The Nieman Lab feature frames the project as necessary infrastructure journalism: no US agency maintains an authoritative registry of data center locations, capacities, or water draws, so a volunteer network is building one. The crowdsourced map aggregates permit filings, utility interconnection requests, and local reports into a single searchable layer that regulators have failed to provide.
By submitting the Nieman Lab piece to HN, cratermoon surfaces the project to a technical audience that understands hyperscaler infrastructure. The 227-point score signals broad agreement that the absence of a federal registry is itself the story.
Brockovich frames data centers as 'the new Hinkley' — a diffuse, technically-justified industrial buildout whose costs (aquifer depletion in semi-arid basins, delayed coal plant retirements, residential rate hikes quietly approved by PUCs) are borne by people who never consented. Her map exists to make those spatial concentrations visible and politically actionable.
The editorial argues the noteworthy angle isn't the activism itself but how it's structured: a distributed, crowdsourced data layer assembled from permit records, interconnection queues, and citizen submissions is doing the work that institutional regulators have refused to do. For a developer audience, the project is a case study in open-data infrastructure substituting for missing public registries.
Erin Brockovich — yes, that Erin Brockovich, the environmental advocate whose PG&E fight became a Julia Roberts movie — has teamed up with journalist Karen Hao and the nonprofit Data Center Watch to publish a public, crowdsourced map of every known data center in the United States. The map, hosted on Brockovich's site and detailed in a Nieman Lab feature, pulls together permit filings, utility interconnection requests, local news reports, and citizen submissions into a single searchable layer. The premise is blunt: no US government agency maintains an authoritative registry of data center locations, capacities, or water draws, so a 65-year-old activist and a small volunteer network are doing it instead.
The numbers driving the project are not subtle. The Department of Energy's December 2024 Lawrence Berkeley National Laboratory report projected that data centers will consume between 6.7% and 12% of total US electricity by 2028, up from roughly 4.4% in 2023. Goldman Sachs put global data center power demand growth at 165% by 2030. In specific markets, the load is already grotesque: data centers consume about 26% of Virginia's electricity, and Dominion Energy's interconnection queue has more pending hyperscale projects than the rest of the country combined.
What the map surfaces — and what the hyperscalers don't want surfaced — is the spatial concentration. Clusters in Loudoun County, central Ohio, Phoenix, Atlanta, and the Texas triangle are pulling water from aquifers in semi-arid basins, drawing on coal plants that were scheduled for retirement, and triggering residential rate hikes that local PUCs are quietly approving. Brockovich's framing in the Nieman piece is that this is the new Hinkley: a diffuse, technically-justified industrial buildout whose externalities are borne by people who never opted in.
The interesting thing for our audience is not the activism. It's the information architecture of the activism. Brockovich and Hao have correctly identified that the bottleneck to public accountability for AI infrastructure is not policy — it's a database that doesn't exist. Hyperscalers file under shell LLCs ("Project Rosie," "QTS Mountain View"), sign NDAs with municipalities for tax abatements, and route power purchase agreements through utility filings that nobody reads. The result is that even reporters covering this beat full-time can't tell you, with confidence, how many active data centers operate in a given state, let alone their MW capacity or cooling architecture.
That's a solvable problem, and the map is the proof-of-concept. It's effectively an OSINT project: scraped FERC interconnection queues, state environmental impact statements, county zoning variances, and Glassdoor reviews that accidentally name a site. Anyone who has built a data pipeline knows this is tractable — it's just nobody had the incentive to fund it.
The counter-argument from the industry is that disclosure creates a security risk. A public map of high-value compute targets is, in fairness, a target list. Microsoft, Google, and AWS have all argued in zoning hearings that even capacity figures should be redacted on national-security grounds. The map's contributors push back that physical address obscurity is performative — drone footage, OpenStreetMap, and FAA filings already expose every site that matters; the secrecy is about avoiding rate-case scrutiny, not foiling sabotage.
There's also a real epistemic problem the map confronts: capacity numbers in press releases are theater. A "500 MW campus" might be a single building drawing 80 MW with three more phases that may or may not break ground. The map's volunteers are doing the unglamorous work of distinguishing IT load from utility-side nameplate from "announced" from "energized." This is the kind of data hygiene that frontier-model training reports gesture at and never deliver.
Community reaction on HN was telling. The top comment thread wasn't about Brockovich — it was about whether the crowdsourcing methodology could be poisoned by hyperscaler PR teams submitting bad data, and whether GitHub or Sourcegraph-style provenance attestations could be bolted on. That's the right question. A map is only useful if its corrections are auditable.
If you're building AI features at any scale, three things change. First, "the cloud" is about to stop being an abstraction in your architecture diagrams and start being a specific zip code that a state legislator is angry about. Expect carbon-intensity APIs (Electricity Maps, WattTime) to become a routine input to scheduling decisions — and expect procurement to start asking your team for region-by-region compute footprint reports. Pick regions now where the underlying grid story is defensible, because retrofitting that argument after a contract is signed is expensive.
Second, the inference-vs-training distinction is going to matter politically in ways it hasn't mattered technically. Training is bursty, geographically flexible, and easy to defend as "R&D." Inference is sticky, latency-bound, and a 24/7 baseload draw on whatever grid you've parked next to a population center. If your product is inference-heavy — which most production AI features are — you're the one on the local-news chyron, not OpenAI. Start building the disclosure muscle now: a public methodology page, a quarterly footprint report, a real number for tokens-per-kWh.
Third, this is a hiring signal. The companies that will navigate the next three years cleanly are the ones with at least one person on staff who can read a FERC docket. That used to be a niche utility-industry skill. It's now adjacent to your platform team, and the comp band hasn't caught up yet.
The map will be wrong in a hundred ways at launch and better in six months, because that's how OSINT works. The structural shift it represents — that the most credible registry of AI infrastructure in the US is being built by an environmental lawyer and a volunteer crew, not by the DOE or the hyperscalers themselves — is the part worth watching. Regulation follows visibility. Visibility now exists. The bill, in both senses, is coming.
The Erin Brockovich page itself repeats the canard, on the front page, that these sites endanger ecosystems with their water consumption.
It’s interesting how many more community reported data centres there are compared to operational and proposed. I’m wondering if this is because of over reporting? Like - does the public mistake any new, big building as a data centre, or are the other categories under reported (or something else)?
This datacenter stuff is such populist brainrot.
Direct Link: https://www.brockovichdatacenter.com/
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
Why go through the effort when such work has already been done?https://www.datacentermap.com/datacenters/Not being negative. But isn’t there existing highly reliable data that already exists for this?