Between 107 and 163 people. Revenues that have “significantly exceeded” 200 million dollars as of 2023, with 2024-2025 estimates ranging between 300 and 500 million. That’s a ratio on the order of 3 to 5 million dollars per employee — roughly twice the ratio of a well-performing large technology company, and far beyond that of a typical industrial SME.

Midjourney is not a statistical anomaly. It is the clearest signal of a structural transformation: the AI-native cognitive enterprise has no vocation to employ. It has a vocation to produce. And these two objectives, long conflated, are dissociating at a speed that economic institutions have not yet truly absorbed.

The Essentials

  • Midjourney has “significantly exceeded” $200M in revenue as of 2023; 2024-2025 estimates range between $300M and $500M, with a workforce of 107 to 163 employees, representing a ratio on the order of 3 to 5 million dollars per employee, according to Forbes and CB Insights
  • The median headcount of young American startups has significantly declined over two decades, a trend documented by researchers in entrepreneurship economics
  • This model is not unique to image generation: AI firms in legal services, automated accounting, and data analysis reproduce the same architecture with similar ratios
  • The central question is no longer that of destroyed employment, but that of the redistribution channel: if the company employs far fewer people at equivalent revenue, the wage bill ceases to be the principal vector for diffusing economic gains

3 to 5 Million per Head, or the End of an Implicit Convention

To understand what this figure represents, it must be compared to what we know. Google (Alphabet) generates approximately 1.91 to 2.11 million dollars in revenue per employee (2024-2025). Apple exceeds 2.5 million. Performances that, when announced, already looked exceptional within industrial capitalism. Midjourney surpasses them by a factor of approximately 2.

This is not a matter of exceptional talent or technological luck. It is a matter of architecture. Midjourney is built around a diffusion model trained on billions of images, accessible via monthly subscription, without a sales force, without massive support, without human infrastructure proportional to revenue. The software does the work that, in a classical company, would be done by creative teams, technical teams, sales teams.

The implicit convention of twentieth-century capitalism could be summed up in one sentence: growth creates jobs. The more a company prospers, the more it hires, the more it distributes salaries, the more prosperity spreads. This mechanism is not dead — it still functions in human-intensive service sectors, in construction, in healthcare, in education. But in the cognitive economy, it is breaking down.

NYU Data: A Trend, Not an Accident

The Midjourney case would be anecdotal if it remained isolated. It does not. J.P. Eggers, professor at the Stern School of Business at NYU, has documented the evolution of median headcount among young American enterprises over two decades. The result is striking: the starting headcount of American startups has significantly diminished, moving from much larger teams in the 2000s to clearly more reduced structures today at the time of their initial funding round.

This observation captures something precise: it is not about size at maturity, but about initial architecture. The previous generation of founders built teams to build products. The current generation orchestrates agents to deliver services. The work has not disappeared — it has been externalized to the model, to the cloud, to APIs, to systems that bill for usage and do not appear in any HR registry.

The acceleration of AI in 2025 has made this movement visible in the short term. What models did not know how to do in January 2025 — write production code, analyze contracts, produce commercial-quality visuals — they were doing routinely nine months later. Each new capability is a layer of cognitive work that ceases to require recruitment.

Sectors Where the Ratio Reproduces

Image generation is the most spectacular case because the product is immediate, visible, and directly monetizable. But the same ratio is building itself in other sectors.

In law, automated document review platforms process volumes that ten years earlier would have had to be entrusted to teams of junior analysts. In accounting, automated closing tools reduce to a few hours processes that once occupied entire teams at quarter-end. In medical diagnosis, radiological image-reading systems reach levels of precision comparable to specialists for a fraction of the fixed cost.

This is not a sectoral trend. This is a transversal recomposition of the relationship between revenue and wage bill. Acemoglu and Johnson, in their analysis of the distribution of technological gains, posed the question already in comparable terms: technology does not spontaneously redistribute its benefits — it captures them where architectural choices permit. Midjourney did not choose its architecture by ideology. It is the optimal structure for its value creation.

What is changing here is scale. Historical examples of high productivity per employee — investment banks, certain investment funds — remained sectorially circumscribed and required considerable barriers to entry. Generative AI lowers these barriers. A team of 5 engineers can today build a product that, in 2015, would have required 50 people.

What the Wage Bill Redistributed, and What Replaces It

The issue is not sentimental. It is structural. The tax and social systems of industrial democracies were designed around a hypothesis: wealth produced transits through wages before reaching households and public coffers. VAT strikes consumption, corporate income tax strikes profits, but it is the social contribution on wages that finances in large part protection against unemployment, illness, retirement.

A small-sized company that generates several hundred million dollars pays social contributions proportional to a few dozen or hundreds of salaries, not to the entirety of its revenues. The delta is considerable. And if the model generalizes — if one hundred, then one thousand companies reproduce this architecture in adjacent sectors — the funding base of social protection erodes without necessarily diminishing overall economic wealth.

This is not futuristic projection. The disconnect between productivity and wage bill is already observable in American macroeconomic data since the 2000s. AI does not invent this trend: it accelerates it and gives it a visibility it did not have when it was progressing gradually in industrial processes.

The resulting question is precise: if the company employs far fewer people at equivalent revenue, what channels ensure the diffusion of gains? Three leads coexist in the current debate, none being sufficient alone. The first is fiscal: tax the value added generated by automated systems, either via an extension of corporate income tax, or via a mechanism for taxing AI usage, as several tax administrations in Europe and Asia have been discussing since 2023. The second is patrimonial: if wealth transits less through wages and more through the value of digital assets, broadening the base of popular savings — citizen sovereign funds, participation in productivity gains — becomes an alternative redistribution path. The third is structural: direct AI toward augmenting workers in sectors with high social value (health, education, assistance), rather than toward their replacement in sectors with high cognitive productivity.

These three leads are not exclusive. They are only so if each problem is treated separately.

Tomorrow’s Reduced Teams: Irreversible or Reconfigurable?

Is the trajectory inevitable? In the short term, the answer is probably yes. The tools that enable a reduced team to deliver large-scale services are available, stable, and will continue to improve. The learning curve for entrepreneurs adopting them is rapid. There is no structural reason why the median startup headcount should increase again.

In the medium term, the question is more open. Technological history suggests two parallel scenarios: in sectors where AI performs on delimited tasks (vision, writing, document analysis), the compression of headcount continues. In sectors where value resides in human relationship, contextual judgment, legal responsibility, or interpersonal trust, AI augments workers without replacing them — and may even create jobs by making services accessible to populations that previously had no access to them.

A surgeon augmented by a robotic guidance system treats more patients. A teacher assisted by an adaptive tutor can personalize instruction to a class of 35 students as they would for 5. In these configurations, productivity per worker increases without overall employment declining — provided demand follows, that is, provided productivity gains actually serve to extend access to the service and not to reduce existing teams.

The difference between these two scenarios is not technological. It is political, in the proper sense of the term: what uses of AI a society decides to favor, to fund, to regulate. This choice will not be made spontaneously in the socially optimal direction, as several decades of innovation economics demonstrate.

What Midjourney Reveals About the Architecture of Future Gains

Midjourney is a profitable enterprise, founded without external venture capital, which never sought to raise public funds and built its growth on a direct subscription model. Its founder, David Holz, explicitly rejected the logic of growth-at-a-loss in favor of a company profitable from dollar one. This is a decision worth noting: in a sector dominated by the race to scale, a choice of deliberate austerity produced one of the highest productivity ratios in the digital economy.

This model is not universal. Foundational AI infrastructure — large language models, training systems, data centers — require billions of dollars in investment and thousands of engineers. The infrastructure layer remains intensely capitalistic and intensely human. It is the applicative layer that is compressing, because it can rely on infrastructure without replicating it.

This distinction is important for public policy. The question of employment poses itself differently depending on whether one is discussing investment in foundations (where jobs remain numerous and qualified) or applications built on top (where the team can remain reduced in size). Conflating the two levels produces faulty diagnoses and inadequate policies.

The stakes at ten years are as follows: the cognitive economy will continue to produce wealth with fewer employees. This trend is measurable, documented, and accelerating. The question is not to stop it — no one has the means or reasons to do so. The question is to build the channels by which this wealth, currently captured in the balance sheets of very small very productive teams, reaches the millions of people not part of these teams.

This channel existed. It was called the wage. It was called the social contribution. It was called the tax on work income. Others must be found — and designed before the gap becomes unbridgeable.

Sources

  1. Paul Baier, Forbes — “AI-Native Firms Lead in Revenue Per Employee” (March 2026): forbes.com
  2. J.P. Eggers, NYU Stern School of Business — data on median headcount of American startups, cited in Forbes / CB Insights
  3. Daron Acemoglu and Simon Johnson, Power and Progress (2023) — on the distribution of technological gains
  4. CB Insights — data on revenue/employee ratios of AI-native enterprises
  5. The Information — Confirmation of Midjourney revenues (March 2026): theinformation.com
  6. CB Insights — Midjourney $200M ARR Analysis: cbinsights.com
  7. Bullfincher — Revenue per employee Google/Alphabet: bullfincher.io
  8. Bullfincher — Revenue per employee Apple: bullfincher.io
  9. NYU Stern — Official J.P. Eggers Profile: stern.nyu.edu
  10. Sacra — Midjourney Report (December 2025): sacra.com
  11. Forbes — Midjourney Profile Page (secondary source): forbes.com