For two years the story about AI capacity was told in chips. It is now told in megawatts, substations and permits.
The US Energy Information Administration expects electricity use to grow about 1% in 2026 and 3% in 2027, and notes that this would be the first time since 2007 that power demand has risen four years running — the strongest four-year stretch since 2000, driven primarily by large computing facilities.[1] On the supply side, Lawrence Berkeley National Laboratory counts roughly 8,200 projects actively seeking interconnection at the end of 2025, representing 1,312 GW of generation and about 749 GW of storage, with 549 GW already holding a draft or executed interconnection agreement but not yet in commercial operation.[2]
Every number a compute investment turns on — queue position, headroom, fuel mix, water, wage floors, export exposure, the rule that lands next quarter — is public. It is also scattered across a dozen federal registries that share no keys, no vocabulary and no shape, and each answer is assembled by hand, from scratch, every time.
Today we are opening a second market inside Knitify: AI Infrastructure & Energy.
Seven specialists, briefed like colleagues
You do not pick a database. You describe the decision, and the right specialists run — in the workspace or through the API — typically in two to eight minutes.
| Specialist | The question it takes | What comes back |
|---|---|---|
| DC Siting | Where can a large load physically land, and how contested is that market already? | Interconnection queue depth at the point of interconnection, nearby generation and grid headroom, permitted emission points, and county water availability — each line traceable to the filing it came from. |
| Grid Supply | Who actually supplies power here, and at what realised revenue? | The generator fleet by plant and technology, wholesale counterparties from quarterly FERC filings, and utility scale with sector sales split. |
| Clean Power | What zero-carbon supply can this load actually contract? | Plant fuel and reported emissions, federal offshore wind leases, the nuclear fleet, and who is already buying green power in that market. |
| Chip Supply Chain | Where does this semiconductor input physically come from, and who is exposed? | Mineral production sites and their operating companies with status, federal award history into the chain, and restricted-party screening hits. |
| AI Policy Watch | What rule is coming, when does it bite, and who is restricted? | Federal rules and notices with docket numbers, effective dates and comment deadlines, plus US and multilateral restricted-party designations. |
| AI Workforce | Where is the technical talent, what does it cost, and who is competing for it? | Attested wage floors from certified visa filings by employer, role and worksite; federally funded R&D; research output as a talent signal. |
| Physical AI Scout | What is the state of the art in robotics and embodied AI, and who is building it? | The research front with per-paper links, the trained-model landscape with compute and parameter counts, computed material properties, and the patent position. |
The generalists come with them. Deep research, company mapping, contact finding and US patent search work in this market too — they simply stop speaking in drugs and trials and start speaking in utilities, substations, dockets and process nodes.
Three views, because three different people are asking
The market splits into positions along the value chain, and each gets its own starting set of questions:
- Data Center & Power — “Where can I site a 500 MW data center with available grid capacity?” · “Interconnection queue position and timeline for a region” · “Which utilities have capacity headroom in the next three years?”
- Hardware & Supply — “Chip supply chain exposure for an accelerator programme” · “Who supplies the transformers and switchgear for this build?” · “Export-control exposure on a semiconductor supply chain”
- Investor & BD — “Competitive landscape for AI infrastructure operators” · “Diligence a counterparty’s build and operating record” · “Company landscape around a compute or power technology”
A fourth set covers policy and workforce — state incentives, siting rules, clean-power procurement risk, and whether the people to build and run the thing exist within commuting distance.
What sits behind an answer
Twenty-four datasets, more than 14 million records, all of them published by the body that collects them. Counts below were measured on 2 August 2026.
| Area | What is indexed | Records |
|---|---|---|
| Grid & interconnection | Interconnection queue requests (all seven ISOs/RTOs plus non-ISO utilities); annual and monthly generating-unit inventories; balancing authorities and control areas | 38,066 · 34,855 · 39,513 · 71 |
| Wholesale & retail power | Wholesale power contracts from quarterly FERC filings; utility rate plans; retail rates, sales and revenue by utility and customer class | 2,529,131 · 247,756 · 2,828 |
| Generation & emissions | Power plant profiles with fuel, capacity and reported emissions; point-source emission units; the nuclear fleet; federal offshore wind leases; green-power buyers | 12,612 · 2,834 · 96 · 52 · 492 |
| Sites, water & facilities | The federal environmental facility registry; county-level water withdrawals by use category | 5,300,149 · 3,223 |
| Minerals & materials | Mineral operations tied to commodity and company; mine master records; annual commodity statistics; DFT-computed inorganic materials | 6,785 · 91,933 · 8,886 · 93,723 |
| Policy & trade | Federal rules and notices, indexed passage by passage; the consolidated US restricted-party lists | 2,309,940 · 113,386 |
| Capital, talent & research | Federal award actions; certified labor condition applications; robotics research papers; notable trained models | 1,708,348 · 1,347,103 · 330,075 · 1,042 |
Reports cite the record behind each claim — the filing, the generating unit, the lease, the award, the passage of the rule — and link straight to it wherever the publisher issues a per-record link. Where a publisher issues none, the citation names the dataset and the record rather than inventing a URL.
An example of why the record matters more than the summary
Ask most tools how much generation is queued in a market and you get a headline gigawatt number. Here is what the queue actually looks like across the 38,066 interconnection requests we index, spanning request years 1995–2025:
| Status in the queue record | Records | Share |
|---|---|---|
| Withdrawn | 24,221 | 64% |
| Active | 8,513 | 22% |
| Operational | 4,654 | 12% |
| Suspended | 668 | 2% |
| Unknown | 10 | <1% |
Roughly two in three requests were withdrawn. Berkeley Lab puts the capacity-weighted version of that even more starkly: of the capacity that requested interconnection between 2000 and 2020, only 13% had reached commercial operation by the end of 2025, while 75% had been withdrawn.[2]
A queue position is a request, not a plant. That distinction decides whether a site memo is worth anything, and it is the kind of thing the specialists are built to say out loud rather than average away.
What this deliberately will not tell you
A diligence tool earns trust by refusing to answer past its evidence. Each of these is a rule the specialists hold to, not a caveat buried in a footer:
- A retail revenue figure is not a tariff. Reported utility revenue per MWh is a historical average across existing customers. It is not what a new large load would be quoted, and the reports say so rather than implying a price.
- Emissions mass is not emissions intensity. Where the source reports annual CO2 mass without a generation figure, intensity cannot be computed — so it is not stated, estimated or implied.
- Lease acreage is not capacity. An offshore wind lease records area. Turning acres into megawatts is a modelling assumption, and it is left unmade.
- A certified labor filing is permission to hire, not a hire — and the attested wage is a legally required floor, not a salary. Units vary across filings, so a raw comparison can rank an hourly rate below an annual one; the reports handle that explicitly.
- The structured coverage is US federal. These are US registries and US federal awards. They describe the domestic picture, and a report will not generalise them into a global supply map.
- Radar boards do not exist for this market yet. The curated monitoring boards are healthcare-built, so this market ships without them rather than with an empty catalogue pretending otherwise.
Getting in
Choose the market when you set up your workspace, or switch to it from the specialist panel; the roster, the starting prompts and the vocabulary change with it. Work in the workspace, or call the same specialists from the API and pull the results into whatever your team already uses. Access takes about a minute.
The healthcare and life-sciences market is unchanged, and remains the deeper of the two — a much larger specialist roster, curated Radar boards, years of tuning. The AI Infrastructure & Energy market is new, and we would rather say that than pretend the two are the same age.
See the AI infrastructure solution page for what a first week looks like.