650 Premature Deaths: The Real Bill for AI and Crypto's Power Habit
An MIT-led study puts roughly 650 premature US deaths in 2023 on the grid electricity data centers actually consumed. The number is small next to total air pollution deaths, but it reframes the whole data center debate around marginal megawatts instead of renewable credits. Here's who pays, who gets blamed, and what the industry should do about it.
Six hundred and fifty people. That's the estimate from an MIT-led research team for premature deaths in the contiguous US in 2023 tied to the electricity data centers pulled off the grid.
Not the electricity they bought contracts for. The electricity that actually got generated to meet their demand, in the hours they ran, at the margin. That distinction is the whole ballgame.
What the study found
The mechanics aren't complicated. When a data center flips on load, something upstream has to answer. In most US power markets that's a gas peaker, a coal unit that stayed online longer than it wanted to, or a combined-cycle plant running harder than its schedule called for. Those plants emit fine particulates, sulfur dioxide, nitrogen oxides. The compounds travel. Downwind communities breathe them.
The research traced that chain and put a human number on it. Roughly 650 premature deaths across states in 2023, with the pollution burden landing in different places than the load itself.
That's the part nobody puts in the press release. A county hosts the servers, collects the property tax, gets the ribbon cutting. A county 80 miles downwind gets the asthma. Ohio's load. Pennsylvania's air. Texas powering racks for California customers.
What you need to know: this isn't a rounding error, and it isn't a solved problem.
Scale matters here. Data centers were about 4.4% of US electricity consumption in 2023, according to Lawrence Berkeley National Laboratory. By 2028 they could be somewhere between 6.7% and 12%. So 2023 is the floor. If the grid mix doesn't shift faster than demand grows, the 650 number gets bigger, not smaller.
The fight over marginal megawatts
Here's where the industry's standard defense falls apart. The answer to "your load is dirty" has been a power purchase agreement and a pile of renewable energy credits. Sounds clean. Isn't.
Most RECs are annual and unbundled. You can buy a certificate from a wind farm that was built a decade ago and already fully subscribed, then claim your 2am load in Virginia is green. Ask yourself this: if the wind farm would have run exactly the same number of hours without your check, what did you actually clean up?
Not much. And the market is starting to figure that out.
The serious operators already know. Microsoft signed a deal to restart Three Mile Island. Amazon went to Talen's Susquehanna site. Those are real, physical, location-specific megawatts. But nuclear restarts take years, small modular reactors take longer, and the grid is short right now. The gap gets filled by whatever's cheapest and fastest, and that's usually gas.
Price is telling the same story. PJM's capacity auction cleared at $269.92 per megawatt-day for the 2025/2026 period, up from $28.92 the year before. That's roughly a ninefold jump. New load is expensive, and the bill lands on ratepayers who never asked for a hyperscale neighbor.
Which brings up the blame problem. Bitcoin miners are maybe 1% to 2% of US power demand. AI data centers are the growth story, the ones adding gigawatts to interconnection queues. But headlines don't split hairs, and "crypto is killing people" is a much better click than "load growth is outpacing clean generation."
My take: miners will absorb blame that belongs mostly to AI. That's unfair and it's also predictable. The industry spent years fighting aggregate emissions statistics instead of publishing the granular data it already collects. Every ASIC on a rack reports its uptime. Every site knows its hourly consumption and its zip code. Publish it. Fund air quality monitors in host counties. That costs real money and it buys something more valuable than a press cycle.
Here's the second thing. The 650 figure is small against total US deaths from fine particulate pollution, which most estimates put north of 100,000 a year. Attribution studies like this carry wide error bars and depend heavily on which plants get modeled as marginal. So a skeptic can reasonably argue the precision is overstated.
But the direction is right. And arguing about the decimal places is a losing strategy for an industry that needs political goodwill to keep its interconnection agreements.
On policy, the tailwind is real. Permitting reform and a friendlier posture toward baseload generation are the only paths to adding supply fast enough to absorb AI demand without pushing marginal emissions up. Fewer hurdles for gas, nuclear, and transmission means less dirt per megawatt-hour. That's the economic case, and it's the one that matters.
The number that matters today
650 is a research estimate. The number that actually runs the industry is 6.7%, the low end of the projected data center share of US electricity by 2028.
Double that share without changing the generation mix and you don't need a new study to guess what happens to the death count. You just multiply.
One thing to watch: hourly clean energy matching. A handful of large buyers are already writing 24/7 carbon-free energy requirements into contracts, matched to the actual grid region and the actual hour. If that becomes the standard for new data center deals, the marginal plant answering your load stops being a coal unit in the next state. If it stays a niche, the accounting stays cosmetic and the pollution stays exactly where it's.
The industry can shape which one happens. It won't get to choose whether people keep counting.