How to read Silicon Almanac.

• Start here
• Reading a figure /record/money.html
• The overview /
• The map /map/
• Fabs and sites /fabs/
• Compute and equipment /ai/
• Energy /energy/
• Filings /filings/
• Feeds /feeds/
• The record /record/
• Stagger charts /record/stagger.html
• The revisions table /record/revisions.html
• What the site withholds, and why /method/

About this help

Help is on by default and can be turned off with the button in the bar above. The setting travels in the address bar and nothing is stored about you.

Every entry explains the instrument: what an axis is, where a number came from, what a shape means, why something is absent. No entry says what any of it means for anyone, and each one states what it deliberately leaves out.

What this does not tell you: How the industry's cost and capacity structure works, and how to build monitoring like this from public sources. Those are a different piece of work.

Why these rules exist

The left column is a record: it states who published what, and when. The right column is a rating: it adds an assessment of that publisher.

The distinction is not about politeness. A site that states no view is not making an investment recommendation. A site that characterizes a forecaster's revisions is doing something else entirely.

The rule that follows for every chart: the vertical axis already carries direction, so nothing else may.

What this does not tell you: Nothing is withheld here. This section is the rule set itself.

Using the converter

Enter an amount, its scale and its currency, and the date the figure carries. The date matters: the conversion uses the exchange rate that date carried, not today's rate.

The ladder underneath shows the same amount in the other currencies this industry reports in.

The sentence below the result names the rate, the date of the rate and the publisher. If the rate used is not from the date you asked for, it says which date it is from and why.

Some dates return no result. That means no reference rate was published for that currency on or near that date, and no rate was substituted for it.

What this does not tell you: Any judgement about which currency to think in, or what a conversion implies about a company or a country.

What this chart shows

One unit of each currency, expressed in euro, day by day across the period the exchange-rate file covers.

Each currency is indexed to 100 on its own first covered day, because the raw rates differ by four orders of magnitude and one scale per chart is a rule this site keeps. Indexing is arithmetic on published rates and is marked derived.

The percentage in the legend is the change across the covered span for that currency, computed from the two end points.

What this does not tell you: Whether a currency is strong or weak, where a rate is heading, or what a movement means for anyone holding it.

Why the conversion rules are written this way

A figure published in one year converted at another year's rate is a number nobody ever held. Converting at the figure's own date is the default for that reason.

The single-rate setting applies one rate to every period. It exists because holding the currency still is the only way to see how much of a movement was the currency rather than the underlying figure. It is always labelled derived.

A cross-rate is a rate no single publisher printed. The Taiwan dollar figures here are chained through the US dollar using two publishers, both named on every figure.

What this does not tell you: Advice on which setting to use for which purpose. Both are stated; the choice is the reader's.

What the tiles are

Each tile is one official series, with its current value, the change against the same month a year earlier, and a small chart of the recent path.

The publisher is named under every tile and links to the page the figure came from. Where a figure is computed here rather than published, the tile is marked derived.

A tile marked stale means the last fetch failed. The value shown is the last good one, with the date it was good.

What this does not tell you: Which series matters most, how the tiles relate to one another, and what any movement means for the industry.

The value-chain ribbon

The ribbon runs left to right in the order material moves: design tools and intellectual property, then chip designers, then manufacturing, then assembly, test and packaging, then the equipment and materials that supply all of it.

Each block links to that layer on the map page, filtered to it.

Position in the ribbon is sequence, not size. A wide block is not a large market.

What this does not tell you: Where value or margin sits along the chain, and which layer is a good place to be.

The section cards

Each card is a route into one of the other pages, with a sentence on what that page holds and a count of the records behind it.

The counts come from the same files the pages themselves read, so a card and its page cannot disagree.

What this does not tell you: Which page to read first for any particular purpose.

Where the headlines come from

The most recent items from the feeds listed on the Feeds page, newest first, in the publisher's own words.

Nothing is selected, ranked or summarized. Items are fetched server side and rendered as plain links; no third-party script runs on this page.

An item appearing here is not an endorsement of the publisher. The full list of feeds, and the ones that were rejected, is on the Feeds page.

What this does not tell you: Which stories matter, whether a claim in a headline is true, and what any of it means.

Using the map

Every company in the record, grouped into ten layers of the value chain. The buttons filter to one layer; the search box matches name, ticker, role and country.

A card carries what the company files publicly: where it is domiciled, what it does in the chain, and the figures from its own filings where it has them.

A company appears in one layer only, the one its own description supports. Companies that span layers are placed where their revenue is reported, and the card says so.

What this does not tell you: How companies compare with one another, which are larger or better placed, and which layer is more attractive.

Reading the fab map

Every fabrication and advanced packaging site in the record, plotted at its published location on a Web Mercator projection. Pan and zoom are local to the page; no tile server is contacted and no third-party script runs here.

Shape carries the status: under construction, in production, paused, announced. Colour follows the shape rather than leading it, and the good, warning and critical palette used elsewhere on the web is deliberately absent. A paused site is a state, not a judgement.

Cloud regions are an optional overlay, drawn as diamonds by provider. Provider colours are not brand colours, on purpose, so no provider is visually favoured.

Every site links to the source that announced or confirmed it.

What this does not tell you: Which sites matter, which region is ahead, and what a cluster of sites means for supply.

The country breakdown

A count of sites by the country they are in, from the same records plotted on the map above.

It counts sites, not capacity, not spending and not output. A country with many small sites can appear ahead of one with few large ones.

What this does not tell you: Which country leads in semiconductor manufacturing. Site counts do not answer that question and are not offered as an answer to it.

The timeline

Sites arranged by the year volume production began, where a year is published. Sites without a published start year are absent rather than estimated.

The count for recent years will keep changing as sites that have been announced begin producing and as start dates are confirmed.

What this does not tell you: Whether construction is accelerating, and what the shape of the timeline means for future supply.

The site table

Every record behind the map, sortable, with location, operator, status, process nodes and announced investment where each is published.

An empty cell means the operator has not published that figure. It is not zero and it is not unknown-to-us: it is unpublished, and the distinction is kept.

What this does not tell you: Which sites are most significant, and how announced investment compares between operators.

What the two series are

Two separate measurements drawn on one page: capital spending by the largest cloud operators, and revenue at the companies that make semiconductor manufacturing equipment. Both are reconstructed from regulatory filings.

They are drawn side by side and no relationship between them is asserted. They are different companies measuring different things on different calendars.

Capital spending is reported year to date in quarterly filings, so the quarterly figures here are de-cumulated: each quarter is the filed figure minus the one before it, assigned to the calendar quarter its midpoint falls in.

What this does not tell you: Whether one series drives the other, what the gap between them means, and where either is heading.

The filer panel

Every company in the chart files with the SEC, and the note names which filers are included in the total for each period.

The panel is chosen so that the total covers the same set of filers across the whole period drawn. A company that started filing later is drawn as its own line rather than being folded into a total it was absent from.

Summing whichever companies happen to have filed produces a total that falls when the calendar, not the spending, changed. That is why the filer set is stated.

What this does not tell you: Which company spends most, and whether any company's spending is high or low for its size.

What this page covers

The figures are what these companies filed, for the segments they file. Capital spending covers everything a company builds, not only what it builds for computing.

Companies that do not file with the SEC are absent from both series, and the method page lists them.

What this does not tell you: Nothing is withheld here. This section is the scope statement itself.

The national series

Metered electricity consumption attributed to data centres by the national statistics office of each country shown, on that office's own definition.

Definitions differ between countries, so the series are drawn separately rather than added together. A national total here is that country's measurement, not a comparable international one.

Where a share of national consumption is shown, it is computed here from two published figures and is marked derived.

What this does not tell you: Which country's data centres consume too much, and what any trend means for policy or supply.

What operators publish

Figures each operator publishes about its own energy and water use, taken from its own sustainability reporting, one row per report year.

Operators use different boundaries, different reporting years and different definitions of what counts. The rows are not comparable with one another and are not totalled.

An absent operator has not published, or has published in a form that could not be read without interpretation.

What this does not tell you: Which operator is more efficient or more responsible. The figures are not comparable and cannot support that reading.

What this page covers

National series measure metered consumption in one country. Operator disclosures measure one company across its whole estate. They answer different questions and are kept apart for that reason.

Neither covers the electricity used to manufacture semiconductors, which is a separate question with no comparable public source.

What this does not tell you: Nothing is withheld here. This section is the scope statement itself.

Reading the treemap

Area is revenue, from the most recent annual figures each company filed. Blocks are grouped by layer of the value chain.

Layers are capped at seven plus an Other block, because the colour set holds eight distinct values and a ninth would repeat a hue, creating a visual grouping that does not exist in the data.

Companies that do not file publicly are absent, which means whole parts of the chain are lighter here than in reality. The method page lists them.

What this does not tell you: Which layer is most valuable or most attractive, and what the relative areas mean about the industry's structure.

The two ranked charts

Capital intensity is capital spending divided by revenue, both from the same filing. Gross margin is as the company reported it.

Both are computed here from filed figures and are marked derived. Companies that do not file both figures are absent from that chart rather than estimated.

Rank on this page is a sort order, not a score. The bars are sorted so the values can be read, and colour follows the company rather than its position.

What this does not tell you: Whether a high or low value is good, which business model is better, and what any company's position means for its prospects.

The filings table

Every company in the record that files with the SEC, with the figures taken from its XBRL company facts, sortable by any column.

Each row carries the period the figures cover and the date they were filed. Companies report on different fiscal calendars, so two rows in the same table can cover different twelve months.

Figures are as filed and are not restated into a common calendar or currency.

What this does not tell you: Which company is performing better, and whether any figure is high or low for its type of business.

The source-health grid

One cell per feed, showing whether it answered the last time it was fetched and when it last carried a new item.

A feed that fails is kept in the list with the reason, rather than being removed quietly. A list that shows only working sources claims more coverage than it has.

Some publishers produce malformed feeds that need a lenient parser, and some rate-limit repeated requests. Both are recorded as their own states rather than as failure.

What this does not tell you: Which sources are worth reading, which are more reliable as journalism, and what a publisher's silence means.

What this section is

Each tile is one forecaster's current published figure for one target period. The slider above sets a date, and the tiles revert to the figure that was on the record on that date.

The slider stops only on dates when something was actually published, so every position corresponds to a real event rather than an arbitrary day.

When a tile says nothing was on the record, that is because no forecast for that period had been issued yet. The date of the first one is given.

What this does not tell you: Whether a figure is high or low, whether a revision was large, or what any of it means for a company or a market.

The date slider

Dragging it left removes everything published after the chosen date. Figures issued later still exist in the file; the tiles count them and withhold their values.

The chosen date travels in the address bar, so a link you send opens on the same date you were reading.

What this does not tell you: Any suggestion that an earlier figure was wrong, or that a later one is better.

How this list is built

Every entry that arrived in the record between the two dates, in date order. Nothing is selected for importance and nothing is left out.

Each line names the publisher, the target period, the figure as published, the figure it replaced and the date that earlier figure was issued. The link goes to the publisher's own page.

A first forecast, a restatement, a withdrawal and a result are labelled differently because they are different kinds of statement.

What this does not tell you: Which entries matter, which are surprising, and what any of them imply. The list is complete rather than curated, which is the only editorial position this site is able to take.

The feed and the digest

The same list is published as an Atom feed at /record/record.xml, so a reader can follow changes to the record in whatever reader they already use.

The digest is the same content as plain text, written so that someone who never opens the site still gets the whole fact including the value that was replaced.

What this does not tell you: Any summary, ranking or highlight. The feed carries everything in date order.

How to read a stagger chart

The horizontal axis is the date a forecast was ISSUED, not the period it is about. The vertical axis is the value forecast.

One line per target period. A circle is a published forecast. An open square is the result once the period elapsed, published by the same body as another number.

Colour identifies which period a line is about. It carries nothing else: no colour on this page means good, bad, up or down. The vertical axis is the only thing that carries direction.

A single point with a short dash through it is a target that has been forecast once. It is a fact, not an unfinished line.

The frame is Andy Grove's stagger chart: forecast the same target repeatedly and keep every version.

What this does not tell you: Whether any forecaster revises a lot or a little, whether a revision was justified, and what the shape of a line implies about the future.

How to read the matrix

Rows are the period being forecast. Columns are the date a forecast was issued. Reading down a column is what stood on the record that day. Reading across a row is one forecaster restating one period over time.

Bold with a dot is a value restated on that date. Grey is the last value on the record, not restated in that release.

There are two different kinds of blank, and they are not the same fact. A middle dot means nothing had been published for that period yet. An empty cell means the period had already elapsed, so no forecast for it was still being issued.

The result column carries the outcome once the period is over. Where it says the result is not yet published, the reason is recorded on the method page.

What this does not tell you: Any count, score or comparison of how often a publisher revises. A revision frequency presented as a number is a rating of the publisher, and this site does not rate publishers.

Choosing a series

Three series are in this file, each collected the same way from the same archive: semiconductor production, computer and electronics production, and new orders for computers and electronics. The choice governs all three sections on this page at once.

The unit differs between them. Two are index numbers with 2017 set to 100, and one is millions of US dollars. The figures are the agency's own in each case, and no series is rescaled to sit alongside another.

Period coverage differs too, because the archive holds a different history for each. A series with fewer periods is a shorter archive, not a gap in collection.

What this does not tell you: Which of the three is the better measure of anything, how they relate to each other, and whether a restatement in one says anything about another.

What this list is

Every period that the most recent release restated, in full. The list is complete for that release rather than a selection from it.

Each line gives the period, the value that stood before, and the value that stands now. Both are the agency's own figures.

A release that restated nothing says so rather than showing an empty space.

What this does not tell you: Which restatements are large, which are surprising, and what any of them mean for anyone using the series.

How to read the grid

Rows are the period being measured, grouped by year. Columns are the date the agency published. A number in a cell is how many months of that year were restated in that release.

Depth of ink follows that count and nothing else. There is no hue here, so no cell can be read as good or bad.

An empty cell is a release that left that year alone. Horizontal bands appear because agencies restate on a published schedule rather than continuously.

One archived vintage carries the series' whole history, so the row count is the length of the series, not the window of vintages collected.

What this does not tell you: How often this agency revises, whether that is a lot, and whether any restatement was justified. The unit here is the period, never the publisher.

Reading one period's history

Every value the agency has published for the chosen period, oldest first, with the value each one replaced.

The first row is the first release recorded in this file, not the first the agency ever made. The file begins where collection began.

A row marked withdrawn means the agency published a value and later published none for that period. That is different from never having published one.

What this does not tell you: Whether an earlier value was wrong. An earlier figure is not a mistake; it is what the number was when somebody had to use it.

What a gap entry means

Each row is something the record knows it is missing. The columns state what exists, what is missing, why it is not recorded and what recording it would unblock.

The statuses distinguish a figure that has been read but lacks a required field, a publication known to exist but not located, a source that is not free, a gap where no publisher produces the data at all, and code that has been written but never run against a live source.

Several of these could be closed by guessing and are deliberately left open. A guessed date puts a real number on the wrong day, which is a worse error than an absent point.

What this does not tell you: Any estimate of the missing values. An estimate presented next to published figures is indistinguishable from one, which is the failure this list exists to prevent.

The source ledger

One row per source: what it publishes, its licence, how often it updates, and when this site last fetched it.

Everything on this site is traceable to a row here. If a figure cannot be traced to a listed source, that is a defect rather than a shortcut.

What this does not tell you: Which sources are best. The ledger records what is used, not a ranking of publishers.

Sources that were considered and not used

Each row names a source that was assessed and the reason it is absent: a licence that does not permit redistribution, a paywall, a definition that could not be reconciled, or coverage that varies so much by country that the map would show mapping effort rather than the thing being mapped.

Rejected sources are listed rather than forgotten, so the reasoning can be checked and revisited if a licence changes.

What this does not tell you: Any judgement about the quality of the rejected sources themselves. The reasons given are about fit with this site's rules.