How to use the KXCO Ontology

Most market research hands you a conclusion and asks you to trust it.

This does the opposite. Every line on the map is a claim with a source attached, a date, and an honest confidence level. If you disagree with a conclusion, you can walk back through the claims that produced it and find the exact one you think is wrong. That is the entire design.

The live map is at kxco.ai/ontology-live. Open it in another tab and read this alongside it.

01 · OrientationWhat this thing is

An ontology is an agreement about language. It answers three questions: what things exist, how they relate, and what that means. That sounds academic until you try to answer a question like "if ASML stopped shipping tomorrow, what breaks" and discover that nobody has written down the dependencies in a form you can actually traverse.

It is not an oracle. It is an instrument. The value is not that it hands you conclusions, it is that it puts structure in front of you in a form your own judgement can work on. Every discovery on this map was made by a person looking at it.

So we wrote them down. The AI sector map currently holds:

• 247 entities: chipmakers, foundries, labs, clouds, investors, data centres, governments, people, and the concepts they argue about.
• 544 claims: typed, directional relationships between those entities, each with a source.
• 21 findings: the conclusions the structure supports, ranked by severity.
• 13 listed majors with consensus valuation data, refreshed against the market.

The map was not assembled by listing famous companies. It was grown from a seed set by following dependencies outward, which is why the most important entity on it is a Dutch company most people outside semiconductors have never heard of.

Worth being precise about how that happened, because it is the whole point of the thing. The engine did not decide ASML mattered. It laid the dependencies out visually, and the convergence became obvious enough that a person looking at it recognised what it meant. The data had been sitting in public filings the entire time. Without the display, nobody would have gone looking. That is the function of an ontology: not to make discoveries, but to render structure legible enough that people make them.

02 · The unitA typed claim, not a thing

The smallest unit here is not a company. It is a claim. Every relationship carries:

The two that matter most are disc and conf. A claim marked filed came out of a regulatory filing. One marked reported came from press coverage. One marked estimate is our arithmetic and is labelled as such. We would rather show you a weak claim honestly labelled than a strong-looking claim with nothing behind it.

03 · NavigationThe six views

The left sidebar switches between six ways of looking at the same data. They are not different datasets. They are different questions asked of one graph.

Findings is where to start. It is the map's conclusions, ranked critical down to opportunity, each with the entities that support it. If you only ever look at one view, look at this one. It answers "what should worry me" without making you assemble it yourself.

Analyst Outlook holds the thirteen listed majors with price, market cap, multiple, consensus rating and twelve-month target. This is sell-side consensus that we have collected and dated, not our own forecast, and the page says so.

Ontology Map is the geographic view. Where the physical things are: fabs, data centres, the political capitals. Useful for the questions that turn out to be about geography, which in this sector is most of the hard ones.

Network Graph is the force-directed graph of everything. Filter by relationship type using the chips. The single most useful filter is Circular flows, which isolates the fifteen loops where money leaves a company and comes back as revenue.

Entities is the catalogue, grouped by kind, with a search box. Fastest way to find one company and see everything attached to it.

Revelations is thirty short pieces of narrative for the things that need a paragraph rather than an edge.

There is also a chat box, and you should know its current limits before you trust an answer from it. It answers against the graph rather than the open internet, which is the right design. But the prompt it receives today contains the entities, the claims and the revelations, and does not contain the findings or the valuation table. Ask it about either and it will fall back on general knowledge and sound confident doing so. We would rather tell you that than have you find out. Use it to explore entities and relationships; read the Findings and Analyst Outlook views directly. It also runs on a free allocation, so it can be unavailable.

04 · VerificationHow to check a number

This is the part that separates a map from an opinion. Take any figure and walk it back:

1. Click the entity. The panel lists every claim touching it.
2. Click the claim. You get the predicate, the magnitude, the as-of date, the discovery method and the confidence.
3. Follow the source link. It goes to the filing or the report, not to our own page.
4. If the source does not say what we say it says, we are wrong. Tell us.

There is also a machine-readable copy at /ontology-live/data.json if you would rather check the whole thing at once than click through it.

05 · TimeWhy every claim carries two dates

Research decays. A map that quietly overwrites last month's truth with this month's is lying about its own history, and you cannot audit it.

So every claim carries two independent time axes. Valid time is when the thing was true in the world. Record time is when we asserted it. They are not the same, and collapsing them into one date is how research quietly becomes unfalsifiable.

When something changes, we do not edit the old claim. We close it off and add a new one that supersedes it. The old claim stays in the graph, marked as history. This means you can ask what the map said on a date in the past, and you can see where our own view changed and when.

This is also why a snapshot is not a scorecard. Keeping the old claim intact lets you see what was said and when, which is more useful than marking it right or wrong on a date the map never chose.

The clearest example on the map right now is Elon Musk on AI regulation. In March 2023 he signed the letter calling for a pause on frontier training runs. By July 2026 he was arguing that an industry summit should substitute for government rulemaking. Those are two separate recorded events, the second superseding the first. An ontology that stored "Musk's opinion on AI regulation" as a property of Musk would have silently thrown away the more interesting half of that.

06 · HonestyWhat we do not know

The discipline is that a gap is left empty rather than filled. Of the 580 claims in this map, 515 carry a source URL and 65 do not. The ones that do not are visible as such, with their discovery method and confidence still recorded, rather than being dressed in a citation that does not support them. Closing that 65 is ongoing work.

Two honest notes on the state of this, rather than the intent. The schema supports a first-class opaque-dependency relationship, for a link that almost certainly exists but nobody has published, and this map does not use it yet. And record time, the second of the two time axes, is populated on 36 claims; the other 544 carry valid time only, because backfilling a recording date onto older claims would have invented provenance we do not have.

The refusal to fabricate is real and enforced. The machinery for flagging a specific suspected-but-unpublished link is defined and not yet populated. Those are different things and it is worth not conflating them.

07 · PracticeThree ways to actually use it

Trace a dependency before you take a position. Pick the company you are looking at, open the Network Graph, and walk upstream. Most people stop at the first hop. The interesting risk is usually three hops up, at a supplier nobody in the room has heard of.

Find the shared exposure. Select two or more companies and look at what they both route through. Portfolios that look diversified by sector frequently converge on one fab, one memory triopoly, or one lithography vendor.

Read the circular flows. Filter to circular capital. Fifteen loops, roughly a trillion dollars, where a vendor invests in a customer who spends it back with the vendor. The newest one is a credit backstop rather than equity, which changes the kind of risk rather than its size.

08 · LimitsWhat it will not do

It does not tell you what to buy, and it is not trying to. What it is built to surface is where the gaps are: the dependency nobody has a second source for, the layer that is commoditising while value migrates upstream, the thing everybody needs and nobody can currently verify. Four are named on the map right now as the sovereignty gap, the verification gap, upstream chokepoint value, and provenance for defence AI. Gaps are actionable in a way a price target is not, because a gap tells you what is missing from the market rather than what a stranger thinks a share is worth.

It will not give you a verdict on a short window. The consensus targets it carries are twelve-month views, so a position marked after three weeks is a position with eleven months left to run, not a result. A claim that has not played out yet has not been falsified, and nothing here is downgraded for failing to happen on a schedule the map never set.

This public map is not complete, deliberately. It covers the spine, not every capillary, and it is limited to what can be sourced from public material. It is a demonstration of a method, not the working system.

And this public map is not live. It is a periodic snapshot, refreshed on purpose, with the refresh date stamped on the page. If the stamp is old, treat the numbers as old. The ontologies KXCO runs internally and builds for clients are live, which is a different product and the next section says so plainly.

It will not settle an argument about motive. We record what people said, where and when. We do not record why they said it, because no source can attest to that.

09 · ScopeThe ones you cannot see

This map is a sample, and it is the smallest thing KXCO runs. It is public because the AI supply chain can be sourced entirely from public material, which makes it a fair demonstration of the method without exposing anyone's data.

The difference is not one of degree. What KXCO runs internally, and what we build for clients, is live rather than snapshotted, and materially closer to full reality than a public map constrained to what has already been published can ever be. Those systems ingest continuously, carry the relationships that appear in no filing and no news feed, and hold the internal material a public map has no right to: holdings, counterparties, exposures, contracts, obligations, the dependencies a business only knows about itself.

Same engine, same typed claims, same two time axes, same refusal to fill a gap with a guess. Different completeness, different cadence, and none of it ours to publish. If you want to see what that looks like on your own domain, the contact page is the way in.

If that is interesting, the contact page is the way in.

10 · OriginWhy we built it

This was built by Shayne Heffernan and John Heffernan as a different way to research and monitor markets.

The problem we kept hitting is that market research arrives as prose. Prose is good at persuading and bad at being checked. You cannot query a PDF for every position that routes through a single Dutch lithography vendor, and you cannot ask a newsletter what it believed six months ago and whether it has quietly changed its mind since.

A graph can answer both. Structure the claims instead of the conclusions, attach a source and two dates to every one of them, and the analysis becomes something you can interrogate rather than something you have to believe. The conclusions then fall out of the structure, and when the structure changes the conclusions change with it, visibly.

That is the whole idea. The AI sector was the test case because it is fast, opaque and consequential. The method is not specific to it.