● AI Bubble Vs AGI Shock

Anthropic Employees’ 2030 AI Economic Scenario: What Really Matters Between the “AI Bubble” and the “Arrival of AGI”

The core point of this piece is not simply that “AI gets better.”

Inside Anthropic, employees see a much higher likelihood that AI could replace a significant share of human work by 2030, while the general public has not yet felt that pace.

What matters here is not AI performance alone, but the future outlook diverges completely depending on how often and how deeply AI is used.

In this article, centered on Anthropic’s report “The AI Expectation Gap”, we’ll summarize in a news-style format how AI investment, the labor market, productivity innovation, GDP growth, and the big tech competitive landscape could change.

1. The market is split in two: AI bubble or the beginning of AGI

In the current global economic outlook, views on AI are split extremely sharply.

On one side, some warn that if AI companies fail to deliver earnings as expected, the market could face a major correction.

In its October 13 World Economic Outlook, the IMF noted that financial markets could take a major hit if profitability expectations for AI weaken.

Simply put, the stock market has priced in AI investment optimism too quickly, and if actual profit generation fails to keep up, AI bubble concerns could intensify.

On the other side, some believe AI has already entered a turning point where it surpasses human-level intelligence.

Elon Musk has referred to 2026 as the year of the singularity, and Sam Altman has also made remarks implying that AGI is already near or effectively underway.

In other words, the market simultaneously holds both the AI bubble collapse thesis and the AGI arrival thesis.

But to determine which is more likely, we need to look at the perceptions of the people actually building AI, rather than the intuition of the general public.

That is where Anthropic’s report becomes important.

2. The “AI expectation gap” shown by Anthropic’s report

On October 16, Anthropic released a report titled “The AI Expectation Gap”.

The report compares responses to the same questions asked of 1,300 Anthropic employees and 10,000 U.S. adults as of August 2026.

The core takeaway is simple.

People who see AI up close believe its disruptive power by 2030 is far greater.

• Question 1: Could AI make a Nobel Prize-level discovery within four years?

  Only 39% of the general U.S. public said yes.

  By contrast, 91% of Anthropic employees said yes.

• Question 2: By 2030, what percentage of my work could AI perform?

  The average answer from general workers was 25%.

  The average answer from Anthropic employees was 75%.

This gap is not just simple optimism.

It could be argued that AI company employees exaggerate to boost their company’s value, but Anthropic has repeated similar forecasts for a long time.

In particular, Anthropic has pointed to the possibility of powerful AI arriving much earlier than Google DeepMind has.

3. Why is Anthropic so aggressive in its outlook?

In March 2025, Google DeepMind CEO Demis Hassabis said it could take five to ten years for powerful AI to appear.

By contrast, around the same time, Anthropic stated in a submission to the White House that powerful AI could emerge by late 2026 or early 2027.

At the time, that was a very radical forecast.

Anthropic CEO Dario Amodei went further, making very strong claims about developer work.

He said that within three to six months, AI would write 90% of code, and within a year it would effectively write nearly all code.

The developer community reacted immediately.

• “AI-written code has too many mistakes.”

• “It takes longer to review than to write.”

• “It still can’t be used in real work.”

But inside Anthropic, something else was already happening.

Before the launch of Claude Code, the share of code written by AI inside Anthropic was reportedly in the single digits.

But as of May this year, Claude is reportedly writing more than 80% of the code.

In other words, what outside developers dismissed as “nonsense” has already become reality inside Anthropic.

4. The moment AI becomes better than experts: humans can become the bottleneck

A particularly important example in the original text is the accounting workflow comparison.

One setup involved accountants and AI working together, while the other used only AI.

Most people naturally assume that if a human expert reviews AI output, the result should be better.

But the result was the opposite.

Humans became the bottleneck.

People incorrectly revised results that AI had handled correctly, lowering overall quality.

AI-only work was more accurate, and the time required dropped to less than one-tenth.

This case has major implications for the labor market.

Until now, we have viewed AI only as a “tool.”

But in some knowledge-work domains, AI is already entering a stage where it can make faster and more accurate judgments than the average expert.

Just as humans can’t easily beat calculators in mental arithmetic, a similar dynamic could emerge in intellectual and cognitive labor.

5. AI expanding beyond coding into security, mathematics, and research

AI’s impact does not stop at developer work.

Claude Opus, released in April, is said to have identified more than 10,000 security vulnerabilities.

This has enormous implications for cybersecurity.

It could be similar to having tens of millions of faster, more persistent hackers moving at once.

OpenAI is also said to have run an experiment using 10,000 agents for 88 hours to prove complex mathematical problems.

Because of this trend, more people are calling AI doctoral-level intelligence.

Dario Amodei’s description of data centers as a “nation of geniuses” reflects this context.

Here is the key point.

AI companies are not just chatbot sellers.

They are becoming companies that commercialize intelligence itself.

From that perspective, it becomes easier to understand why the valuations of big tech and AI startups no longer fit the logic of traditional industries.

6. Are Anthropic employees simply optimists? In fact, they are more sober about current AI limits

An interesting point is that Anthropic employees are actually more conservative than the general public when it comes to current AI capability.

When asked whether AI can make a Nobel Prize-level discovery right now, 12% of the general public said yes.

By contrast, only 4% of Anthropic employees said yes.

In other words, Anthropic employees are not people who blindly overestimate current AI.

Rather, they understand more precisely what AI still cannot do.

What they see differently is the pace of progress.

People who use AI deeply feel a huge difference between AI from a year ago and AI today.

They even feel the difference between a model from one month ago and the current version.

People who use Claude Code or Codex at the scale of billions of tokens each month immediately notice productivity differences even with small version changes.

By contrast, people who only ask a few questions occasionally may say:

“I tried it, and it didn’t seem that different.”

But that is more likely because they haven’t used AI deeply enough, not because AI is underwhelming.

7. AI usage volume shapes perceptions of the future

The Anthropic report also shows how forecasts change depending on AI usage experience.

The share of people who believe AI will be able to perform seven or more out of eight tasks by 2030 varied greatly by usage experience.

• People who have never used AI: 28%

• General public average: 41%

• Regular AI users: 56%

• Anthropic employees: 96%

Perceptions of work speed also differed.

When asked whether AI can finish work within one-fifth of the usual time, 26% of general AI users said yes.

Among Anthropic employees, 59% said yes.

This is extremely important.

The gap between people who use AI and people who don’t is becoming smaller than the gap between people who use AI deeply and people who use it superficially.

This gap could affect not only personal productivity, but also wages, promotions, entrepreneurship, and investment returns.

8. AI polarization has already begun: $13-a-month users and $7,500-a-month users

Even among people using AI in companies, the usage gap is huge.

At the median, monthly AI spending is only about $13.

That is roughly 15,000 won in Korean currency.

But the top 1% of users spend about $7,500 per month on AI, or more than 10 million won.

More extremely, the top 0.001% developing AI at OpenAI, Anthropic, and big tech are said to use more than 100 million tokens per month.

This gap is not just a difference in usage volume.

In the economy ahead, AI usage volume could equal the scale of productivity innovation.

People and companies that use AI more learn faster, run more experiments, and produce higher-quality results.

Conversely, individuals and companies that barely use AI will inevitably lose competitiveness over time.

9. Inside Anthropic, AI labor time has already surpassed human labor time

According to the original text, as of August 2026, the total time AI agents worked inside Anthropic was six times the total time employees worked.

Anthropic believes that 51% of the work involved in building advanced AI models is already being done by AI.

And by the end of 2030, it forecasts that this share could reach 95%.

This is highly significant.

If even the work of making AI is done by AI, the speed of technological progress could shift from linear to accelerating.

From that point on, deploying thousands or tens of thousands of AI agents in parallel becomes far more powerful than adding one human at a time.

Even so, Anthropic is not yet conducting large-scale layoffs.

In fact, its employee count has tripled over the past year.

But the change is showing up in new hiring.

The share of entry-level roles among new technical hires fell from 27% to 9%.

Simply put, companies are increasingly coming to the conclusion: “Rather than hiring and training juniors, let’s use AI.”

In Silicon Valley, the shift toward hiring AI instead of junior developers is already becoming reality.

10. A shocking scenario in which 12.2% unemployment and 17% GDP growth are both possible in 2030

The most shocking part of the Anthropic report is the separation between the labor market and economic growth.

The current U.S. unemployment rate is around 4%.

But U.S. workers who believe AI will replace more than 80% of their job expect unemployment in 2030 to average 12.2%.

12.2% is far higher than during the 2008 global financial crisis.

Normally, if unemployment rises that much, consumption falls, corporate earnings weaken, and the economy slows.

But when Anthropic ran the model again reflecting that unemployment rate, the economic growth rate reportedly reached as high as 17%.

Considering that current U.S. GDP growth is roughly 3% to 4%, that is an enormous number.

The core takeaway of this scenario is blunt.

Even if AI replaces humans, the productivity gains could be so large that total GDP might actually grow explosively.

Of course, this is an extremely uncomfortable outlook.

The whole economy may grow, while individual jobs disappear.

That is likely to become the most important social conflict in the AI economy ahead.

11. AI companies’ revenue model is different from traditional companies: they sell not products, but “intelligence”

The number of knowledge workers worldwide is estimated at about 900 million.

If you multiply that by the average wage, the total annual wage bill for global knowledge workers comes to about 8 quadrillion won.

If Anthropic replaced just 1% of that market, annual revenue could reach roughly 800 trillion won.

That is a scale that could instantly make it the world’s largest revenue company.

Even more important is profitability.

According to a recent SemiAnalysis report, Anthropic’s gross margin on inference services is estimated to exceed 90%.

There are also projections that operating margins could exceed 40%, even after computing and marketing costs.

From this perspective, talk of Anthropic or OpenAI reaching a $2 trillion valuation may not be an exaggeration.

This is also why some research institutions argue that by 2027 they could surpass Nvidia’s market capitalization.

If Nvidia sells the AI infrastructure, Anthropic and OpenAI are selling the intelligence that runs on top of that infrastructure.

12. The real core point that many news reports miss: the issue is not whether AI replaces jobs, but who captures the gains

Many news outlets and YouTube channels discuss whether AI will replace jobs.

But there is another, more important question.

Who will capture the gains from the productivity revolution created by AI?

In a paper Anthropic released in September, it discusses an extreme scenario.

In the past, labor may have captured about 60% of national income, but if AI cheaply replaces human labor, that share could fall to 45%.

In other words, GDP growth could be high while the worker’s share declines.

In that case, corporate profits would surge.

In particular, AI model companies like OpenAI and Anthropic, as well as big tech firms with cloud infrastructure, could be the biggest winners.

By contrast, the bargaining power of human labor could weaken.

That is the truly frightening part.

The economy could grow, but households could become poorer.

Companies could make money while consumers have less room to spend.

Human beings are still the ones raising children, buying homes, paying education costs, and sustaining society, yet a paradox could emerge in which the human share of income declines.

13. Not everything has materialized yet: technological progress and adoption speed are different

That said, at present there is not yet clear economic evidence of large-scale unemployment caused by AI.

Technological progress is very fast, but the speed at which companies and society actually adopt AI is slower.

Most people still have not used AI properly.

There are even surveys showing that nearly half the world’s population has never used AI at all.

So it would be excessive to say all jobs will disappear immediately.

But it is also dangerous to feel reassured just because “nothing major has happened yet.”

AI does not spread slowly like older technologies; in certain industries and jobs, it can suddenly cross a tipping point.

In particular, knowledge-work areas such as coding, accounting, research, translation, legal review, security analysis, and data analysis are likely to be affected first.

14. Strategy for individuals and investors: how to survive the AI era

Whether this scenario turns out to be right or wrong, the strategy needed now for individuals and investors is relatively clear.

• First, pay attention to companies building AI.

  AI model companies, cloud infrastructure companies, semiconductor companies, and data center-related companies are likely to remain core pillars.

  From an AI investment perspective, do not just look at revenue growth; look at who controls the bottleneck in intelligence production.

• Second, keep building your ability to use AI.

  Going forward, the important question is likely to be less “Can you use AI?” and more “How deeply have you integrated AI into your work?”

  People who redesign their entire workflow around AI, not just monthly chatbot users, may create a much larger productivity gap.

• Third, staying at junior-level tasks is risky.

  The areas AI is most likely to replace first are repetitive, reviewable entry-level tasks.

  Therefore, you need to build higher-level capabilities such as problem definition, result verification, strategic judgment, and organizational decision-making.

• Fourth, watch labor-market changes and policy risk together.

  If unemployment rises while GDP growth remains high, governments will have no choice but to consider policies such as taxes, universal basic income, labor regulation, and taxation of AI companies.

  This shift could directly affect the stock market and the global economic outlook.

15. Conclusion: the core question for 2030 is not “Will AI replace humans?”

The most important question in the 2030 AI scenario is not simply, “Will AI take people’s jobs?”

The real question is this.

How will individuals, companies, governments, and society divide the gains from the productivity created by AI?

Anthropic employees do not think AI can do everything right now.

But because they observe AI progress most closely, they believe 2030 could bring a world completely different from today.

The AI bubble debate matters too.

But even more important is the speed at which AI actually changes the economic structure.

The labor market, GDP growth, corporate profit margins, personal productivity, and global capital flows could all be reshaped at once.

So the right attitude now is neither fear nor blind optimism.

Understand the companies building AI, build your ability to use AI, and first examine how your role will be redesigned by AI.

The difference between people who prepare for that and people who do not could become far larger than expected by 2030.

< Summary >

According to Anthropic’s report, there is a large gap between the general public and AI developers in their 2030 outlook for AI.

The general public expected AI to perform 25% of work, while Anthropic employees believed it could reach 75%.

For the question of whether AI could make a Nobel Prize-level discovery within four years, the general public answered 39%, while Anthropic employees answered 91%.

The core point is not AI performance alone, but the gap in pace of progress and usage volume.

People who use AI deeply are far more likely to see major changes by 2030.

By 2030, even if unemployment rises sharply, a scenario is possible in which GDP growth remains high thanks to AI productivity.

The most important issue is whether the wealth created by AI will go to workers or become concentrated in AI companies and big tech.

Individuals should build AI usage skills, and investors should pay attention to the business models of AI companies that sell intelligence.

[Related Articles…]

• Core Takeaway on AI Investment and the Global Economic Outlook

• Analysis of Labor Market Change and Productivity Innovation in the AI Era

*Source: [ 월텍남 – 월스트리트 테크남 ]

– 앤트로픽 직원들이 본 충격적인 2030년 상황..ㄷㄷ

---