Eighty percent of enterprise revenue from one percent of customers. That is what Ramp's data showed for OpenAI and Anthropic, a concentration the analysis described as unseen in any other software category. Most product teams would call that a triumph of focus. The same shape can erase a quarter in a single week.
The spread underneath the average is the part worth staring at. The median business spent $11.95 per employee per year on AI. The top one percent of spenders came in at a median of $7,400 per employee, a gap of roughly six hundred times. Two medians from the same category, belonging to almost different species of buyer. Any report that blends them into one clean line describes neither one.
The pattern is not confined to model labs.
| Company | Concentration | What the number really means |
|---|---|---|
| OpenAI and Anthropic | 80% of enterprise revenue from 1% of customers | A few accounts set the whole trajectory |
| Anthropic | about $1.2B of $5B from Cursor and GitHub Copilot | Two integrations, roughly a quarter of revenue |
| Doximity | 84% of revenue from 116 clients spending $500k+ | A narrow base of very large contracts |
| Nvidia | 44% of FY2027 revenue from three customers | Demand concentrated at the top of the stack |
| AI software spend | $11.95 median versus $7,400 top percentile | Six-hundred-fold difference in intensity |
These are unrelated industries sharing one profile. A small set of accounts stops behaving like customers and starts behaving like infrastructure. Remove one and the model does not bend slightly; the whole structure reorganizes around the hole. Doximity's 116 clients at half a million dollars or more are not a segment in the usual sense. They are the floor the business stands on.
Here is the part that makes strategists uncomfortable. Concentration is neither a defect nor a virtue by itself, and diversification is not automatically the cure. Serving a tiny elite extraordinarily well is a coherent strategy, often the only one that makes sense when the product needs deep integration and long implementation. The danger is mistaking that choice for an accident. A team that never named its dependency cannot price it, insure it, or build around it.
The liability shows up in a specific form: correlation risk. A single ownership change at one lab can move billions between competitors in a week, with no change in product quality and no warning in the funnel. Standard churn models built on average behavior do not see this coming, because nothing about the average user changed. Only the whale moved. A dashboard that tracks retention percentage will report calm right up to the cliff.
Power users also resist the obvious lever. Pendo's data put new-feature adoption among power users at just twenty-four percent without a campaign. The heaviest users are not the easiest to move; familiarity breeds a stable routine that treats a new button as friction rather than promise. Sending more announcements to the top cohort often produces less change than sending fewer, better-timed ones. Attention is not the same as influence, and volume is not the same as relevance.
For a host platform, the practical response is governance rather than worship. A runtime that vets and distributes third-party services should know which services carry a disproportionate share of activity, and which owners sit behind them. That is not an argument for starving the winners. It is an argument for seeing the dependency early enough to decide whether to deepen it, duplicate it, or make the relationship less fragile through portable data and shared identity.
A shared event contract is what makes that visibility possible. The same schema that records a service's transactions also records how many distinct buyers produced them, so a host can tell the difference between genuine breadth and one very enthusiastic merchant. Concentration becomes a measured input instead of a quarterly surprise. The goal is not a flatter curve for its own sake. It is knowing which accounts the whole business is quietly standing on, and having a plan for the day one of them moves.
Sources
- OpenAI and Anthropic revenue concentration — the report behind the eighty-percent-from-one-percent figure.
- Ramp data on extreme concentration — spending spread and the six-hundred-fold gap.
- Concentration and competitive structure — why a narrow customer base shifts bargaining power.
- Customer concentration in enterprise software — a second angle on the same dataset and its caveats.