What bitcoin miners know about power

At MARA I worked on the finance side of a bitcoin mining fleet that drew 1.8 gigawatts across 18 sites, and modeled the economics of converting that capacity to AI and HPC. For a miner, power is the biggest cost by far. The bitcoin network sets what a miner earns for its computing, so the main thing a miner controls is what it pays for power, and when it uses it.

Miners built a few habits around that cost. AI data centers are starting to need them too.

Hourly prices

ERCOT, which runs most of the Texas grid, sets real-time prices every five minutes and settles them in 15-minute intervals. The swings are lopsided. At ERCOT’s West hub in 2025, the median hour cost $27.34 per megawatt-hour. The most expensive hour, 7 to 8 a.m. on January 15, cost $1,963.41, about 70 times as much. The 88 priciest hours, 1% of the year, made up 8.5% of what it cost to run all year.

Shutting off

A miner compares the price of the next hour with what that hour of mining will earn, and shuts down when power costs more. Because the spikes are packed into so few hours, this costs very little running time. A site at the West hub that sat out every hour above $100 in 2025 would have run 96.9% of the year and cut its energy bill by 16%. That works out to about $46,000 a year for every megawatt that can switch off.

Getting paid to stand by

Grid operators need load that can drop on short notice, and they pay for it. ERCOT runs programs that pay large customers to stand ready to cut their usage when the grid runs short. For a miner, the ability to turn off quickly is a second source of revenue.

The four summer peaks

In ERCOT, much of a large customer’s transmission bill for the next year depends on its usage in four 15-minute intervals: the highest-demand interval in each of June, July, August and September. Miners try to be off during those intervals, which means forecasting when they will land. A load that guesses right can cut a large share of next year’s transmission charges.

How AI data centers differ

Data centers were built to run every hour of the year. Their customers expect it, and they keep backup generators on site so they never go dark. That made sense when power was easy to get. Now new large loads can wait years for a grid connection, and Texas passed Senate Bill 6 in 2025, which sets new rules for very large loads, including when they can be cut off in a grid emergency.

In August 2025, Google agreed with two utilities, Indiana Michigan Power and the Tennessee Valley Authority, to scale back machine learning work at some of its data centers when the utilities ask.

What I’d ask on an AI data center deal

If I were underwriting one, these are the first things I’d ask.

• How much of the power is hedged, and what happens in the hours that aren’t?
• Can any of the load move or stop? What is that worth, and who gets paid for it?
• How exposed is the site to 4CP and other peak-based charges?
• When does the grid connection go live, and what is the plan until then?
• What will the backup generators cost to run, and what do their permits allow?

Price figures are from ERCOT’s real-time settlement point prices for 2025 at the West hub (HB_WEST), averaged to hours. The calculator page has the source and the method.