516 companies, technologies and constraints in one model of the semiconductor and AI economy. Every figure carries a source and a date. Every conclusion names what would overturn it.
Ask about any name. The model returns what it depends on, ranked by how much of the position each dependency explains. Underneath sits a directed graph of who supplies whom, with weights. Change a premise and every conclusion resting on it recomputes. You can check the arithmetic yourself.
I · How it works
Evidence goes in one end. A dated, scored view comes out the other. In between are four steps, each doing one job, each open to inspection. The opinion comes last, and it is the smallest step.
Four auditable stages. Nothing is typed straight to a conclusion.
News, SEC filings, prices and fundamentals routed to the entities they touch. Fields that go stale refresh on their own.
538 entities of the chip and AI economy with typed, weighted dependency edges. This is the graph a shock propagates through.
Each thesis's premises are refs into the model. A checker verifies the model still holds them; a flipped premise flags the conclusion stale.
Every surviving conclusion becomes a dated claim with a probability. Each is graded against the outcome when it resolves.
II · The whole stack
Most industry models cover a slice. This one runs from the photoresist on a wafer to the price of a token, and every layer carries numbers that an argument actually depends on. A layer with entities and no argument is decoration, so the third column is the one that matters.
| Layer | Entities | Carrying numbers | In an argument | |
|---|---|---|---|---|
| loading… | ||||
A shock, traced end to end
natural‑gas → iren → nvidia → b200 → energy‑per‑token → inference‑token‑demand
euv → sk‑hynix → hbm3e → energy‑per‑token → inference‑token‑demand
silicon‑capacitor → nvidia → gb200 → cost‑per‑token
These are real paths in the graph, not illustrations. A fuel price reaches token demand in five hops because the model holds the conversion at every step: gas to power, power to accelerator, accelerator to watt-hours per token, watt-hours to tokens served. Most models stop at the third hop.
III · Query the ontology
Type any company, technology or theme. You get what it depends on, what it feeds, who it competes with, and where the risk sits. Runs live in the browser. Watch for names that appear in both directions at once. In this industry your supplier is usually also your rival. This is the trial view. The gated tool and the build-on-your-own-data version are paid.
Runs the real 538-entity graph client-side (trial). Paid tiers add gated live access, your watchlist, risk alerts, and a custom ontology built on your own data.
IV · Live demo
Not what breaks — what moves. Pick a shift the desk is tracking and watch where value goes to, and who it leaves behind. Down to the layer that decides it: HBM base dies, ABF substrate, custom ASICs, foundry second-sources. Each one maps to a scored thesis you can open and argue with.
Client-side view over a fixed sub-graph; the winners/losers are the desk's thesis-derived mapping, not a mechanical output. The live engine runs the full 538-entity model on your names. That is the engagement.
IV · Where it binds
"Energy is a bottleneck" is true of almost nothing in particular. A constraint binds a named subject at a stated scale, and two companies in the same business can sit on opposite sides of it. Below: 10 constraints, 63 scored subject relationships. The utilities that sell into a shortage score none — they own the scarce thing.
| constraint | coreweave | crusoe |
|---|---|---|
| Heavy-duty gas turbine delivery slotsgas-turbine | moderate | severe |
| GPU residual value as loan collateralgpu-accelerator | severe | moderate |
| Large power transformer lead timeshv-transformer | severe | mild |
| Grid interconnection queue positionelectricity | severe | mild |
CoreWeave leases grid-connected capacity, so the queue and the transformer bind it hardest. Crusoe sited off-grid to dodge the queue and took on turbine risk instead. Neither is simply "exposed to power" — they are exposed to different things, relieved by different parties on different clocks. A single chokepoint rating cannot express that, and would rate both names the same.
Every conclusion is composed from its premises, each carrying a probability and the basis for it. and means all must hold — conjunctive arguments are weaker than they read. The desk publishes the number even when it undercuts its own stated conviction, because a premise you can argue with is worth more than a mood you cannot.
Therefore the cascade defence for depreciation weakens, and useful life is a power question rather than an accounting one
Same correction: entered at 0.65 against conclusion 2's composed 0.3185. A depreciation argument denominated in chips does not reach this conclusion at all, but the dependency cannot be cited more confidently than the thing it depends on.
That is the whole product in one card: not "we are confident", but a number, the rule that produced it, and the specific unsourced estimate capping it. Fix that premise and the number moves — which is also how the research agenda gets set.
VI · The conviction board
Conviction is earned by evidence, not asserted. High-conviction views rest on a physical or structural rate-limit the market cannot compete away quickly; low-conviction ones are honest contrarian bets on an uncertain future. Every view resolves on a dated criterion. Below is a selection of the 23 calls in the open book.
VII · Pricing
The scored record is public and always free. Paid tiers point the engine at your names. You get the exposure and the concentration a single-name view hides. Early pricing, while the forward record seasons.
Most engagements start here
Research you buy goes stale the day it is written. This is a pipeline: entities with sourced fields, typed edges, derived quantities that recompute when an input moves, and theses whose probabilities are composed from checkable premises rather than asserted. It checks itself — a change to the model that has not reached the product blocks the commit; a thesis whose falsifier cannot be scored is refused; a premise citing something the model does not hold is rejected before it is written. Those guards exist because each one failed here first, and they come with it.
Also open to
Everything on this page was built by one person: the ontology, the graph engine, the scoring, the validation work that retired four of its own signals, and the site. If your desk needs someone who can build the model and then argue the position from it, I am open to the right role. Semis and AI infrastructure, buy side or corporate strategy.
Early rates are for the first design partners and rise as the public record settles. Not investment advice. This is a research and tooling subscription.
✶ · About the founder
The master-economist must possess a rare combination of gifts. He must reach a high standard in several different directions and must combine talents not often found together… He must understand symbols and speak in words. He must contemplate the particular in terms of the general, and touch abstract and concrete in the same flight of thought.
John Maynard Keynes, Alfred Marshall, 1842–1924, The Economic Journal, 1924
The truth is the whole.
G. W. F. Hegel, Phenomenology of Spirit, 1807
On the shelf: the reading behind the model
Bespoke exposure mapping and thesis stress-testing for the AI-compute complex — the demo above, run on your book: what a shock actually traverses in your names, and what would prove your thesis wrong. Or follow the scored calls as they resolve.