Infrabench / Where the power is
Tool 03 — Power-constrained siting

The chips are the easy part. Where is the power?

The US interconnection queue holds more capacity than the entire installed grid, and most of it will never be built. Meanwhile a new hall lives or dies on one date: the day the utility energizes it. Put in the megawatts you need and the year you need them, get the regions that can actually deliver, ranked by the earliest realistic window. Every constant is in the method.

The screener

While you wait
I need MW of new grid capacity, energized by .
While the queue moves, I .

The answer
Assumptions
MW per extra study year

Every megawatt past the first 200 makes the transmission study bigger and the network upgrades longer. One extra year per 250 MW is a planning heuristic read off the gap between published timelines for 100 MW and gigawatt-class requests.

years minimum, bridged

On-site generation is not instant power. An air permit, a substation and the gas turbine order book set the floor: majors are quoting deliveries years out, so two years from decision to first fire is aggressive but done.

Earliest realistic energization window for each region
The regions, ranked
Reality check
Why the tool exists

The site selection question inverted

Ten years ago you picked land and ordered power. Now you pick power and the land follows, because the grid, not the building, is the long pole on every schedule.

01

The queue is not a pipeline

Over 2,000 gigawatts sit in US interconnection queues, more than the entire installed grid, and historically only about one queued megawatt in eight reaches commercial operation. A queue position is a lottery ticket, not a delivery date.

02

Time-to-power beat price-per-kilowatt

PJM's capacity price rose ninefold in one auction and hit the regulatory price cap in the next, with data center load named as the reason. Expensive power hurts; power that arrives three years late kills the project. The date is the number to shop.

03

Loads queue like generators now

A few hundred megawatts of demand gets studied the way a power plant does. ERCOT's large-load queue hit 239 gigawatts of requests, three quarters of it data centers, of which about 7.5 have actually connected. You are in a queue whether you like it or not. The question is which one.

GW
Active in US interconnection queues
Of queued capacity ever gets built
YRS
Median wait, request to operation
How it works

Four steps from an ask to a shortlist

Each region is carried through the same four steps. You can do the arithmetic yourself, and the two constants it leans on are editable in the assumptions panel.

Step 01 / The published band

Start from what each region admits to

Every region has a band of years between a large-load request and energization, assembled from utility filings, regulator dockets and trade press. ERCOT connects loads in 2 to 4 years; a constrained coastal market can take 6 to 8. The band is the tool's starting point and its biggest judgement call, so it is printed in full in the table below.

Step 02 / Size

Scale for the ask

A 100 MW request rides existing headroom. A 700 MW request triggers network upgrades that get studied, procured and built. The tool adds one year per 250 MW beyond the first 200, applied to both ends of the band.

Step 03 / The bridge

Credit on-site generation where it is legal and practical

In regions that allow large behind-the-meter builds, on-site gas turbines can carry the load while the grid catches up. Bridging takes 1.5 years off the wait in those regions, floored at two years, because turbines have their own queue: the order book.

Step 04 / The verdict

Check the window against your deadline

The clock starts mid-2026. Each region gets a window, earliest to latest realistic energization. If the whole window lands before your year, the region makes it. If your year falls inside the window, it could slip. If even the optimistic end is late, it misses, and no amount of negotiation fixes that.

Worked example, the default settings
300 MW, needed by 2030, waiting on the utility, screened against ERCOT
ERCOT's published large-load band = 2 to 4 years
size surcharge: (300 − 200) / 250 = +0.4 years of study
adjusted band = 2.4 to 4.4 years from mid-2026
window = 2028 to 2030, and 2030 ≤ the deadline, so ERCOT makes it
the same arithmetic puts PJM at 2030 to 2033: could slip
and CAISO at 2031 to 2034: misses, before politics is even priced in
The surprise is not that Texas wins. It is the spread: for the same 300 MW, the gap between the fastest region and the slowest is longer than the entire construction schedule of the building.
The regions

Ten grids, ten different decades

The lead-time bands the screener starts from, with the queue depth, grid carbon and industrial power price behind each one. The bands are the model's biggest judgement call, so they are printed here in full.

The ten regions the screener ranks, with large-load lead-time bands, active interconnection queue capacity, average grid carbon intensity, industrial electricity price, and whether on-site bridging is practical
RegionLead timeQueueGrid CO2Power priceBridgeNotes
The method

Every constant, in the open

This is the whole model. If you disagree with a constant, you know exactly which one to argue with, and the two that matter most are editable on the page.

The model6 equations
Lead bandper region, published years from large-load request to energization
Size surchargemax(0, MW − 200) / 250, in years, added to both ends
Bridge credit−1.5 years in bridge-friendly regions, never below the 2-year floor
Windowmid-2026 + adjusted band, floored to calendar years
Verdictmakes it if the window closes by the deadline; could slip if it straddles; misses if it opens after
Rankmidpoint of the window, then carbon or price if you re-sort
The judgement calls6 notes
  • The lead bands are the model. They are assembled from utility integrated resource plans, regulator dockets and trade press, not from a single published dataset, because no such dataset exists. That is why they are printed in full.
  • The clock starts mid-2026. Windows are floored to the calendar year they land in: a mid-2028 energization prints as 2028.
  • Queue depth is a congestion proxy. The gigawatts are generator requests, not loads, but they occupy the same study engineers and the same transmission headroom you need. The two Western regions share LBNL's single 567 GW non-ISO pool, and figures marked ~ are read off public trackers rather than the report.
  • Carbon and price are anchors, not accounting. Carbon is the eGRID 2023 annual average for the subregion where you would actually site, not the marginal unit your load turns on, which is almost always gas. Price is the EIA 2023 industrial average for the region's anchor state.
  • Bridging means gas. The bridge credit assumes reciprocating engines or turbines you own, permitted on site. Solar and batteries shave peaks; they do not carry a 24/7 hall through a winter.
  • A screening heuristic, not a siting study. The output is which three conversations to start, not which contract to sign. Nothing here replaces a utility letter with a date on it.
Built by
Adrian Mucha

Product designer for the software most designers avoid: operator consoles where a wrong click costs real money.

Focus
Data centers · AI tooling
Base
Los Angeles, CA
Status
Open to new work
Copied