AI Boom, Deepens Polarization

● AI Agent Boom Deepens Corporate Polarization

“40–70 billion won in revenue per employee” The secret of AI-native companies… and how corporate “polarization” grows even more

One-line summary of the core point of this article

As AI agents (outsourcing-style AI) begin to take off, there’s a view that while only a small number of AI companies see their revenue surge, ordinary companies (especially the middle layer) may move their “operating body” more slowly—widening the gap even further.

In particular, the core point is that AI-native companies greatly outpace general SaaS and legacy companies in revenue per employee, and in the process, even profitable companies are limited due to the compute cost and infrastructure war.

The “real important points” included in this article are five below.

AI-native revenue per employee: Estimated to be at a level of 40–70 billion won (with mentions such as OpenAI, Anthropic, Cursor, etc.)
Persistent losses rather than profits: Even with high revenue, losses continue due to compute costs
The differentiator isn’t “coding” but “operating approach”: Developers don’t just write code; the AI performs real work (prompts, orchestration)
Spread of solo entrepreneurship: Not fully solo, but a growing trend of “solo + α (agents/assistants)”
Worsening corporate polarization: Large companies can endure thanks to assets, while the middle layer may become the most vulnerable

1) The AI agent era: A view that “the whole game changed since late last year”

There’s a common message that comes up in recent interviews and expert commentary.

After agentic AI like Claude Code and Open Clo appeared at the end of last year, about 3–4 months passed— evaluations repeat that “that was truly the big moment.”

Even non-developers can switch tasks in a way that effectively “replaces coding”
Using agents means one person can handle what is essentially “the workload of multiple people”
So the momentum for both solo startups and team startups speeds up

The result of this trend is summarized as moving not just toward “a fascinating technology,” but toward changing organizational operations, workforce structure, and the way products are produced.

2) Revenue of AI-native companies: Does revenue per employee surpass Google, Meta, and Apple?

The most striking data is “revenue per employee.”

The original text calculated estimates by looking at multiple AI-native companies (OpenAI, Anthropic, CursorAI, Midjourney, Lovable/Pleplexity, etc.).

OpenAI: revenue per employee of 5.5 million dollars → roughly 7–8 billion won
Anthropic: 4.4 million dollars → around 6 billion won
Midjourney: 3.8 million dollars → around 4.5–5 billion won
CursorAI: 3.3 million dollars → around 4 billion won
Pleplexity also has a minimum estimate of 800,000 dollars (around about 1 billion won)

On the other hand, in the context described, revenue per employee at existing SaaS/legacy companies is often around 200–300 million won, and even the best performers are in the several hundred million won range.

So the conclusion is:

AI-native revenue per employee is at least 10x–20x
Depending on conditions, it can be interpreted as up to 40x

What’s important here is the claim that it’s not just because “the technology is better,” but because of the difference in the 3) operating approach below.

3) Still, why isn’t “profitability” common? Compute costs and the infrastructure war

There’s a saying that even with big revenue, most companies remain unprofitable.

OpenAI, Anthropic, Cursor, etc. are loss-making
Cases mentioned as profitable are limited—roughly only one, Midjourney

There are three main reasons commonly cited.

Computing costs: costs keep occurring throughout training and inference
Investments to preempt infrastructure for the next decade: large upfront spending on semiconductors, data centers, and infrastructure
Labor costs R&D expenses: the unit cost of AI talent is high

On top of that, some model-based business models even discuss scenarios where API costs exceed revenue (CursorAI is mentioned as a representative example).

In other words, the current stage is:

Optimization to reduce cost is still underway
A funding/investment stage in which pricing and monetization approaches of business models continue to evolve

From a consumer standpoint, there may be models where “the more you use, the worse it gets,” so it naturally follows that later there could be price adjustments or usage limits.

4) The differentiator of AI-native: “Prompt + orchestration” becomes a person’s new kind of labor

This is where the real secret (= operating approach) starts.

In the original text, it draws a line to clarify that calling something an AI-native company doesn’t mean they all “only work with AI.”

Companies like Anthropic are said to be larger in scale, with organizations similar to the existing model
Companies like Midjourney are mentioned as being relatively small, with AI-native methods stronger there

The core is that an AI-native organization is defined like this.

The members’ mode changes from “practitioners of real work” to “roles that command and operate AI.”

There’s an analogy.

It’s like becoming a “Roman farm owner,” where the AI does the work and I give instructions
Or like an “overseer from the Industrial Revolution,” where you manage the outputs created by AI

So differentiation moves from “the developer’s ability to write code” to the following.

Prompting ability: instructing the desired outcome precisely
Orchestration ability: coordinating multiple agents without conflict
Data-driven decision-making: the AI keeps accumulating the rationale and execution records for its judgments
Ethics/legal/strategy risks: AI is weak here, so final judgment by humans is needed

As a result, there are scenarios where meeting styles, decision-making, and hiring culture inside the company could also change.

Meetings: the AI creates “alternatives and scenarios,” and converges using data
Hiring: “the ability to instruct AI” matters more than experience/amount of knowledge

5) Why “solo entrepreneurship” is increasing: technology got easier, and what matters more is “task decomposition”

The original text also mentions the trend of more solo entrepreneurship with data.

Based on startup equity management platforms like Carta
The share of solo founders increases from 23.7% in 2019 to 36% in 2025

However, it can’t be concluded solely that it’s because of AI; the article adds an interpretation that “AI likely had an impact.”

And in practical terms, the realistic form of solo entrepreneurship is closer to solo + agents (or assistant/tools) rather than a completely “single-person” setup.

In the past, one person had to handle planning/operations/marketing all by themselves
With AI, development, content production, and operational support roles ease the “time constraint”

The examples in the original text are also pretty direct.

Headshop Pro: one person runs service → mentions monthly revenue of 300,000 dollars
Headleime: mentions recording monthly revenue of 1 million dollars in 8 months
A service built on Open Clo is framed as a “service built by oneself”
Moltbook also mentions a flow where one person creates and sells an SNS (acquired by Meta)
In a domestic example, ReleaseAI is mentioned as a two-person startup (summary AI) and that it grew in scale

6) The divide: “easy fields for AI startups” versus “still hard fields”

The original text lays out quite clearly the conditions where AI startups fit well.

Areas where physical assets/regulation are not entry barriers
Repetitive white-collar work
Bottleneck areas where there is plenty of manpower but productivity improvements don’t immediately follow

And it says “software + content” is close to the optimal combination.

AI is strong in software because there’s abundant training data and execution history
For content/news/translation/research/production, productivity increases by leveraging AI
Professional services are more likely in the long-tail/middle stage than in the “high-end stage”

As an example, it mentions legal and tax advisory areas where it’s hard for “small businesses/individuals” to hire experts on a constant basis.

Meanwhile, the weak areas for AI startups are summarized like this.

Trust-dependent areas (advanced legal, M&A, investments, etc.)
Areas where physical mastery is key (construction, beauty, crafts)
Industries where infrastructure/network matters (local wholesale/retail, real estate brokerage)
Areas requiring licensed and legally final judgment (medical, accounting/auditing, etc.)

So rather than “AI replaces everything,” the view is that things are likely to be reorganized starting from software and content, step by step.

7) “Big companies don’t collapse” vs “the middle layer is the most dangerous”

Even here, the perspectives split sharply.

The original text says that it’s difficult for a large company to make the kind of “conversion that overhauls a massive organization” with AI-native methods.

Because organizational legacy is so large, it’s hard to stop the system immediately and transition

Still, here’s why big companies don’t collapse:

Assets that AI can’t replace (brand trust, sales networks, regulatory/legal capabilities, infrastructure, etc.)
AI-native startups’ strengths are likely to first appear in primarily B2C/digital domains

However, the most dangerous area is identified as the “middle layer.”

Small/mid-sized IT and “middle” companies that lack both agility and resources to keep up with AI
There was a similar pattern in the Internet and mobile era too (some firms/middle markets were shaken first)
So polarization could worsen further

Additionally, it’s described not as the “middle layer disappearing completely,” but as a possibility that generational change will happen quickly.

8) VC (venture capital) gets shaken too: is investing less attractive for solo startups?

From an investment perspective, it turns to more realistic talk.

The share of solo entrepreneurship is increasing, but the share getting funding is still low (around 10% is mentioned)
VCs prefer teams so that an early MVP can be made “properly”

The core logic is this.

It’s possible for one person to build a product, but to make it properly—like a real pro—and increase the chance of being #1 (including marketing/operations), usually more resources are needed
So at the early stage, raising investment may become necessary
But solo/ultra-small teams may have a worse founder experience due to equity dilution and pressure to deliver short-term results

Even in Silicon Valley, there’s a brief talk like “the end of VC,” and the original text also mentions that this issue is spreading in Korea too.

That said, there’s a rebuttal that VC won’t completely disappear, and that roles like selecting “A-grade founders” and supporting founder growth (organizing growth teams such as marketing/legal) will become more important.

9) AI strategy of Korean big enterprises: from “HyperCLOVA X” to “summary + agent experience”

The original text connects also to the movements of Naver/Kakao.

Naver

Rather than actively pushing the old LM-centered strategy (such as HyperCLOVA X),
they are focusing in search results on “experiences” like AI briefings (summarized answers)
Changes like ending related search keywords can be interpreted as “answering so users don’t need to search more”
In the future, AI agents will perform tasks from searching to reservation and decision-making
Strategically, there’s a shift not to insist only on its own models (interpreted as that other models could also be used behind the scenes)

Kakao

A representative strategy is to embed ChatGPT into KakaoTalk (conversational touchpoint)
Extend by inserting its own/small on-device models like Kana into the conversation context to enable “recommendation, connection, and booking”
Also mentions aiming to build an AI agent platform (the dream of connecting with MCP)

In conclusion, even Korean big tech also needs to go toward “AI agents,” but the nuance is that transitions may not be smooth because strategies diverge or due to reasons like internal resources/model selection and regulatory/infrastructure costs.

10) Conclusion: The turning point of K-shaped growth—right now, the “felt gap” is the more dangerous one

At the wrap-up, the tone becomes quite practical.

In places where there’s infrastructure and preemption resources like semiconductors and AI, the chance of going toward the “upper part” of K-shaped growth is higher
Where that’s not possible, it can get pushed down to the lower part
So the gap grows between “some people transition too quickly” and “some people consume it only at the level of something novel”

And it’s seen that this gap isn’t just an issue of personal sentiment—it could lead to real-world outcomes in jobs, earnings, and employment.

There are also discussions that the government should increase entrepreneurship (to buffer unemployment rates, tax revenue, and jobs), but the original text also addresses the balance that “to avoid turning entrepreneurship into mass production of personal doom, there will need to be a culture like safety nets/education/coupons after failure.”

My most important one-line takeaway that differs from the original perspective (separate summary)

The real weapon of AI-native is not model performance, but the speed at which the organization changes (the ability to turn prompts and orchestration into a work language).

So, more than whether the company goes under right away, the biggest possibility is that if the “middle layer” misses the timing for a body-type transition, polarization accelerates.

SEO core keywords (naturally reflected in the text)

The core themes that connect naturally in this article are agentic AI, AI-native, compute costs, solo entrepreneurship, and the venture capital (VC) flow.

Main points intended to convey

The AI agent era changes the “operating approach to work,” beyond “automation of development.”
AI-native companies have high revenue per employee, but unprofitable structures are common due to compute costs and the infrastructure war.
Solo entrepreneurship is increasing, but the realistic form is “solo + agent” rather than a fully “solo” setup.
Corporate polarization can grow even more, and especially middle-layer companies are at risk.
In Korea, big tech is shifting strategy from search/conversational touchpoints toward AI briefings and agent experiences.

< Summary >

1) Since the end of last year, AI agents have become fully established, rapidly changing the landscape for work and entrepreneurship.

2) AI-native companies can show differences of 10–20x (up to 40x) versus general companies in estimated revenue per employee at a level of 40–70 billion won.

3) However, even with large revenue, most companies remain in a loss-making structure due to compute costs and infrastructure preemption investment.

4) The differentiator isn’t coding, but an organizational/operational approach to commanding AI through prompts and orchestration.

5) Solo entrepreneurship is increasing, but the success likelihood is higher in formats like “solo + agents (or a small team).”

6) While big companies can endure with assets, the middle-layer companies with insufficient agility and resources are most likely to be shaken.

[Related articles…]

*Source: [ 티타임즈TV ]

– 1인당 매출 수십억원은 기본, 이런 회사의 비결은?

NextGenInsight.Net