Job destroyer? Here’s what you need to know about AI and labor markets

History tells us that technological regime shifts are often associated with job obsolescence. In past regime changes over years and decades, an average of 200,000 to 300,000 workers were displaced per year.

The rise of personal computers in the 1980s and the internet in the 1990s deeply affected the labor market, most visibly in businesses such as print media, brick and mortar retail and advertising. Economists estimate that the rise of the internet alone displaced ~6 million workers.³

Labor market adjustment usually occurs through a combination of task re‑composition, sectoral expansion, geographic mobility, and new firm formation. The internet may have displaced 3.5 million workers, but it also contributed to nearly 20 million new jobs that would have been hard to imagine in the early 1990s (for example, e-commerce logistics and digital marketing).⁴

This time is different, the AI maximalists argue, because the technology is improving at such a furious pace that labor markets will be unable to recalibrate.

We disagree. We believe that the three constraints we mentioned—the capability of AI models, the demands of physical infrastructure investment and regulatory, organizational and social-political resistance—will moderate the speed of the transition, enabling the technology’s benefits to outweigh its costs for investors.

Constraint #1: Model capability: Improving fast, not human yet

Turning to the first constraint, model capability, we acknowledge what every large language model (LLM) user knows: AI model performance continues to improve rapidly. Anthropic’s latest LLM, Claude Opus 4.6, can complete tasks that take human experts over 12 hours, a more than 2x improvement since December 2025. Similarly, its “Claude Code” tool, which effectively turns plain language instructions into code, is growing exponentially. Github Copilot now writes nearly half of all code written by developers who actively use the tool.⁵

Agentic AI, where AI tools can complete tasks entirely on their own, is here. For example, C.H. Robinson, a transportation logistics company, has developed AI agents that can deliver the same price quotes that took human workers 15 minutes in 30 seconds. The firm’s employee headcount has fallen around 10% a year since 2022.⁶

But a job is a portfolio of tasks, not a capability benchmark. If knowledge work were only about processing tokens, the GPU (graphics processing unit, a computer chip that helps drive AI models) would have an insurmountable advantage.⁷

Knowledge workers reason. They act on vague or incomplete instruction. They use training, social intuition, emotional intelligence, pattern recognition and institutional knowledge to decide whether to apologize, troubleshoot or escalate an issue. Reasoning AI models try to replicate human cognition by retrieving historical data and iterating different approaches before deciding on the best one. Their track record is mixed.

Reasoning models are improving rapidly, but their gains look decidedly less impressive when greater success rates are required.

Consider Anthropic’s Claude Opus 4.6. When attempting a task that would take a human expert two hours, Claude can complete it 80% of the time. That success rate falls to 50% when Claude attempts more complex tasks that would take a human expert 12 hours. In other words, if you need AI to complete a task nearly all of the time, that task will need to be quite simple. That limitation may partially explain why Anthropic is seeing less task displacement than what the models could theoretically achieve.

Market pricing for both human labor and GPUs provide perhaps the most compelling evidence that AI models are still unable to outperform a knowledge worker.

A white-collar worker who makes $75,000 per year costs roughly $50/hour; depending on the workload and type of chip, a GPU can rent for $2.50/hour (a 20:1 ratio).⁸ Employers thus have a basic incentive to switch from human labor to AI labor. If AI were already capable of replacing knowledge workers at scale, it would have done so.

So far, the evidence suggests that task disruption is much more widespread than job disruption. That provides a positive signal for productivity.

Constraint #2: Physical infrastructure: Massive capex spend, huge power demand

Believing in mass labor displacement requires believing in a profound (and fast) transformation of computing and power infrastructure. The transformation may come, but it will take time.

Artificial intelligence demands massive and ongoing investment in its infrastructure. A semiconductor server (GPUs and memory) is housed in a data center that must be powered, cooled and connected to the broader network. So how many GPUs would we need to replace 10 million people?

If you assume one GPU could perfectly replicate one hour of human labor, the economy would only need 2.4 million GPUs to replace 10 million workers.⁹ However, the current 20:1 ratio implied by knowledge worker wages and GPU rental rates would imply a GPU fleet of nearly 50 million would be required to replace the hours worked by 10 million people.

Epoch AI, a research firm, estimates that there are 24 frontier data centers that are either operational, under construction or planned in the United States.¹⁰By the end of 2028, Epoch AI predicts that the U.S. frontier computing fleet will be a little over 25 million “H100” equivalent chips.¹¹ For the 25 million planned chips to replace 10 million workers—as Amodei and others have discussed—the chips would need to perfectly replicate one hour of human work in a little over 10 hours.

If, on the other hand, we assume that the current exponential growth in chip sales continues, hyperscalers and AI labs would own nearly 75 million H100-equivalent chips by 2028.¹² This would surpass the amount needed to displace 10 million workers at the current 20:1 ratio, but that assumes no computing power will be used by consumers or to train new models. The assumption of exponential growth is hard to square with current price/earnings multiples for major semiconductor manufacturers. Investors seem to be pricing in a decline in sales and earnings over the next few years.

Let’s assume only one third of the computing fleet will be allocated to labor displacement (as is currently the case).¹³ In that scenario, the chips would need to replicate one hour of human labor in anywhere from 10 hours to a little under 3.5 hours (depending on which GPU growth assumption used). That would represent a major improvement over the current 20:1 state implied by GPU rental rates.

To believe in massive labor market disruption, you not only have to believe in continued improvement in both model and semiconductor performance, but also a multi-year infrastructure investment cycle.

We can see the scale of infrastructure investment in the largest frontier data center in the pipeline: Microsoft’s Fairwater facility in Wisconsin. Construction began in early 2026 and the data center is scheduled to open in the fall of 2027. The facility will reportedly cost over $100 billion, house more than 5 million H100 equivalent semiconductors and draw more power than the city of Los Angeles.¹⁴

Demand for power is one of the many physical bottlenecks in the AI infrastructure ecosystem. Taiwan Semiconductor Manufacturing Company (TSMC), which produces over 90% of leading-edge chips, has only just started to increase its capacity to meet surging demand.¹⁵ On the energy front, the wait times for grid interconnection are between 3–5 years, and the backlog for “behind the meter” generation options from players such as GE Vernova and Caterpillar are between 6–7 years.¹⁶ Those companies, remembering the fallout from the tech bust of the early 2000s, will likely hesitate to scale up their own capacity.

If it is indeed possible to replicate human cognition, what will it take? The intellectual and physical computing requirements (semiconductor and model design, energy generation and transmission, memory bandwidth, silicon packaging, networking equipment, data center cooling systems, etc.) will be tremendous. 

That’s a key reason we find investment potential among the companies that will be building out AI’s physical infrastructure. In fact, based on metrics such as profit margins and revenue growth rates, they are currently some of the most attractive opportunities we see across markets.

Constraint #3: Institutions: Regulatory, organizational and social-political resistance

Perhaps the institutional limits to the AI transition will prove the most powerful constraint to AI disruption. These limits include regulatory and legal requirements, organizational decision making and implementation, and the potential for social and political backlash.

Regulatory bodies that oversee healthcare, finance, legal, education and government sectors already impose significant friction on AI deployment. The Food and Drug Administration takes 12–24 months at a minimum to approve diagnostic tools driven by artificial intelligence.¹⁷ Financial services regulators such as the Federal Reserve and Office of the Comptroller provide clear guidelines on how AI models can be used in decision-making.¹⁸

Concerns about accountability and liability are growing across sectors and industries. Human auditors must understand and evaluate the methods that models use to generate their outcomes.

Attorneys must independently assess and defend legal advice. Accountants must sign and certify financial results. Government employees must justify decisions within administrative and legal frameworks.

These requirements suggest that AI will be deployed as decision support rather than decision replacement. Future regulatory regimes and norms may eventually accept autonomous AI outcomes in areas where accountability currently resides with humans, but the timeline is likely measured in years. 

Corporate systems, legal liability

Regulatory compliance is an external factor that limits institutional adoption. The internal factors could be even stronger. Enterprises take quarters if not years to redesign workflows, operating structures and decision-making hierarchies. Management teams will likely be conservative about automating workflows until they can understand the potential liability for AI errors. If an AI agent gives poor investment advice in a fiduciary context, who is liable?

Corporate studies of tech adoption suggest that it takes 18–36 months from pilots to scaled deployment, and that is for more mundane tech solutions such as sales management software.¹⁹ Finally, data readiness and integration present a material challenge for enterprises that currently house records across multiple legacy systems and formats.

Policy choices: Who captures growth?

Finally, social and political resistance to AI could prove to be both potent and destabilizing.

Policymakers and citizens will be asking the same difficult question: If technology sparks economic growth, who captures that growth? To rephrase the question in the language of economists: Will capital’s share of GDP rise even further relative to labor?

More highly paid, highly educated workers could find themselves most at risk of AI displacement. Many are politically engaged. Today U.S. politicians on both ends of the spectrum, from Vermont’s Independent Senator Bernie Sanders on the left to Missouri’s Republican Senator Josh Hawley on the right, are sounding the alarm on AI. (Hawley has positioned himself as an “Anti-AI” presidential candidate in 2028.)²⁰

Lessons of history: Resistance to new technology is an old story

Periodically over the last century, economists and futurists have argued that innovation will lead to human obsolescence. Resistance to a groundbreaking technology, like AI, is nothing new.

In 1930, the influential British economist John Maynard Keynes warned that the economy would be “afflicted” with the disease of “technological unemployment.”²¹ In 1964, an independent commission urged U.S. President Lyndon Johnson to adopt universal basic income to mitigate the impact of technological change.²² In 1983, Nobel Laureate Wassily Leontief argued that human labor would be eliminated as horses were when the tractor appeared on the scene.²³

If investors had taken these warnings as signals to sell risk assets, they would have made a catastrophic error.

What do these warnings miss? They fail to account for the jobs that are created when technological change spurs innovation. Tasks and jobs can become obsolete—that’s easy to imagine. It is much more difficult to envision the new sectors and industries that will require high value human labor in the future.

Creating new sectors and markets

Artificial intelligence is reducing the constraint of expertise. This will likely lead to new sectors and addressable markets that are very difficult to foresee, even as AI disrupts what we currently view as valuable knowledge work.

A clerical worker at a bank before the introduction of mainframe computing likely did not think of their job as a routine cognitive task, though we view it that way today. How might that worker react if they could see a computer tracking transactions for the first time? How do we feel today when an AI model does in seconds what once took us hours?

Humbled, perhaps. But also hopeful.

AI could expand sectors like law in ways that are broadly beneficial. Some 92% of legal needs for low-income households go unmet and 40% of small businesses with a legal issue cannot afford an attorney.²⁴ If AI can lower the cost to serve these cohorts, the total revenue and employment for each sector could expand, not contract.

Finally, capital markets could provide the ultimate constraint to AI disruption. AI investments are ramping up because corporate executives and investors believe that the capex will lead to positive returns driven by productivity gains or new revenue streams. But if those fail to materialize because the costs to labor markets, consumer spending or business models outweigh the benefits, the capital needed to finance the physical infrastructure build-out will disappear.

Monitoring real time layoff announcements and labor market data for the most AI sensitive sectors will help us determine if the labor market fallout is happening faster than we expect.

For now, our own AI narrative is more optimistic than the consensus.²⁵ We think the prevailing market view underappreciates the most positive case for risk assets. Productivity and profits can rise, labor markets can adjust. The technology transition will disrupt, but it need not completely destroy.