John Bistline

Ed has a working paper version here: edrub.in/Papers/draft...

Downwind and out: The strategic dispersion of power plants and their pollution | Journal of the Association of Environmental and Resource Economists: Vol 0, No ja

This is from a "coal" new paper by John Morehouse and Edward Rubin in JAERE @jaereaere.bsky.social: www.journals.uchicago.edu/doi/10.1086/...

Wild stat: the "home" county often isn't the one breathing it.

Authors estimate that within ~6 hours, ~99% of coal emissions leave the source county.

For more, check out our Powering Intelligence white paper dashboard here: powering-intelligence.epri.com/dashboard/

Why this improves on older approaches:
• Trend extrapolations struggle when geography/techonlogies are shifting fast
• Chip shipment methods can be informative nationally, but don't tell you WHERE the load lands

Pipeline-by-state and explicit load mapping gets you closer to "grid-relevant truth."

Another state-level angle planners care about: We compare our implied peak increases to planning forecasts (FERC Form 714).

This helps answer how the scenarios compare with what utilities are planning for (regionally).

State-level results: Virginia is already in a different universe with its DC share >20% today, which could rise dramatically by 2030.

And the "next Virginias" aren't all the usual suspects; several new clusters emerge.

Sanity check/validation for credibility: We reconcile modeled DC demand with publicly reported state electricity sales, especially in high-concentration states.

If your estimate implies more load than the state's data can plausibly support... your estimate is the problem.

Conversion #3: utilization + load shape (peak vs. average)

Real-world facility data show:
• Data center load is relatively flat over the year
• But realized peak is meaningfully below "headline/nameplate," which means peak utilization often lands well under 100%

Conversion #2: ramp-up (the "hidden lag")

New sites don't show up as 100% utilized on day 1. We assume a staged ramp where only a fraction of nominal capacity is active initially, increasing over time.

This is why pipeline ≠ immediate peak.

Conversion #1: IT → nameplate (cooling & infrastructure)

That's where PUE comes in. Fleet average PUE is ~1.3 today, and new hyperscale builds can be lower. This changes annual TWh and peak MW.

Now the part many forecasts skip:
Announced MW ≠ peak MW

You have to translate "nominal IT capacity" into what the grid actually sees.

We do that with an explicit mapping:
IT capacity → nameplate → active (ramp) → realized peak + annual energy.

Step 1 of the method: build the pipeline.

We sort planned projects into:
1. Under construction
2. Announced (advanced)
3. Announced (early)

Then we ask: What fraction of each bucket becomes real by 2030? These three scenarios help to bracket uncertainty.

Why "state-level" matters: a national trendline can look continuous while a handful of states are doing something... discontinuous.

So we built scenarios from state-level operational capacity, what's under construction, what's been announced.

First, the headline results: we estimate U.S. data centers at ~4-5% of electricity today, rising to ~9-17% by 2030 (scenario range). That's the national story.

But the real action is local. States ≠ average.

Modeling Monday 🧵

Data centers are the new load forecasting problem

So we updated EPRI's Powering Intelligence methodology to move beyond trend extrapolations and "chip shipment math" and into something planners can actually use: a state-level project pipeline to grid load mapping

Here's a link to the report: powering-intelligence.epri.com

❌ "We know how much electricity data centers use"

There is currently little public reporting of DC electricity. We cross-checked our estimates against state C&I sales data and found they couldn't be much higher and still be consistent with observed totals.

Data transparency is a major gap.

❌ "Data centers run at full capacity all the time"

Our facility data shows peak utilization of 62-80% of nameplate. And during rapid buildout, fleet average utilization can actually drop, since new capacity takes years to ramp up.

❌ "AI is the whole story"

AI workloads are estimated at 15-25% of DC electricity today. Growing fast, yes. But non-AI data center demand (streaming, cloud, enterprise) continues steady growth, too. The base is bigger than people think.

❌ "Data centers announced X GW, so the grid needs X GW of new capacity"

Announced nominal MW is a pipeline indicator, not a peak forecast. Non-IT loads, ramp-up schedules, load shapes, and onsite generation all mean realized peak is lower than the headline.

Now that our updated data center electricity analysis is out, a few things I see people often get wrong about DC load growth:

Co-authored with Geoff Blanford, Tom Wilson, and Nils Johnson. This builds on the 2024 edition that was EPRI's most-downloaded deliverable.

Happy to discuss what we found, plenty more in the full paper and site: powering-intelligence.epri.com

We also compared DC load to EV charging load. Data centers: ~184 TWh in 2024. Light-duty EV charging: ~12 TWh.

Under reference projections, EVs don't surpass DC load nationally until the mid-to-late 2030s. In heavy DC states, maybe never.

One finding I think deserves more attention: how the generation mix to serve this load depends entirely on the policy environment.

Under current policies → natural gas dominates.
Under 24/7 carbon-free energy targets → renewables, storage, nuclear.

And some states barely on the radar are emerging fast (Indiana, Louisiana, Mississippi, Ohio) driven by land availability and power access for large AI training facilities.

Meanwhile, CA and NY have significant existing capacity but relatively little projected growth.

The state-level story is where it gets really interesting.

Virginia is currently the only state where DCs exceed 20% of electricity. By 2030, seven more states could join that club: Oregon, Iowa, Nebraska, Nevada, Wyoming, Arizona, Indiana.

But a critical nuance that gets lost in headlines: announced MW ≠ peak demand.

We mapped the chain from nominal IT capacity → nameplate (add cooling/PUE) → active capacity (ramp-up lags) → realized peak load.

The peak is often meaningfully below the headline number.

The past 18 months saw record levels of data center development activity. The pipeline of projects under construction or in advanced planning grew dramatically.

New from EPRI: Our updated Powering Intelligence analysis of U.S. data center electricity demand.

Data centers could consume 9-17% of U.S. electricity by 2030, up from 4-5% today.

Our projections are ~60% higher than last year's. Here's why. 🧵