Key Takeaways

  • Data centers consume between 5% and 10% of total electricity in the United States, and multi-trillion dollar infrastructure build-outs will drive a massive share of marginal grid growth.
  • Bitcoin operators like Riot make money during grid stress by cutting power to their ASICs because crypto hardware depreciation is only 1.5 to 3 times energy expenses.
  • AI hardware flips this math: depreciation on graphics processing units like the Nvidia H100 runs 5 to 10 times higher than the electricity required to power them.
  • Because idle silicon loses substantial monetary value every hour it sits dark, AI data centers must run as 24/7 baseload power consumers rather than interruptible loads.

The Balance Sheet Problem Behind Demand Response

When electrical grid operators manage sudden changes in power usage, they look for flexible consumers. Power grids must maintain instantaneous balance across baseload generation, intermediate plants, and peaker facilities. For years, cryptocurrency mining facilities served as useful partners. When a summer heatwave arrives and electricity prices spike to thousands of dollars per megawatt-hour, operators like Riot simply switch off their machines. They get paid by grid operators to stop consuming power, then turn their machines back on once temperatures cool and prices drop back to normal.

Grid planners assumed the rapid build-out of artificial intelligence facilities would provide similar flexibility. That expectation ignores the balance sheet. Patrick McKenzie explains that the underlying economics of crypto mining and machine learning compute pull operators in opposite directions.

Why Crypto Miners Pause and AI Chips Cannot

The difference comes down to the clock on hardware obsolescence. A Bitcoin mining machine runs on specialized application-specific integrated circuits. The depreciation on those chips is roughly 1.5 to 3 times the cost of the electricity they burn. If wholesale power prices spike above the expected value of the coins being mined, turning off the machine saves money. The owner eats a modest depreciation loss while avoiding a massive power bill.

AI infrastructure operates under an entirely different cost ratio. McKenzie points out that for training and inference hardware like Nvidia H100s, depreciation runs 5 to 10 times higher than the electricity bill. A server rack filled with GPUs pulls heavy electric current, but the financial decay of those chips happens every hour regardless of whether they are running active inference workloads or sitting idle.

If an AI data center operator shuts off an H100 cluster during peak demand, the saved energy represents pocket change compared to the lost value of the idle hardware. The silicon loses value every minute because newer, faster chips will inevitably make it obsolete within a few years. As McKenzie explains: “Why do you expect AI data centers to run 24/7, where you don't expect that necessarily from crypto miners? It's because the expense associated with doing AI inference is mostly depreciation on the chips and not electricity.”

Data centers already account for 5% to 10% of total United States electricity consumption. McKenzie notes that as companies commit trillions of dollars to AI infrastructure across global supply chains, data centers will swallow a huge portion of marginal power growth. Yet because an idle GPU destroys capital faster than high electric rates do, operators cannot participate in standard demand response programs. They must act as persistent baseload consumers, competing for continuous generation and creating bottlenecks in grid interconnection queues.

What to Do With This

Audit your compute infrastructure this week. If your company pays for reserved GPU instances on cloud providers, calculate your hardware spend per idle hour alongside your actual model query volumes. If your utilization rate dips below 70% during night shifts or weekends, stop buying reserved capacity and re-architect your inference stack to spin down to spot capacity or offload low-latency requirements to shared endpoints. Letting an H100 sit idle burns your runway faster than your utility bill ever will.