Methodology

Where the figures come from, how the ranking works, what a blank means, and how peers are chosen.

Everything on this site is a claim about a real company, so this page says where each kind of claim comes from, how confident it is, and what it means when there is nothing there. If a figure here disagrees with a company's own reporting, the company is right and we would like to know — send a correction.

Who gets listed

A company is listed when it is AI-native, or AI is central to its product's value. That rule has held since the first companies were added and covers a frontier-model lab exactly as it covers a decades-old enterprise vendor whose product line has become an AI product — Notion, Canva, Microsoft, Oracle, IBM and Salesforce are here on that basis, and a newer entrant such as Figma or monday.com is judged the same way, not as an exception to it. A company that merely uses AI internally, or bolts one AI feature onto a product AI does not define, does not qualify.

A company outside AI entirely earns a record only when it clears two tests at once: it is itself AI-native, and it owns more than one company already in this directory. Owning several is not enough by itself. SpaceX clears both — it ships its own AI product line (Starmind) and owns two companies here, xAI and Cursor (Anysphere) — and is listed. SoftBank owns two companies here as well, Arm Holdings and Graphcore, and is not listed: SoftBank is a financial holding company, not an AI company, so the first test fails even though the second one passes. That distinction was checked explicitly rather than assumed, because the two cases look identical on the "owns two" test alone. An acquirer that owns one AI company here and nothing else is not listed on any strength: Reliance Jio, which owns Haptik, and TransPerfect, which owns Unbabel, are each named only in their subsidiary's own description, because minting a sourced acquired relationship — the kind that renders Acquired by X (year) under a company's name — needs the acquirer to be a record in its own right, and neither is one. That will keep happening — an acquisition outside AI does not, by itself, admit the acquirer, and neither does simply owning several.

Where the figures come from

Company figures were compiled in August 2026 from public sources:

Product records were drafted from the company description and then checked against the company's own site: every product link and pricing link in the dataset was requested on the date recorded against it. That check is not cosmetic — it caught eight product pages that had silently moved, including a whole vendor's documentation shifting to a new domain.

Figures for private companies are frequently estimates, and are treated as approximate throughout — total funding, valuation and employee counts on a private company's own page carry an est. marker with the month of their source, unless the source is the company's own disclosure or a regulatory filing, which are reported rather than estimated. Nothing on this site is investment advice or a valuation.

How records are dated, and how to check one

Every company and product page carries a Record checked date in its footer. That date is the honest scope of one claim: when the record was last reviewed, not when the company last changed.

Individual figures are dated separately, and by their own source. Where a figure cites something, the line under it names the kind of source, the date that source is as of, and links to it — so a headcount read from a company’s last annual filing is dated to that filing and not to the day we read it. 1075 figures across 380 companies carry such a line today.

A figure can be in one of three states, and the page distinguishes them:

The Sources figure in each record’s statistics counts how many of that record’s own shown figures cite one. There is no automated refresh. A figure ages from the moment it is written, and a dated figure at least says how much.

How the ranking works

The grid opens A–Z and prints no figure, so it makes no ranking claim until a reader asks for one. The Value sort is that ranking, and it is a cascade with an explicit end:

  1. Valuation, where there is one. 219 of 426 companies have one.
  2. Total funding raised, for companies with no valuation figure. That covers 69 more.
  3. Neither, for the remaining 138. They sit together at the end of the list, sorted alphabetically among themselves — never by a number the data does not support — with nothing marking where they start. The grid used to break here with a labelled band; Jae asked for one uninterrupted list instead, so the boundary is gone and the ordering underneath it is not.

So 288 of 426 companies (68%) are ranked on one of the two figures. Every tile says which one it is using — $5.3T market cap beside $12.9B funding — because a single column holding one company's market capitalisation next to another's money raised is only honest if it says which is which.

And a valuation is labelled for what it actually is. For the 108 public companies the figure is a market capitalisation: a price the market sets and resets every day. For the 122 private companies that carry one it is a valuation: the price implied by the last funding round, which may not have moved in a year. The two are not comparable in the way one shared word would suggest, so the tile, the drawer and every company page name them separately. Money raised needs no such split, and funding means the same thing for everyone.

A third case, ruled on 2026-09-14: an agreed but not yet closed acquisition price is not a valuation either. A private company can be the target of a signed, publicly announced deal for months before it closes, and the price the acquirer agreed to pay is not what the market — such as it is — has priced the company at; it is what one buyer offered. So it stays out of valuation and out of the Value sort until the deal actually closes, the same rule that keeps status from moving early. The figure is not dropped — it is recorded as its own dated, sourced fact alongside the standing valuation, so a reader comparing the two sees both numbers and how old each one is. Hugging Face is the first case: its recorded valuation is a 2023 funding round, while NVIDIA's signed 2026 agreement prices the company nearly three times higher — a real, useful fact that still is not a valuation until the acquisition closes.

The cascade deliberately stops there. Continuing it through revenue and headcount was proposed and measured against the real data, and it made the ranking worse in two specific ways: it mixes units, which put a large public company below almost every startup in the file, and it empties the unranked group, because headcount is recorded for everyone — so “we do not know” becomes indistinguishable from “very small”. An honest gap beats a confident wrong answer.

Revenue is recorded, and shown, and deliberately not part of any of this. 116 of 426 companies carry a revenue figure — every listed company and a handful of private ones that have disclosed a number — and each appears in the Sources section of its page with its source and its date. It is not in the statistics grid and nothing is ranked on it, for two reasons. It is the thinnest figure here, so a row for it would be blank on most pages and both sides of a comparison carry it barely one time in six. And it is not one measure but two: a listed company reports revenue actually earned over the last twelve months, while a private company’s figure is usually an annualised run rate — a recent month or quarter multiplied out, which runs ahead of what the company has earned. Every record says which of the two it is. Putting them in one column a reader scans would invite exactly the comparison the label exists to prevent.

What a blank means

Three different things look similar and are not:

—
Not recorded. The figure was not found in a public source, or has not been checked. It is not zero and it is not small.
Undisclosed
Same meaning, shown on a company tile where a figure would otherwise be. These are the 138 companies with neither a valuation nor a funding figure on record.
$0
A real, recorded zero — Midjourney and Surge AI raised no outside funding. It is shown differently from a blank on purpose.

The same rule runs through the product records. A product whose charging model could not be established from an official page is recorded as having no pricing model plus a note saying what was checked and when — never a plausible guess. 515 of 1892 products are in that state.

Companies and products

1,892 products are recorded across 420 of the 426 companies. The remaining 6 have no product record, and that is a finding rather than a gap: each ships nothing it names as a product, names only models, or sells entirely through a separately branded subsidiary.

A product and the model it runs on are separate things. ChatGPT is a product; GPT is a model. The names collide constantly — Claude, Grok, Kimi and SenseNova each name both an assistant people use and a model family that products are built on — so where the two share a name, the record here is the product: the thing a person can open and pay for. Models get their own records later.

Every company page ends with one Related companies rail. It used to be two sections — Competitors and Closest by capability — and that was the problem: two words were carrying five different relationships. A company's largest customer, its direct rival and a reseller of its hardware could all appear under one heading, and nothing on the page told you which was which.

So the relationship moved onto each pair. Every row states what the two companies share and where they diverge, and carries badges saying what that claim rests on. Nothing is collapsed into a verdict: the site shows you the relationship and leaves the judgement to you.

Competitors come first, and only a source puts a company there. When a company has at least one recorded rivalry, its page opens with a Competitors tier above the related list. A company is in that tier for exactly one reason: somebody read a filing, a vendor document or a measured-market report naming a product of each as alternatives for a named job, and recorded it with the quote. The badge on the card names the job — AMD is NVIDIA's rival once on buying data-centre compute and once on programming GPUs, and the card says both — because a company rarely has one competitor; it has different competitors per line of business. No score, category or size cut puts a company in the tier, and that is deliberate: measured on this data, every such cut either listed twenty companies or listed none, and a competitor the site cannot source is not one it will assert. 235 of the 426 companies have a Competitors tier today; the rest say in a sentence that no rival has been sourced yet and show the related list, so the absence is stated rather than implied.

Where a related row can come from. Three things, and none of them is an opinion. A sourced competitor row exists because somebody read a source naming the two as alternatives for the same job and recorded it with the sentence that says so; 235 of the 426 companies have at least one, and those rows are the Competitors tier above. A sourced partner row is the same kind of record pointing the other way: a source naming the two as working together rather than competing, which matters most where the companies also overlap — Dell and NVIDIA sell against each other in AI servers and Dell is a named NVIDIA partner, and a rail that showed only the rivalry would be telling you half of it. A capability overlap row is computed from what the two companies' products do. All three sit in the same rail and each row says which it is.

The badges are literal. Named in filing appears only when the source is a company's own regulatory filing — the strongest evidence available — and a curated row from a measured market says sourced competitor instead, because "named in a filing" would not be true of it. Capabilities, scale and layer badges are computed from recorded fields.

What the pivots do. The chips filter the rail by axis: capability overlap, same stack layer, comparable scale, named in filings, same stage. They filter what is already there; none of them recruits a company the rail did not already contain.

The honest limit. A capability overlap is only as good as the product records behind it, and 130 of the 426 companies have one product record or none. For those the rail leans on scale and stack layer, and the reason sentence says so rather than implying a match it cannot support. Where a company's scale is not public, the row says the scale is undisclosed instead of guessing a band. The same limit governs a comparison's capability boxes: a capability one company's records carry and the other's do not is shown as Not verified for the second company, never as the first company's alone — a missing record is not a missing capability. A comparison says X only only when the other company carries a dated, sourced record that it does not offer the capability, and that record is shown under the box.

What money cannot buy

No payment changes a figure, a ranking, a peer set or the order of any list. That is a rule about this page as much as about the ranking. Every ordering on this site is computed from the recorded fields described above, and it has to be, because a ranking that could be paid for would make every claim on the page unfalsifiable: you could never tell whether a company appeared because it fit or because it paid. Should the site ever take vendor money, it will be for something that has no effect on ordering, and Terms will say what it is before the first sale.

Corrections

Reader corrections are the fastest and most accurate refresh this dataset has. Every company and product page carries a link for one, and quoting the record id from that page is what makes a correction cheap to apply. A correction backed by a public source gets applied faster, because it can be checked.