AI Deconstructed

AI companies do not grow on a smooth diagonal. Their expansion is a sequence of distinct regimes — research demo, narrow product fit, distribution flywheel, defensible scale — and each regime rewards different bets. Teams that confuse the regime they are in for the one they want to be in tend to stall halfway up the curve.

1. 01

   Finding capability–market fit

   Before raw model quality matters, an AI product has to land on a real piece of work that someone would notice losing. The signal is not benchmark improvement but retention inside a specific workflow — users returning because the system actually closes a loop they care about. Spending on ads or sales before this regime is reached is pouring water into a cracked beaker. Measure activation, repeat use, and how loudly early users describe the product to peers. Revenue is a secondary instrument here.

2. 02

   Building repeatable acquisition

   Once capability–market fit is real, expansion finally pays off — but expansion here does not mean raw revenue growth. It means holding cost per acquired user roughly constant while volume grows. The work of this regime is finding a reliable acquisition channel (developer onboarding, product-led free tier, partner integrations, content surface) and designing the team and tooling that can run it without breaking.

3. 03

   Building self-reinforcing loops

   AI businesses become hard to dislodge when usage itself improves the product — more queries surface more failure modes that feed evals; more enterprise deployments yield more domain-specific traces that sharpen fine-tunes; more developers building on an API expand the integration surface that newcomers must replicate. These loops do not appear by accident. They are designed in: telemetry, feedback collection, model–data flywheels, and developer-platform structure are all the same problem viewed from different angles.

4. 04

   Turning advantage into a moat

   Once the business is large enough to matter, the question becomes what is still here in five years. Switching cost, capacity advantage, proprietary data layers, distribution embeddedness, and trust in regulated domains — the deeper these stack, the harder the company is to clone. Compute alone is not a moat: it can be rented. Models alone are not a moat: open weights catch up. Moats in AI are almost always composite, and the moment just before peak growth is the cheapest time to deepen them.

A business model is a choice about who pays, in what unit, at what frequency. In AI, that choice also encodes a position on compute exposure, gross margin trajectory, and how exposed the company is to the next price drop in inference. Read the shape of the revenue before you read the size.

1. 01

   The four pricing primitives

   Most AI revenue resolves to four shapes: per-token / per-call usage pricing (variable, tracks compute), per-seat or per-workspace subscription (predictable, decoupled from inference cost), outcome- or volume-based contracts (charge per resolved ticket, per generated asset, per closed lead), and hardware-led models where a device or chip carries the value, with software and services trailing. The same workflow can be sold in any of these four shapes — and the resulting income statement looks completely different in each case.

2. 02

   Unit economics in an inference-priced world

   In AI, unit economics has an extra term most software businesses do not: cost of serving a customer is not flat. Every active user generates inference cost that can move with model size, prompt length, and retrieval depth. Healthy unit economics in this category require either holding inference cost down (smaller models, caching, distillation), pricing it through (usage-aligned billing), or extracting enough downstream value (workflow lock-in) that gross margin survives anyway. Companies that price subscription but pay for inference per token live or die on this math.

3. 03

   Two-sided structure in AI platforms

   Model marketplaces, agent platforms, and AI developer clouds are not selling a product directly — they are arranging traffic between two sides. Model providers and application builders. Tool authors and agent runtimes. Annotators and labs. Once each side reliably pulls the other in, the platform itself becomes the standard, and the cost for a newcomer is no longer building a comparable product but rebuilding the surrounding ecosystem.

4. 04

   Fixed-cost leverage in compute-heavy businesses

   The reason a mature AI lab can swing from cash-burning to high-margin so quickly is the same reason it bled cash on the way in: pre-training, research staff, and reserved capacity are mostly fixed. Below break-even those fixed costs dominate; above it, every additional dollar of API revenue drops disproportionately into operating profit. This asymmetric curve explains why so many AI businesses look terrifying for a few years and then suddenly look obvious.

AI operators come from a narrower distribution than software founders generally — research backgrounds, infrastructure backgrounds, applied-ML at scale, or domain experts who only became technical because their problem demanded it. The way they decide is recognisably different from the operating playbook of an enterprise-software CEO, and reading that difference makes the next move of a company much easier to anticipate.

1. 01

   Why they started

   AI founders cluster into three motivations: a research conviction that a particular capability is now within reach, a product conviction that a specific painful workflow can finally be automated, or an infrastructure conviction that someone has to build the missing layer (tooling, evals, agents, governance) before everything else stabilises. The motivation persists into every later decision — what they over-invest in, which trade-offs they refuse, and whose criticism they take seriously.

2. 02

   Deciding under model uncertainty

   AI operators have to bet before the substrate they depend on has settled. Tomorrow's frontier model could obsolete this quarter's product. Strong operators handle this by being explicit about which assumption is load-bearing — "this only works if context windows keep growing," "this only works if open weights stay six months behind frontier" — and by running the cheapest experiment that would falsify that assumption first. They do not run a balanced analysis; they sequence kills.

3. 03

   Designing risk, not avoiding it

   What looks from outside like a bold bet — pre-training your own model, going closed-source, committing to a single vertical — is usually the visible end of a series of smaller, contained experiments. AI operators treat risk as something to allocate: where can we lose, and how do we keep that loss reversible? The skill is not boldness; it is structuring exposure so that one bad call does not end the run.

4. 04

   What they do with failure

   In AI culture, failed training runs, abandoned model lines, and shipped-then-pulled features are not stains — they are tuition already paid. What matters is not the count but the density: which decision improved because of what was learned, and how quickly did that learning propagate to the rest of the team. An operator's relationship with failure quietly sets the entire organisation's tolerance for ambition.

The earliest stage of an AI company is not a coding sprint and not a fundraising sprint — it is a sprint to find out whether the problem you are pointed at actually exists in the shape you imagined. The quality of work done in these weeks compounds into every later decision.

1. 01

   Falsify before you train

   The fastest way to learn whether anyone needs what you are about to build is to ask them before you build it. Sit with the intended users, ask how they handle the workflow today, ask what they would pay to make it disappear, ask where their current tools fail. The goal of this stage is not "find the right answer" but "narrow the surviving hypotheses." Training a model before this work is done is the most expensive way to discover you were solving the wrong problem.

2. 02

   Co-founder composition

   The combination of co-founders sets the cultural and technical durability of an AI company more than any single hire afterward. The strongest pairings cover three angles — research / model work, product and applied engineering, and the customer or domain — and have shared experience under pressure, not just shared interests. The capacity to disagree sharply and recover by next week is rarely visible in casual interviews; it is worth the months it takes to confirm.

3. 03

   Demos before training

   Minimum-viable in AI does not mean a smaller model — it means the cheapest thing that creates real usage signal. A wrapper around an existing API, a curated prompt, a thin agent, a Notion form behind a "magic" backend that a human runs — any of these can produce more learning per week than a from-scratch fine-tune. If shipping the first version feels embarrassing, you have probably already missed the optimal moment to ship it.

4. 04

   Living close to the first users

   Early users in AI are not revenue — they are co-developers of the product's direction. Who you choose to deploy with first determines which failure modes you see, which evaluations you build, and which traces fund your eventual fine-tunes. The intimate conversations with the first thirty users produce the insight that lets you reach the next three thousand. Outsourcing this contact to a sales team too early is the most common avoidable mistake of this stage.

Understanding an AI company is not the moment you have collected the most information about it — it is the moment you can describe its mechanism and its reason to persist in your own words. The four questions below apply to any AI business, regardless of stage, sector, or open-vs-closed positioning.

1. Q1

   Revenue design — where does money enter?

   Who is paying, for what unit, at what frequency? Is this per-token, per-seat, per-outcome, or hardware-led? If there is more than one stream, which is the engine and which is the rider? You should be able to describe the revenue surface in one sentence before you read anything else about the company.

2. Q2

   Cost shape — what carries the operation?

   Is this a research-heavy business burning fixed cost on training, an inference-heavy business whose cost scales with usage, or a human-in-the-loop business where labelling and review dominate? Where do the GPUs go? Where do the headcount dollars go? This axis tells you where growth will first break and what level of margin is even structurally possible.

3. Q3

   Demand — who keeps coming back, and why?

   New users are easy to acquire when a category is hot. Returning users are the only honest signal. Who is the cohort that uses this every week six months in? What problem actually got solved for them — and would they pay more if it disappeared? Acquisition and retention are entirely different problems, and the latter is what determines long-run intensity.

4. Q4

   Moat — why can't the next team catch up?

   When several products do the same thing and one keeps winning, something is impeding replication. In AI, that "something" is usually a stack: head-start in capacity, proprietary or contracted data, deep workflow integration, brand trust in a regulated domain, or a developer ecosystem with switching cost. Single-layer moats erode fast. Composite ones — three of these at once — are what hold for a decade.