Physical-Economy Intelligence · Interactive
The Power Draw
The largest industrial buildout in a generation is nearly invisible in public data. From orbit, it isn't.
estimated new AI capacity online across five megasites — measured from open satellite imagery
Press releases slip, interconnection queues are unreliable, and utilities don't disclose — but you cannot hide a gigawatt of construction from a satellite. This reads five AI megasites from open Sentinel-2 imagery and turns new building footprint into the one signal several desks trade on: how much power is actually coming online, and when.
Power & gas traders
Energy
“Where and when does new load hit the grid?”
Regional power prices
US data-center load: 31 → 66 GW by 2027
Utilities & grid operators
Energy · Infra
“Is this interconnection request actually being built?”
Capacity planning
A wrong forecast = stranded capex or a shortfall
Infra PE & data-center REITs
Infrastructure
“Which sites are real and on schedule?”
Deployment & deal intel
≈ $45–55B in capex per gigawatt of campus
Equity analysts
Financial services
“Does hyperscaler capex match the press release?”
The earnings read
≈ $725B of hyperscaler capex in 2026
Five sites↓
How it's measured
Measured, not asserted
For each site, the least-cloudy Sentinel-2 scene per quarter (2020–2026) is pulled from the Microsoft Planetary Computer and reprojected onto one fixed grid so the time-lapse registers pixel-for-pixel.
“Built” is bright + low-vegetation + spectrally neutral (man-made roofs and concrete read grey-to-white in blue light, where soil and desert do not), filtered to large contiguous structures. Subtracting each site's 2020 footprint isolates new construction; a rolling median plus a build-to-date floor turn it into a clean trajectory.
Footprint (hectares) is the measured quantity. Megawatts are a deliberately conservative estimate — roughly 1 MW per hectare of new build — shown against published capacity below. It is an order-of-magnitude read, not a meter.
Honest limitHonest limit: at 10 m, a new data hall and a new warehouse look alike. That ambiguity is precisely what Planet's 3 m detail and embeddings resolve — the limitation is the argument for the commercial-grade layer.
Validation
Estimate vs. the public record
Built-to-date footprint converted at ~1 MW/ha, against each site's announced capacity. The estimate sits below the announced figures by design — it counts what's on the ground today, not the ultimate buildout.
What it's worth
The signal reads against real money
These demos run on free imagery, but they index decisions measured in the hundreds of billions. The figures below are illustrative estimates from public sources — the point is the order of magnitude, and that satellites get there first.
≈ $13B
construction capex implied by the ~1.1 GW measured here (≈ $40B all-in with compute)
≈ $725B
hyperscaler capex in 2026 — the number this signal helps verify
31 → 66 GW
US data-center power demand, 2025 → 2027 (Goldman Sachs)
≈ $15B / yr
what hedge funds already spend on alternative data
Sources: data-center build cost ≈ $10–12M/MW facility, $30–40M/MW all-in (industry benchmarks, Epoch AI); hyperscaler 2026 capex ≈ $725B (company guidance); US data-center power demand (Goldman Sachs Research); alternative-data spend (industry estimates). Estimates, not measurements.
The engine · embeddings + semantic search
Search Earth by example
The footprint detector behind the sites above is classical image processing. This is the AI layer the role asks for, running for real: every Sentinel-2 tile across Phoenix's West Valley embedded with RemoteCLIP — a satellite-tuned vision-language model — so the imagery is searchable by meaning. Click a tile to find the places most like it (image-to-image), or pick a concept to rank every tile by it (text-to-image). Change detection, image classification, and embeddings for semantic query — the geospatial-AI stack, demonstrated on open data.
Honest readHonest read: distinctive land uses — golf courses, airports, stadiums — rank cleanly. Data centers and warehouses surface together, because at 10 m they're near-identical big flat roofs. Separating them, and running this across the entire daily global archive, is exactly what Planet's 3 m imagery and production embeddings infrastructure are built for.
Open engine → commercial-grade
Where Planet takes this
Everything above runs on free, open Sentinel-2 at 10 m. That's enough to prove the signal exists — and exactly where it hits a wall is where Planet's product suite takes over.
The standout is the operational beat: Land Surface Temperature flips a campus from “built” to “drawing power” — a signal open optical data structurally cannot produce.
Imagery: Sentinel-2 © Copernicus/ESA and Landsat (USGS/NASA, public domain), via the Microsoft Planetary Computer. Planet products referenced for the commercial-grade layer; the public engine uses open data only.
Find the signal. Build the proof.
Every number here came from free, public satellite imagery and a few hundred lines of Python — built ahead of any product investment, the way you find a signal before committing real engineering. The method runs at far higher fidelity on Planet's daily 3 m archive, which is exactly what makes it commercial-grade.
Open data proves the signal exists. Planet is the version a desk pays for.
The narrative companion: Ground Truth →