How this notebook is drawn

The historical map: EIA-923

Every state’s hachure mix comes from Form EIA-923, the U.S. Energy Information Administration’s monthly survey of power plants. Utilities and generators fill it out and mail it (electronically) to the federal government; EIA compiles, checks, and publishes it roughly one to two months later. That’s why the atlas is always stamped SURVEYED · <month> rather than showing today’s date — the ink genuinely hasn’t dried yet on the most recent month by the time you’re reading this.

“Other” is a catch-all for every fuel outside our seven headline categories (petroleum, waste heat, a handful of small or unclassified sources) plus one deliberate quirk: pumped-storage hydro reports as negative generation in the months it’s pumping water uphill (consuming power) more than it’s generating. We keep that true, signed value in our own records, but never show a bar with negative length — the displayed share for “Other” is clamped at zero, so a pumping-heavy month reads as “none of this,” not as ink drawn backwards.

One thing this map is not: a picture of where anyone’s electricity comes from. EIA-923 counts generation produced inside a state’s borders, and electricity does not stop at them. A state that generates far more than it uses is selling the difference to its neighbours; one that generates far less is buying. Washington D.C. is the extreme case — it makes almost nothing locally and imports nearly all of its power from the PJM grid — but every state on this map is doing some version of the same thing, and none of that trade is visible here. Read a state’s mix as what its power plants burned and spun, never as what came out of its sockets.

One more honest gap: small, rooftop/distributed solar (what EIA calls “DPV”) is folded into the same “solar” bucket as utility-scale solar farms, and EIA itself estimates distributed solar rather than metering every rooftop directly — so the solar share you see is a mix of a hard count and a modeled estimate, not two separately verifiable numbers. Read the source data yourself at eia.gov/electricity/data/eia923.

The long-run chart: what it is, and isn’t

The Trends pages draw twelve years of monthly generation, one line or band per fuel. Everything on those charts is surveyed history — things that already happened, taken from the same Form EIA-923 survey described above. There is no model in it, no forecast, no smoothing that invents a number: each point is a month EIA published. That makes it a different animal from the year slider on the Atlas page, which is explicitly a projection. If a chart on this site reaches into the future, it says so in stamp-red; if it doesn’t say so, it’s the record.

One consequence worth knowing: the most recent month or two is always missing, for the same reason the atlas is. The survey takes time to collect and check. The line stops where the data stops rather than being extended to today.

Why the line starts in 2014

EIA’s monthly state-level generation data actually goes back to 2001, so starting in 2014 is a deliberate choice, not a limit of the source. Here’s the reason.

Solar power arrives on the grid in two quite different ways: big utility-scale solar farms, which are metered like any other power plant, and small rooftop and “behind-the-meter” solar on houses, barns and warehouses, which mostly isn’t metered by anyone who reports to the federal government. EIA estimates that second category (it calls it DPV, for distributed photovoltaics) — but it only started estimating it monthly, state by state, with the December 2015 Electric Power Monthly, backdated to the beginning of 2014.

So if we drew the line from 2001, solar would appear to jump sharply in January 2014 — not because anything happened on the grid that month, but because that’s the month EIA started counting rooftops. That step would be a measurement artifact masquerading as a trend, on the very series people most want to read a trend into. Rather than draw it and footnote it, we start the chart where the definition stops changing. Every fuel, solar included, is then measured the same way from end to end, with no asterisk.

The flip side: the solar line here includes rooftop solar as well as solar farms, added together. That’s the honest total of what the sun contributed, but it does mean part of the solar line is a careful federal estimate rather than a meter reading — the same caveat as on the map above.

What’s in “Other,” and why it can look strange

We ask EIA for six named fuels — gas, coal, nuclear, wind, solar, hydro — and, separately, for its own figure for all generation. “Other” is then simply the difference: everything EIA counted, minus the six we named. We derive it by subtraction rather than by adding up the leftover categories ourselves, because EIA’s own list mixes individual fuels with overlapping subtotals, and adding those up double-counts. Subtracting from EIA’s grand total can’t.

That means “Other” genuinely is a remainder: petroleum, geothermal, biomass, waste heat, and anything EIA classifies outside our six. It also picks up pumped-storage hydro, which is a reservoir used as a battery — water pumped uphill when power is cheap and released downhill when it’s wanted. In a month when more power went into the hill than came out of it, that shows up in the survey as negative generation, and it can drag the whole “Other” remainder below zero. We keep the true, signed number in our records, but we never draw a band with negative height — the displayed value is clamped at zero, so such a month reads as “none of this,” not as ink drawn backwards. If “Other” looks flatter than you’d expect in some months, that clamp is usually why.

The national line is the sum of the fifty states plus the District of Columbia, added up from the same rows the state charts use — not a separately-published national figure. That’s deliberate: it means the national view and the state views can never quietly disagree with each other.

The drafting-ruler slider: projections

Dragging the year slider past 2026 doesn’t show a forecast of demand, weather, or prices — it draws out the arithmetic consequence of already-announced power plants. Every month, Form EIA-860M lists plants that say they’ll come online or retire by a given year. We start from a state’s real trailing-12-month generation, add each announced addition once its year arrives (megawatts × 8,760 hours × a typical capacity factor × a completion discount, below), subtract each announced retirement the same way (undiscounted — retirements overwhelmingly happen, often early), and never let a fuel’s total fall below zero. That’s the whole model. We stop at 2031 on purpose — six years is already stretching what “announced” can responsibly tell you.

This is why every projected year is stamped DRAFTED · FROM ANNOUNCED PROJECTS in the same stamp-red as the historical stamp: the color says “this is the honesty mechanism,” the words say which kind of honesty applies. If an announced project is later withdrawn or quietly drops off EIA’s list, it disappears from the next run — we never delete our own record of it, it just stops contributing.

The completion-discount table

Not every announced project gets built on schedule, or at all — interconnection queues are congested and projects slip or drop out constantly. Each addition’s contribution is discounted by how far along its EIA-860M status code says it is; an unrecognized status code is treated conservatively, as the lowest-confidence tier.

Typical capacity factors

A plant’s megawatt rating isn’t how much it actually generates — a capacity factor is the fraction of the year a technology typically runs at that rating. These are national averages, applied the same way in every state:

A known v1 limitation: every lookup above resolves to the national default today, not a state-specific number. A real state-specific capacity factor would need to know a state’s already-operating fleet capacity by technology, which this site doesn’t yet ingest (only planned additions/retirements from EIA-860M, and fuel-level, not technology-level, generation history) — so a Great Plains wind farm and a Southeast wind farm share the same 0.35 today, even though their real-world output differs meaningfully. The lookup already prefers a state-specific row the moment one exists; this is an honest placeholder, not a bug, and it self-corrects with no code change once that data is seeded.

Battery storage projects appear in each state’s “on the drawing board” list but never move a fuel’s projected generation — a battery stores power generated (and already counted) elsewhere, so including it would double-count. We compute a second, more optimistic scenario internally — every addition at 100% likelihood — and we never chart it. A drawn line gets screenshotted without its caption; a plotted optimistic future would travel as a forecast we don’t stand behind.

What we do publish is its size. Each state’s drawing board carries one sentence naming how much announced generation the board holds by 2031 and how much of it we actually count. Withholding that would leave the most consequential assumption on this site invisible: the discount is the single biggest thing standing between what has been announced and what we draw, and a reader is owed its magnitude. It is published as two numbers in a sentence and never as a second line on a chart — the difference matters, because a sentence cannot be replotted.

The board: announced, permitted, under construction, testing, standing

Each state’s drawing board, and the national board, sort announced generators by how far along they are. The four stages come straight from the status code EIA-860M gives each generator, and they are the same four marks the little drawn pie beside each project row has always used: announced (nothing under way yet), permitted (regulatory approvals in hand), under construction, and testing. A fifth state, standing, means the machine got built and now appears in the operating inventory.

Four stages displayed, five likelihoods used. EIA distinguishes construction that is under half complete from construction that is further along, and our completion-derating table above treats them differently (0.85 against 0.92). The board shows them as one bucket. Splitting them would present a reader with five stages and imply the two construction buckets mean different things to them, which they don’t — the difference is a weighting we apply, not a fact about the plant.

The board is measured in installed megawatts, not generation. Every other figure on this site is gigawatt-hours — electricity actually produced. A megawatt is a machine’s size, and a megawatt of solar and a megawatt of gas do not produce the same electricity in a year. The two never share a total, an axis, or a sentence. Batteries are counted on the board and in the standing inventory, because both are counts of installed machines; they still never move a fuel’s generation, for the reason given above.

What the board cannot tell you, and won’t pretend to. When a generator disappears from EIA’s planned sheet, there are two possible reasons and the workbook does not say which: it got built, or it stopped being planned. We can identify the first, because a generator that got built turns up in the operating inventory afterwards under the same identifier — that is what a “standing” row is. We cannot identify the second. So nothing on this site reports a cancellation, an attrition rate, or a share of announced projects that never happen, and a project that has left the record is described as exactly that and nothing more.

We also can’t yet tell you what has slipped. Until August 2026 each weekly reading overwrote the last, so a project’s earlier promised dates were destroyed as they changed and no history of them survives. We started keeping every reading on 12 August 2026. Once enough of them have accumulated — roughly six months — “this plant has been promised three times” becomes a sentence we can write. Until then it is one we can’t, and the rows built since we started watching are labelled with that window rather than implying we have watched longer.

The clippings: how sources are chosen and weighed

The Clippings page is a reading board, not a publication. Every headline links out to whoever wrote it; we store nothing but the title, the feed’s own one-line description, and a link. We never fetch or reproduce an article’s body.

We carry roughly three dozen feeds, chosen to cover the whole system rather than the part of it that gets written about most: natural gas, nuclear, hydro, coal, wholesale markets and grid operations, and independent analysis, alongside renewables. Renewables still lead the feed most days, and should — but on the strength of what’s being reported, not on how many posts an outlet publishes per day.

Every source carries a stated lens

Each publication is labeled with what it is organized to do — official for statistical agencies and regulators, analysis, academic, trade, independent, industry, or advocacy — and that label is printed on every card next to the source name. It is a description, not a grade. An advocacy outlet is often right and a wire service is sometimes wrong; the lens tells you the vantage point a piece was written from, which is something you would want to know either way.

The lens also sets a weight in the default ordering. Statistical agencies and research shops rank a little above trade press; advocacy and industry outlets rank a little below. The effect is bounded and worth stating precisely rather than leaving vague: across the full range of weights, it is worth about two days. A regulator’s filing will sit above a same-day advocacy post for roughly that long, and after that recency wins regardless of who published what. No source can hold the top of the feed.

What the AI does, and what it does not

A language model reads each headline and its one-line description — never the article — and writes a two-sentence summary, corrects the fuel and state tags, and rates two things: how far the headline oversells what the text supports, and how many concrete specifics the item contains (figures, named projects, filings, dates). It also flags press releases, opinion, and sponsored content. Where a summary appears on a card it is marked in brief and is our writing, not the publisher’s.

These ratings are a model’s opinion, not a measurement. They are not a fact-check and they are not a verdict on whether a claim is true or whether a position is correct — a model is not competent to referee energy policy and we do not ask it to. It is scoring presentation: whether a headline is writing cheques the text can’t cash. It gets that wrong sometimes.

Which is why the ratings only ever reorder the feed. Nothing is hidden, filtered, or removed — a sponsored post ranks last and is still there, labeled. An article the model hasn’t looked at yet ranks as an average one, never as a bad one. And the Newest toggle at the top of the Clippings page returns the raw chronology, with no weighting of any kind, at any time. The ordering is an editorial act, so it is one you can undo.

What this notebook is not

• Not real-time — no live grid data, no “grid stress” theater. The stamp is the honesty mechanism.
• Not a GIS product — no county drill-down, no plant-level dots.
• Not a demand, dispatch, price, or weather model — see the projections section above.
• Not AI-written where the data is concerned — every number, every hachure, and the surveyor’s summary lines are deterministic rules over EIA data, with no model anywhere near them. A model touches only the news clippings, as described above.
• Not a publisher — we link out to every article and reproduce none of them.

Privacy

No accounts, no login, no cookies, and no tracking of any kind run on this site today — every page you’ve loaded so far came from us with nothing set aside to identify you or your visit. If we ever add analytics, it will be a privacy-friendly, cookieless page-count only (no fingerprinting, no cross-site tracking, no third-party ad or analytics scripts) — a commitment, not just a current fact.