Andreessen Horowitz (a16z) has recently published a long-form article, “The ‘AI Job Apocalypse’ Is a Complete Fantasy,” directly targeting the panic narrative that “AI will create a permanent underclass” and that “white-collar jobs will be completely wiped out.” It argues this is not any brand-new insight, but rather an AI-wrapped version of the long-derided economic “lump-of-labor fallacy.”
The so-called lump-of-labor fallacy assumes there is a fixed number of jobs that must be done, so once machines, migration, outsourcing, or AI does more, humans must necessarily do less. a16z believes this premise fundamentally contradicts human needs, market dynamics, and the empirical record of economic history.
What humans want has never been fixed, and the market is not a zero-sum game. As productivity rises and costs fall, people typically do not end up with “nothing to do.” Instead, the saved time, resources, and labor get pushed into the next set of new demands, new industries, and new jobs.
AI makes intelligence cheaper
The article acknowledges that AI will eliminate some tasks and compress some positions—and that this may already be happening. But a16z pushes back on moving from “some jobs are replaced” straight to “the entire economy will experience permanent unemployment.”
The authors argue that what truly happens is not jobs disappearing, but the shape of the labor market being rewritten. Like every time a general-purpose technology has emerged in the past, AI will change what work involves and reallocate industry structures. But the ultimate effect of higher productivity should be to increase labor demand, because human capabilities become more valuable—not less valuable.
Cognitive costs are indeed collapsing. The charts show that from September 2023 to the end of 2025, the LLM “price relative to the intelligence index” has fallen clearly—and it is presented on a logarithmic scale, indicating rapid improvement in the relationship between model capability and usage cost. In other words, AI is making the kinds of analysis, writing, reasoning, coding, and knowledge processing that were once exclusive to human intelligence increasingly cheap.
The reasoning of AI doomsayers is: if AI can think for humans, the human moat disappears, and human economic value becomes zero. a16z counters this: the argument only sees that “AI can perform existing tasks,” while ignoring that “when a powerful input becomes cheaper, markets will create more demand.”
The article uses an energy analogy: when fossil fuels made energy cheap, stable, and available at scale, humans did not simply eliminate whalers and lumberjacks. They invented plastic, reorganized industrial systems, and created new consumer goods and new ways of living. a16z suggests AI may be the same kind of force. When AI takes on more cognitive burden, humans don’t stop thinking—they can tackle bigger, more complex problems that were previously unaffordable.
Labor markets keep changing: agriculture shrinks, while services, healthcare, business, and software emerge
a16z uses a long-term labor market structure chart to point out that the U.S. labor market has already undergone dramatic transformations over the past 170-plus years. Around 1850, agriculture accounted for a very high share of employment in the U.S.; in the modern era, agriculture’s employment share has nearly vanished at the bottom of the chart. Yet the U.S. has not suffered permanent unemployment—instead, it has grown manufacturing, construction, finance, business services, education, healthcare, leisure and entertainment, and all kinds of service industries.
This is the core historical basis a16z uses to rebut the AI doomsday narrative: every sector that once dominated the economy gives way to a larger, more complex follow-on sector. Fewer old jobs does not mean the total amount of work disappears; it means humans redirect the remaining productivity into new areas.
a16z notes that although today’s tech industry is powerful, looking at the history of the U.S. stock market shows that in the past, industries such as finance and real estate, transportation, energy, and materials have also occupied highly dominant positions during certain periods.
Tech’s dominance is currently high, but it is not the most extreme industrial concentration in history. The point of this chart is that markets keep swapping the starring role—technology is not the first sector to dominate the market, and it will not be the last.
Agricultural mechanization didn’t destroy the labor market—it released more labor
The article then uses agriculture as an example. In the early 20th century, about one-third of the U.S. employed population worked in agriculture; by 2017, that proportion had dropped to about 2%. If automation really caused permanent unemployment, then tractors and agricultural mechanization should have destroyed the labor market long ago.
But the actual outcome is the opposite. Agricultural output rose dramatically, enabling the global population to keep expanding. And the labor released from agriculture moved into factories, stores, offices, hospitals, laboratories, service industries—and ultimately into software and the knowledge economy.
The chart a16z pairs with—“More Productive Farming Led to (A Lot) More Workers”—looks at farm product prices alongside world population. Between 1913 and 2024, the real prices of corn, wheat, rice, and more have trended downward for the long run, while the world population has increased sharply. This indicates that rising productivity did not make people “no longer need to work.” Instead, it lowered basic living costs, making it possible for more people, more industries, and more new job categories to emerge.
Electrification wasn’t just swapping energy sources—it redesigned factories and homes
The second historical case is electrification. a16z points out that electricity was not simply replacing one power source with another—it fundamentally changed factory structures: from centralized shafts and belt drives to individual motors driving each machine. This forced factories to redesign workflows, and it also created entirely new industrial and consumer products.
General-purpose technologies do not unleash all their productivity at once when they are introduced. In the 1820s to 1840s, Faraday and Henry established the principles of electricity. In 1879, Edison advanced commercial electric lighting. In the 1880s, Tesla developed alternating-current motors. In the 1900s, unit-drive began to spread. Only in the 1920s, after single-motor setups reshaped factories, did labor productivity accelerate noticeably. The impact of new technology needs time, and AI may be in a similar early diffusion phase.
When technology makes a product cheaper, markets usually don’t shrink—they expand. The same logic applies to automobiles. Between 1900 and 1925, the real price of new cars in the U.S. fell sharply, while annual car production and employment in the automotive industry increased significantly. Making cars cheaper didn’t make the automotive industry disappear—it made the automotive industry be born.
Spreadsheets didn’t eliminate finance jobs—they created the FP&A era
a16z also pushes the argument into the white-collar world. Spreadsheet tools like VisiCalc and Excel did indeed automate a large amount of manual bookkeeping, spreadsheet calculations, and data processing work. But they did not make finance work disappear; instead, they enabled a rapid expansion of higher-level financial analysis, accounting audits, and FP&A (financial planning and analysis) roles.
From 1970 to 2020, the number of bookkeepers and accounting clerks in the U.S. first rose and then fell. Over the same period, accountants & auditors continued to increase, and financial analysts grew substantially starting after the 1980s. a16z summarizes it as: the U.S. roughly lost 1 million bookkeeping staff, but added about 1.5 million finance analysts.
This case is especially crucial for the AI debate. Because the impact of today’s AI on white-collar work is likely to resemble the impact of spreadsheets on accounting and finance: it will replace low-level, repetitive, formatting-heavy tasks, while also creating more work that requires judgment, strategy, integration, and business understanding.
Excess productivity also creates entirely new service industries: tutoring, pet care, nail salons, and sports
a16z further argues that labor surplus created by productivity gains doesn’t always flow only to adjacent areas of automation. Sometimes it forms new jobs in entirely different industries.
Between 1990 and 2025, several categories of service-industry jobs in the U.S. grew quickly, including athletes, coaches, referees and related work, exam preparation and tutoring, pet care, nail salons, and more. These industries were not created because a particular machine directly produced them. Rather, as overall income rose, leisure increased, consumption upgraded, and labor could be reallocated, previously limited demand became a large market.
This is also one of a16z’s responses to the claim that “AI will only make a few people super rich, while others get left behind.” Even if productivity gains initially make some people extremely wealthy, they will spend the money, which creates new service demand. The article concedes that how to judge whether service industries are “serving the rich” may involve value judgment. But from a labor market perspective, new demand still turns into new jobs—and supports broader wage and employment opportunities.
Beyond replacement, the bigger issue is AI augmentation
a16z believes the AI doomsday narrative only talks about “substitution,” but ignores “augmentation.” For some jobs, AI is a threat to survival. For others, AI is an amplifier that makes those roles more valuable.
This isn’t to say AI has no substitution effects—but substitution is not the only effect. For jobs with high judgment, high integration, high responsibility, and high complexity, AI may help people complete more tasks, make faster decisions, and handle a wider range of problems.
Software engineers may be the most typical AI-augmented occupation
The article specifically notes that Goldman’s AI augmentation checklist doesn’t even list software engineers, but software engineers could be the most typical AI-augmented profession. AI coding agents are amplifying engineers’ capabilities, increasing the number of git push actions, new apps, and new company formations.
a16z Growth cites data from Sensor Tower and Wells Fargo Securities showing that the monthly volume of iOS app releases in the U.S. has been broadly flat over the past three years. But it has clearly accelerated since agentic coding tools arrived. The single-month year-over-year growth rate surged rapidly in the second half of 2025, reaching 60% by December 2025. On a TTM basis, it also climbed from near-flat to 24%.
This suggests a wave of “vibe-coded apps” is pouring into the App Store. Apps that previously required full engineering teams may now be quickly built—by individual creators, small teams, and even people without a traditional engineering background—using AI coding agents such as Claude Code, Cursor, and Codex to rapidly create prototypes, modify interfaces, deploy features, and publish to the store. Once the marginal cost of building software falls, the market starts producing a large number of new products that previously were not worth developing, not feasible to develop, or had no engineering resources behind them.
a16z also points out that software development job openings have been rebounding since early 2025, whether measured by absolute numbers or by the share of total openings. The authors admit it is still too early to judge whether this is entirely driven by AI, but logically, as every company thinks about how to integrate AI into its business, demand for software engineers, product managers, and systems design talent may actually increase.
The article also cites Lenny Rachitsky’s observation that product manager job openings have continued to rebound after the interest-rate shock and are now the most plentiful since 2022. a16z argues that the simultaneous rebound in software engineer and product manager roles is exactly why the fixed-work-labor fallacy is wrong.
If AI replaced thinking one-to-one, you might see outcomes like “PMs don’t need so many engineers” or “engineers don’t need so many PMs.” But what we are seeing is that demand for both is rising—because AI lets people accomplish more work and because companies want to do more things. Humans will not just stop creating because AI has appeared.
Looking back to history for a lens on human productivity innovation: a16z rebuts the AI unemployment doomsday theory as fantasy. First appeared on Chain News ABMedia.
