How the map works
Singularity Business Models tracks the new ways value is being created as AI transforms the economy. The unit is the business model, not the company. Companies appear as evidence that a model is working, or that it is under pressure.
What gets on the map
A model enters when three things are true.
- It has a loop. We can say what turns each time the model runs, and why each turn makes the next one easier.
- It is working now. At least two companies are running it, each with a dated traction signal: disclosed revenue, paying customers, deployment at scale, capital committed against it, or a second company copying it.
- The evidence can be checked. Every fact links to its source.
Models with real signals but thinner evidence sit On watch until they qualify.
Three labels on every model
Every entry answers three questions, one label each. A reader sees: Pay per outcome · Rung 2, Rides · Selling the finished work · learning loop.
1. Rung: how aligned is it with the exponential? The rung is stated as loop gain: each time AI gets markedly better and cheaper, how much stronger does one turn of this model's loop become?
| Rung | Loop gain | Plain test |
|---|---|---|
| 0. Fights | Negative | The curve shrinks what it charges for (seats, hours, heads, words, clicks). |
| 1. Flow through | About zero | The curve helps it, but the gain passes straight through to customers and rivals. |
| 2. Rides | Positive | Each step of the curve makes its own loop stronger. |
| 3. Compounds | Rising | The loop makes its users more able, so the gain itself grows. |
| 4. Self-improving | Self-set | The business redesigns its own loop, increasingly by agents. |
| Leveraged | Borrowed | A flag, off the ladder: the gain comes from expectations of the curve and reverses if it slows. |
Leveraged is a flag, not a rung. A leveraged model can look like rung 2 or 3 while the curve holds; what it rests on is the expectation, and it reverses if the curve slows.
2. Value family: where does the money land? Five families: selling abundance (volume of cheap intelligence); selling the finished work (the outcome, not the tool or the hours); owning the scarce inputs (compute, energy, capital, capacity); owning the scarce human (original human signal, expertise, likeness); owning the choke points (where machine choices are made and paid). Retiring models carry the family their value is moving toward, which is what a reader on the retiring side needs to know.
3. Loop: what mechanism makes it compound? Seven loops: cost-curve (cheaper models, lower unit cost, more volume, more revenue); learning (use, data, better product, more use); network (more participants (people or machines), more value per participant); control-point (own where choices are made and paid, capture the flow); capital (expected demand, capital, capacity, demand that justifies more capital); human-signal (more machine output makes original human signal scarcer and worth more); self-improving (each cycle makes the participant, or the business itself, more able). The family says where value lands; the loop says why it compounds. One family can hold several loops.
Placements are judgments drawn from the evidence. They are published with the entry, can be disputed like any fact, and move when the evidence moves.
How evidence is graded
- A. A primary source: a company filing, earnings release, official announcement or regulator record, opened and checked against the claim.
- B. A named, reputable news outlet reporting the specific fact, opened and checked.
- C. An estimate, an aggregator, or a source we could not open. Held back from public pages until it is upgraded to A or B.
Announced is not delivered. A signed partnership is not revenue, and we say which one we have.
What every entry carries
- Its three labels: rung, value family and loop.
- What compounds. What grows as the model runs, and why growth makes the next unit easier.
- Who accelerates. Whether the people inside the model (customers, workers, contributors) get more capable and better off as it scales, or whether the model grows at their expense. This is also what separates rung 3 from rung 2.
- What it displaces.
- When it stops working. The condition, drawn from the evidence, under which the model fails.
- Counter-evidence. Reversals, failures and disputes sit beside the positive case, not below it.
- Named companies, each claim dated, graded and linked.
States and movements
Rising models are Emerging (dated traction, earlier or smaller), Scaling (material, disclosed revenue at two or more companies) or Contested (strong evidence on both sides). Retiring models are Early pressure, Under pressure or Declining.
Every change is a movement: a change of state, a move up or down the ladder, or a change of family or loop. Each is logged with its date, the evidence that moved it, and what it was before. The history of every entry is public.
Disputes and corrections
- Anyone can dispute any entry, its placement included, through the reply door on every page.
- We answer every dispute within seven days.
- When a dispute changes an entry, the correction is published with its date and the reason. When it does not, we reply to the person directly.
- We publish our assessments on this method and on cited facts. We do not send notice before publishing.
What we refuse
- No paid placement, no pay-to-list, no sponsored verdicts. No entry, and no rung, can be bought.
- No model without evidence.
- No private individuals named.
- No rankings of companies by how badly they are doing. On the retiring side, the unit is the model, and what holds it up is stated beside what is pressing on it.
Who runs it
Singularity Business Models is run day to day by AI agents, with human oversight: sourcing, grading, placing models and updating the map. The site is run the way many of the models it maps are run, and that is part of what it shows.
The AI that runs it is Claude, made by Anthropic. Anthropic and its competitors appear on the map as evidence. Their entries are held to the same grades as every other company, and each of those pages says so.