ISSUE 009
What AI Actually Costs to Run
I went looking for one number: what it costs to run the AI tools I use every day. I came back with a range, a revision, and a federal pilot survey that got nine responses out of fifty. None of that is a failure of searching. It is what the public record actually holds right now, and I would rather hand you that than a tidy figure.
An idea worth thinking about
Price the capability, not the model
Epoch AI, a research nonprofit that tracks AI compute and costs, published a data insight in March 2025 that does something most price reporting does not. Instead of asking what the newest model costs, it fixes a level of performance — the score GPT-4 reached on a benchmark of graduate-level science questions, say — and then tracks the cheapest model that clears that bar over time.
Measured that way, the price to reach GPT-4's score on GPQA Diamond fell roughly 40 times per year. Across the six benchmarks Epoch examined, the rate of decline ranged from 9 times per year to 900 times per year. That spread is the finding. Which number someone quotes you depends entirely on which capability they picked, and almost nobody tells you which one.
Two caveats Epoch states itself. These are published list prices for API access, not what the compute costs to produce, so they reflect competition as much as efficiency. And the steepest declines are the most recent ones, which Epoch says makes it less clear they will continue. I would add a third of my own: a benchmark score is not the work you need done, and no price curve tells you whether the cheap model is good enough for your quotes.
A tool I disappeared into
Your own state's electricity profile, kept by the federal energy statisticians
The U.S. Energy Information Administration maintains a profile page for every state, and Missouri's is short enough to read in one sitting. The 2024 edition, released 10 November 2025, puts Missouri's average retail electricity price at 11.06 cents per kilowatt-hour, which ranks 31st among the states.
The summary table fits on one screen: primary energy source, net summer capacity, net generation, emissions in totals and in pounds per megawatt-hour, retail sales, direct use, average retail price. Most rows carry a state rank, so you can see where Missouri sits without doing any arithmetic yourself. Behind it sit nineteen detailed tables and archived profiles going back to 1996.
What it will not give you is the part a business actually pays. That 11.06 cents blends residential, commercial and industrial customers, and a shop on Broadway does not pay the blended rate. For the commercial figure you need the detailed tables, or your own last twelve invoices, which are a better source about your business than any state average ever will be.
Learn something
Read the same federal forecast twice, eighteen months apart
Lawrence Berkeley National Laboratory produces the data centre energy estimate Congress asked for in the Energy Act of 2020. The December 2024 edition put U.S. data centres at 176 terawatt-hours in 2023, about 4.4 percent of national electricity, and projected a 2028 range of 325 to 580 terawatt-hours, or 6.7 to 12 percent.
The June 2026 update, LBNL-2001758, revised the baseline to 192 terawatt-hours in 2024, about 4.7 percent, and pushed the horizon to 2030: a reference case of 649 terawatt-hours, about 11.8 percent of projected U.S. electricity, with combined uncertainty bounds of 521 to 843 terawatt-hours, or 9.5 to 15.3 percent.
The two are not a like-for-like revision. Different end years, different scenario construction, and the authors do not present them as one. That is exactly why reading both is worth an hour. You learn what a careful forecast looks like when the thing being forecast moves faster than the forecasting cycle, and you stop treating any single percentage you see quoted as settled fact.
What we’re seeing
A federal pilot survey of fifty data centres got nine responses back, and that is roughly the quality of the national picture
The Congressional Research Service updated its FAQ on data centres and energy on 12 May 2026. It is a reference document written for members of Congress, not an advocacy piece, and it is unusually candid about what is not known. Overall data centre energy use, it says, is not well described nationally, and federal data collection has been limited.
The specifics are blunt. In 2021 the Energy Information Administration ran a pilot survey of fifty data centre facilities and received nine responses. In 2024 an EIA cryptocurrency data collection was halted after a legal challenge and the agency agreed to destroy what it had gathered. The estimates everyone quotes therefore rest on modelling and equipment shipment data, and CRS notes that academic work has found those estimates vary in accuracy and reliability depending on each source's method — and that because later researchers reuse them, the problems propagate.
CRS does report the figures that exist. Berkeley Lab's 176 terawatt-hours, about 4.4 percent of U.S. electricity in 2023, excluding cryptocurrency. EPRI's independent estimate of about 4 percent for the same year, close enough to be mildly reassuring. EPRI also estimates that AI accounted for 10 to 20 percent of data centre energy use in 2024, which is a very wide band for the number most people assume is the whole story.
The part that matters for a business paying a commercial power bill sits in the ratemaking section, and it runs against the common assumption. Rates are set by state and local regulators. CRS summarises a Berkeley Lab analysis covering 2019 through 2025 which found the main driver of retail price increases was utility grid investment, mostly aging infrastructure and resilience, with natural gas prices, disaster recovery and state policies also contributing. That analysis did not find data centres to be major price drivers in most areas, and CRS reports that states with the largest data centre demand growth generally saw price decreases — one explanation being that higher sales spread fixed costs across more kilowatt-hours.
I want to be careful about how far that travels. Generally saw price decreases is not will not raise your rate. CRS says areas with grid constraints are more likely to see increases from new data centre demand, and that a growing number of regulators are creating separate customer classes for data centres precisely so the cost of serving them stays with them. Whether either applies where you are is a question for your utility and the Missouri Public Service Commission, not for a national average.
Putting the four sources together, and this last paragraph is my summary rather than any one source's claim: the compute is expensive and getting cheaper per unit of capability, the electricity is a real and fast-growing share of national demand, and the line from an AI data centre to your October invoice is not one the best available federal analysis can currently draw.
One question
What has your commercial power bill actually done over the last two years?
I would like to know whether the national picture matches the one on your invoice. Reply to this email with what you saw, and I will keep a tally and report it back.
More in this pillar as it's written.
— Shana
Sources link to the original publisher, never to a summary.