The U.S. Economy’s Omni-Trend

This is an edition of Heatmap Daily, an evening review of the day’s news written by our executive editor. Sign up for it here.

It isn’t exactly a new thought, but I’ve been struck recently by how many trends in America’s economic and environmental data are fundamentally about the data center boom and the return of electricity demand:

First, the Energy Information Administration reported this week that U.S. emissions grew by more than 2% last year, driven by surging electricity demand and an increase in coal-fired generation. What caused that higher power demand? New factories and data centers — as well as record summertime cooling demand.

Second, many of the new factories driving that higher power demand are themselves producing goods that are … let’s say … data center-adjacent. There are the enormous new semiconductor fabs, of course. But Ford and General Motors have also set up new production lines (or repurposed old ones) to manufacture grid-scale batteries to meet power demand.

Third, take a look at the recent U.S. spending on private non-residential construction — in other words, everything American companies are building that is not houses, condos, or apartments.

The construction industry’s spent almost $60 billion on data centers over the past year, which is more than it spent on all other office buildings combined (and more than it spent building warehouses, too). Just a handful of categories — data centers, power plants, electricity infrastructure, and certain kinds of electronics manufacturing — now make up a third of all U.S. private non-residential construction investment. They’ve never made up such a large share of construction spending since data collection began in 2014.

As The New York Times recently noted, the American economy is unusually dependent on the American stock market right now — and the stock market is unusually dependent on artificial intelligence. This week, investors started to balk at the enormous spending hyperscalers are planning to keep building out the AI boom; Alphabet’s shares dropped 8% this week after it boosted its planned 2026 capital expenditure and signaled 2027 will be even bigger. If the data center boom started to slow down in earnest, then more than just that budget will change.

Speaking of which, my colleague Emily Pontecorvo wrote earlier this week about how many businesses are struggling to even estimate their carbon emissions from artificial intelligence. The carbon accounting startup Watershed recently unveiled a new formula to help companies get a sense of their AI-related emissions.

But even that formula is still limited by the amount of data hyperscalers publish — and they don’t publish that much. Google, for instance, is the only AI company that has (laudably) provided estimates of its emissions on a per-prompt basis. Yet no company has published its per-token emissions, or how emissions sync up with particular models or regions.

So Emily asked Google: Why aren’t you — or any other model provider — disclosing this kind of data yet?

The tech company didn’t get back to us until after we’d published Emily’s story. But its response was interesting enough that I wanted to quote some of it here.

The problem is “industry consensus,” Cooper Elsworth, a Google spokesperson, told us. “There is currently very little consensus on how to comprehensively and fairly measure the serving environmental impact of generative AI (such as text generation),” he wrote. “Without standardized, ‘apples-to-apples’ frameworks, it is difficult to compare different providers accurately.”

That’s partly because energy use — and emissions data — can vary from site to site and depend on “custom-built hardware, software compilers, and advanced inference techniques.” And he claimed Google doesn’t always have the measurement hardware in place to provide such specific estimates: “Providing precise, repeatable data requires highly advanced measurement infrastructure,” he said. “For example, software-based energy monitoring tools often suffer from sampling biases. For our study, we had to step away from top-down averages and directly measure actual energy at the physical power supply unit (PSU) level across our deployed fleet. Not all providers have the telemetry or data sets required to benchmark their operations at this level of granularity.”

Read Emily’s story to understand the other reasons why estimating — or even “guesstimating” — AI-related carbon emissions is so challenging.