Where every number comes from

Nothing enters the facts feed without a named source and a date. If a claim cannot be sourced, it is not a fact and it does not appear — not in the feed, and not in the ticker at the top of the page, which renders the same dataset.

Where a figure is a single reading taken on a particular day, the section shows that date rather than implying it is live. Data that ages past its useful window is labelled as old rather than left to pass as current.

Every fact in the archive carries one of three tiers, shown as a chip on its card. The tier does not rank the fact — it says how strong the receipt is.

Primary. Straight from the body that produced the number — a filing, a dataset, a model card, a statistics agency.

Reported. A reputable outlet reporting the figure and naming where it came from. Traceable, but second-hand.

Estimate. A named analyst house's projection. Honest, but it is a guess.

How facts decay. Historical facts — dated events — never expire. Trend facts, a rate or ratio a source will revise, are rechecked every 180 days by default, or sooner when the source itself publishes on a shorter cycle — TSMC's foundry share, for instance, is rechecked every 90 days because TrendForce reissues that figure quarterly. Live facts, bound to an API field, refresh their number and are reframed every 30 days.

No fact ships without a named originator and a resolvable URL. That is enforced in code, not by review.

See every active fact, filterable by topic and tier, on the fact archive.

A new fact does not appear on the site the moment it is proposed. It sits in a queue until a person approves it. An email lists everything currently waiting, with a link to approve or reject each one, or to act on the whole batch at once.

Each approval link expires after 48 hours. A fact whose link expires is not published — it stays in the queue, and the next digest carries a fresh link. Nothing reaches the site by default. Facts already verified and already in rotation are unaffected by this — they keep publishing on their existing schedule.

The fast path. A figure can break between digests — a lab drops a model card, a company files. For that, a separate fast-path queue exists, but it runs the identical gate above: the fact waits in the same queue, appears in the same digest labelled Urgent, and needs the same human tap to approve or reject. Nothing about being fast skips a human. The only difference is timing after approval — once a person approves it, it posts on the next posting check rather than waiting for the next day's queue, and it does not use one of that day's scheduled slots. Holding a digest stops this too: an approved urgent fact does not post while its day is held, the same as any other scheduled post.

A live fact's figure is re-read from its source on a schedule, without waiting for a human to notice it is stale. What that process is allowed to touch is fixed and narrow: a number, its period, and the date it was last checked — never the sentence around it. The wording a human wrote and approved does not change on its own, ever.

A figure that moves sharply is held for a human. Once the process has a prior reading on record for a fact, a fresh one differing from it by more than 15% leaves the site unchanged: the proposed number goes into the same approval queue every new fact passes through, showing the old and new values side by side. A smaller move is applied automatically, because a sentence like “has grown roughly” is not put at risk by a figure ticking up a few percent — it can be by a figure that doubled, or fell.

The exception is the very first check. Before there is a prior reading to weigh a move against, an automatic check has nothing to compare with yet, so it applies whatever the source currently reports without a threshold. Every check after that compares against what was last recorded.

Facts are re-verified on a schedule, not just refreshed. Live facts are checked every 30 days, trend facts every 180 by default — the same intervals described above. A fact that is not re-confirmed by its date leaves the fact archive rather than sit there un-rechecked, and comes back as a plain “is this still true” question in the next digest — the same queue, the same human approval, before it returns.

Source links are checked, and flagged when they break. Each fact's link is tested periodically. A link that fails twice in a row is queued for a human to look at. A broken link never deletes or hides a fact by itself — only a person removes one, and only after actually checking what happened to the source.

Put plainly: an automated process can update a number, but it never writes a claim. Composing a sentence about what a number means is a human judgement, and it stays one.

Source: SEC filings (10-Q, 10-K) via the SEC XBRL API. Each company's cash capital-expenditure line is read from its own filings through the SEC's structured company-concept API — Microsoft, Alphabet and Meta report it as PaymentsToAcquirePropertyPlantAndEquipment, Amazon as PaymentsToAcquireProductiveAssets.

Quarterly figures are derived the way the filings report them. A 10-Q states cash-flow lines cumulatively from the start of the fiscal year, so a quarter is either a directly tagged three-month figure or the difference between two cumulative periods that share a start date. The fourth quarter is the annual 10-K figure less the nine-month figure. Microsoft's fiscal year ends in June, so its quarters are mapped onto the calendar quarters they actually cover — otherwise the four companies would not line up.

What this excludes. These are cash purchases of property and equipment. They do not include finance leases, which Microsoft in particular uses heavily, so the figure here is lower than the headline capex number Microsoft quotes. They are also total company capex, not AI capex: AI infrastructure is what is driving the growth, but fulfilment centres, offices and ordinary facilities are in these numbers too. Anyone presenting this as pure AI spending is overstating it.

Refresh. Re-read from the SEC daily. The chart only extends to the most recent quarter that all four companies have filed — the big four do not report on the same day, and a combined total missing three of them is not a total.

Scores are the Artificial Analysis Intelligence Index and prices are that same provider's blended price per million tokens, both from Artificial Analysis, read on the date shown on the section.

Every figure comes from that one source. Rival leaderboards run different evaluations and score the same models several points apart, so mixing them would produce a table where no two rows mean the same thing. When the board is refreshed, every row is re-read at once rather than patched individually.

The value frontier is the dashed line: the models that nothing cheaper beats on score. A model sitting below the line is dominated — something cheaper is also smarter.

The proxy tickers are ours, not the data provider's. They describe how a public-market investor gets exposure to a lab — a listed parent, a major backer, a strategic investor. Exposure is not ownership, and a ticker beside a model is not a view on that security.

There is no composite 0–100 score. There used to be one, built on placeholder values, and we removed it rather than dress it up. Every figure in the section now carries one of three provenances, shown as a chip on the card:

Company disclosed. The AI share of revenue, but only where the company publishes such a figure — Nvidia's Data Center segment, Broadcom's AI revenue line, TSMC's HPC platform share, Microsoft's AI run rate. Each links to the filing or release it came from, with the period it covers and a note where the figure is not quite AI revenue. Where a company discloses nothing, the card says not disclosed and shows the nearest cited signal instead. We do not estimate the gap.

Derived from filings. Revenue, R&D and capital expenditure for the most recent full fiscal year, from SEC XBRL data, with the ratios computed from them. TSMC files under IFRS as a foreign private issuer, so these are unavailable for it and shown as such rather than substituted.

Our read. Chain criticality and moat durability are editorial judgement on a four-step scale. They are deliberately kept out of any calculated figure, so nothing on the card looks measured when it is an opinion.

Why there is no single number. Exposure means opposite things at opposite ends of the chain: for a supplier it is revenue earned from AI, for a hyperscaler it is money spent on it. A composite of consolidated financials ranks Meta, which buys AI hardware, far above Nvidia, which sells it. A score that inverts the obvious answer is worse than no score.

Refresh. Each earnings season, by hand. Segment and AI-specific figures are not machine-readable — the SEC's structured API returns consolidated facts only — so unlike the capex tracker this section cannot refresh itself.

Prices are 15-minute delayed and come from EODHD. They refresh while the US market is open, and the table stamps the time each snapshot was taken. Outside market hours the last snapshot stands, labelled with its own time rather than presented as live. One listing is quoted in Korean won — Samsung — and is restated in dollars at the USDKRW rate the same vendor publishes, which the table names and dates beside the delay note. The price, the 52-week range and the sparkline are restated at that one rate, so the range is today's dollar equivalent rather than the dollar range as it was actually traded; the percentage column is left as the stock's return in its own currency, because a single spot rate cannot convert a return. Without a rate that row shows dashes rather than won under a dollar sign.

The index tiles, and the one that is not an index. The S&P 500, NASDAQ, DOW and VIX tiles carry those indices' own levels, from the same vendor as the prices. The Russell tile does not: no Russell index is available on our data feed at all, so it shows IWM, the iShares Russell 2000 ETF — labelled as the fund it is and priced in dollars rather than dressed up as an index level. It tracks the index closely. It is not the index.

Changes and the 52-week range are end-of-day. The 7, 30 and 90-day columns are measured in trading sessions, not calendar days, and use the adjusted close, so a stock split does not read as a crash. The sparkline is the last thirty sessions of the same series.

Market cap and P/E are derived by us, not quoted. Shares outstanding come off the cover page of the company's own latest filing and diluted EPS is summed from four consecutive filed quarters, both read from SEC XBRL data; market cap is that share count times the current price, and P/E is that price divided by those trailing earnings. Every figure carries the accession numbers of the filings behind it. A company that files no financial statements with the SEC — Samsung, BE Semiconductor, SoftBank, Tencent — shows no market cap and no P/E rather than a number from an unstated source, and a loss-making company shows no P/E rather than a negative one. Our figure is GAAP diluted trailing twelve months, so it will differ from sites quoting an adjusted basis; we do not tune it to match them. Twelve filers publish no cover-page count the SEC's data interface exposes — companies with several classes of stock answer nothing at all there, and a few others answer with an empty list — so for those we fall back to shares outstanding from the balance sheet, and failing that to weighted-average diluted shares, which is an average over the quarter rather than a count on a date. A market cap built on either fallback is marked with an asterisk that names the basis, and a fallback older than about a year is refused outright rather than used.

Which layer a company sits in is our editorial judgement. A company is placed by where its AI revenue is actually recognised, not by where its marketing points, and every row carries the one-line rationale for its placement. A row marked proxy is a listed route to something you cannot buy directly — the frontier labs are private companies — and exposure is not ownership.

The table is information, not advice. No ranking, sort order or placement on it is a recommendation to buy or sell anything, and a company appearing here is not an endorsement of it.

Every photograph on this site was made by an AI image model — the bands behind the section headings, and the backdrop on each fact card, both here and on the cards posted to X. Not one of them documents a real place, a real building or a real moment. They set a mood for a section; they are never evidence for anything, and nothing on the site rests on one.

What they never contain. No identifiable company, product, logo or livery, no real location, and no faces. One early image came back carrying something close to a real shipping line's name on a container and was regenerated rather than published: on a site whose promise is that every claim resolves to a source, a picture that half-resembles a real company is a claim nobody can check.

Two things on the site are not generated. Company logos in the stock screener are each company's own mark, shown unaltered and never recoloured. And every chart, table and figure is drawn from the data described above — none of them is an image, and no number anywhere on the site was produced by an image model.

AllFactsAI publishes research and educational content. Nothing on the site is investment advice, and no content is a recommendation to buy or sell any security. Scores, rankings and trackers are analytical tools, not ratings.

No company pays for placement, ranking or coverage. Where a figure is estimated, illustrative or excluded from a definition, that is stated on the section itself rather than only here.

Found something wrong? That is worth more to us than a polite email — admin@allfacts.ai.