the machine data instead. Neural nets finally worked too, he said, once there was more compute and memory and the field gave up on making computers think like humans; deep learning, with its layers of hidden nodes, became the basis for generative AI.
we just give a ton of data to the machine and say, you figure the rules out
Nate Givens · 5:38
Givens said machine learning is no longer a theory: it is embedded in everyday products. He pointed to credit risk assessment as a prime example. When an individual applies for a credit card, the decision is made by a machine learning model, specifically a gradient boosted tree ensemble. This same technology powers the recommendation engines of Spotify, Amazon, and Netflix. He added language translation, which only became reliable once deep learning models were trained on millions of translated words, and generative AI built on generative adversarial networks.
The Zero-Dollar Peak Hour
Returning to the mining context, Givens detailed two specific energy optimization strategies where machine learning offers distinct advantages. The first is real-time curtailment. Large-scale electricity consumers participate in day-ahead and futures markets, allowing them to lock in prices. If the real-time price the next day goes very high, they can turn machines off and get paid for not using electricity.
you will actually get paid more for not using electricity than you would for mining
Nate Givens · 10:16
Whether to curtail depends on the price of the coin being mined and how long the spike is expected to last. A model can take that data and decide quickly, around the clock, without a person awake at 2 a.m. He called the result “kind of marginal savings.” The second strategy, which Givens called “a much bigger deal just in terms of the amount of money that's on the table,” is avoiding capacity charges or transmission fees. These are assessed differently in different markets. Where Revolve operates, each month the network looks at the hour with the highest load and charges users on what they drew in that hour, the coincidental peak. This fee can represent 25 to 30 percent of the entire electricity bill.
If you're not using electricity during that one peak hour, then your fee is zero dollars
Nate Givens · 11:46
The catch is that the peak hour has to be known ahead of time, every month. If a miner knows which hour will be the peak, they can curtail their operations for that specific window, reducing their fee to zero. Givens explained that while the two strategies—real-time curtailment and peak hour avoidance—can be run independently, what operators really want is to combine them. When a model is highly confident about the peak hour, it frees up other hours for real-time curtailment opportunities. When confidence is lower, the model may need to curtail multiple hours to hedge against the uncertainty of the peak. He gave the example of a month where the peak is probably Tuesday at 2 but could be Thursday at 4: then both hours get curtailed.
When Turning Off Becomes a Profit Center
What he wants is one model doing both optimizations at the same time, and he named the type.
a reinforcement learner is the type of model that would be really good at that
Nate Givens · 13:04
The two models can run independently, “the state that we're at right now,” he said; the combined reinforcement learner is “kind of our future state.” Givens acknowledged that implementing machine learning at scale is difficult, citing a Google white paper that notes most attempts to put ML models into production fail. A toy model in a Jupyter notebook is easy, he said; a product needs a data pipeline, monitoring and inference infrastructure, and according to the paper, senior to mid-level machine learning talent. “That's why I'm here.”
The Symbiosis of Downtime and Maintenance
The final component of Givens’ argument was the integration of energy optimization with operations optimization, specifically predictive maintenance. Miners who collect operational data—such as temperatures, fan speeds, and environmental conditions—can use machine learning to predict which machines are likely to fail. This allows for preventative maintenance, where repairs are performed before a breakdown occurs. Givens called this an obvious application of machine learning. His last slide was about how it interacts with energy.
these energy and operations optimization techniques, they're symbiotic
Nate Givens · 13:51
The better a miner becomes at energy optimization, the more hours they will curtail. During these curtailment hours, the miners are offline anyway. His example: if the fleet will curtail three hours this coming Thursday to avoid the peak, move the scheduled preventative maintenance into those three hours rather than waste additional time and money. Every hour a fleet is curtailed also makes it more important that the machines work when they are scheduled to run. Givens called a single model that synthesizes energy and operations optimization “kind of the holy grail here.” Revolve Labs was at booth 156.
Intelligence Disrupt 2027 runs March 22 to 24 in Irving, Texas. The full session is on our page above.