30 Economic Impacts of AI
AI is reshaping the global economy in ways we are only beginning to understand. The parallels to previous technological revolutions (electricity, the internet, mobile computing) are instructive but imperfect: AI affects cognitive work, not just physical or informational work, making its economic impact potentially broader and more disruptive. This chapter surveys what we know, what we do not know, and what you should be thinking about.
Economic predictions about transformative technologies are almost always wrong. Experts in 1995 predicted the internet would be a niche tool. Experts in 2010 said smartphones would not change computing. We should be humble about predicting AI's economic impact while still preparing for the range of likely outcomes.
30.1 AI and the Labor Market
30.1.1 Who Is Affected?
Eloundou et al. (Eloundou et al. 2023) conducted one of the most comprehensive analyses of LLM exposure across US occupations, finding that approximately 80% of the workforce could have at least 10% of their tasks affected, and around 19% could see 50% or more of their tasks impacted. The most exposed occupations are not the ones people typically worry about:
- Highly exposed: Writers, translators, accountants, mathematicians, financial analysts, programmers, legal assistants, and data entry clerks.
- Moderately exposed: Teachers, managers, consultants, marketers, and designers.
- Least exposed: Electricians, plumbers, construction workers, athletes, and other occupations requiring physical dexterity and real-world interaction.
Unlike previous automation waves that primarily affected manual labor, AI disproportionately affects white-collar knowledge work: precisely the types of jobs that educated workers pursued to avoid automation. The most “automation-proof” jobs may turn out to be trades and hands-on professions.
30.1.2 Augmentation vs. Automation
AI affects work in two fundamentally different ways:
- Automation: Replacing human tasks entirely. Examples: automated data entry, AI customer support chatbots, routine code generation, basic document summarization.
- Augmentation: Making human workers more productive while keeping them in the loop. Examples: AI coding assistants (GitHub Copilot), AI-assisted medical diagnosis, AI-powered legal research tools.
The distinction matters enormously for workers and policy. Augmentation raises productivity and potentially wages; automation reduces demand for certain types of labor. Most AI applications today are augmentation, not full automation, but the boundary is shifting rapidly.
30.2 Measured Productivity Gains
Unlike many technology hype cycles, AI's productivity benefits have been documented in rigorous studies:
- Customer support: Brynjolfsson et al. (Brynjolfsson et al. 2023) showed that customer support agents using AI assistants handled 14% more issues per hour. The gains were largest for novice workers, suggesting AI helps flatten the skill curve.
- Programming: GitHub's studies (Peng et al. 2023) found that developers using Copilot completed coding tasks 55% faster in controlled experiments.
- Consulting: A Harvard Business School experiment (Dell’Acqua et al. 2023) found that consultants using GPT-4 completed 12.2% more tasks, 25.1% faster, with 40% higher quality.
- Writing: A study on professional writing found that AI assistance reduced task completion time while particularly boosting the quality of writing by weaker writers.
Ethan Mollick describes AI capabilities as a “jagged frontier”: LLMs are superhumanly good at some tasks and surprisingly bad at others, and these boundaries are not intuitive. Workers who blindly trust AI on tasks beyond the frontier perform worse than those who use no AI at all. Understanding where AI helps and where it misleads is the critical skill.
30.3 The Cost Structure of AI
30.3.1 Training Costs
Training frontier models has become staggeringly expensive:
- GPT-4 is estimated to have cost over $100 million in compute.
- Google's Gemini Ultra reportedly cost even more.
- Training clusters now cost billions of dollars to build and require power infrastructure comparable to small cities.
This creates a concentration of AI capabilities in a small number of well-funded organizations. Open-source models (LLaMA, Mistral, DeepSeek, Qwen) partially democratize access by letting others fine-tune and deploy models without bearing the pre-training cost.
30.3.2 Inference Costs: The Real Battleground
While training is a one-time cost, inference happens every time someone uses the model. The economics of inference are improving rapidly:
- The cost per million tokens has dropped by orders of magnitude since GPT-3's release.
- Quantization (Chapter 9) reduces model size 2 to 4x with minimal quality loss.
- Distillation (Chapter 15) creates smaller models that approach the quality of larger ones.
- Hardware improvements (NVIDIA H100/B200, custom inference chips) continue to reduce cost per operation.
- DeepSeek's efficiency innovations demonstrated that clever engineering can dramatically reduce both training and inference costs.
This cost trajectory matters because it determines which applications become economically viable. As inference costs approach zero, AI becomes embedded in everything.
30.3.3 Energy and Environment
AI's energy consumption is a growing concern:
- Large training runs consume megawatts of power for weeks or months.
- Data centers are being built near power plants and hydroelectric dams.
- The total electricity consumed by AI inference is growing as adoption increases.
Responses include more efficient architectures (MoE models, BitNet 1-bit quantization), model compression techniques, and commitments to renewable energy. Whether AI's environmental benefits (e.g., optimizing energy grids, accelerating materials science) outweigh its costs is an open question.
30.4 New Industries and Business Models
AI is not just transforming existing industries; it is creating new ones:
- AI infrastructure: GPU cloud providers (CoreWeave, Lambda), model-as-a-service APIs (OpenAI, Anthropic, Google), and AI-optimized chip design (NVIDIA's dominance, Google TPUs, AMD's challenge).
- AI tooling: A rapidly growing ecosystem of development tools, evaluation platforms, fine-tuning services, observability tools, and prompt management systems.
- AI-native products: Products that could not exist without AI: real-time language translation earbuds, personalized AI tutors, code generation copilots, and generative design tools.
- Data and labeling: Companies like Scale AI and Surge AI provide the human feedback essential for RLHF and evaluation. The demand for high-quality human annotation has created a global gig economy.
During gold rushes, the most reliable wealth came from selling picks and shovels rather than panning for gold. In the AI gold rush, NVIDIA (the pick and shovel seller) has become one of the most valuable companies in the world. For your career, consider: are you panning for gold (building AI applications) or selling picks and shovels (building AI infrastructure and tools)?
30.5 Inequality and Access
AI risks exacerbating existing inequalities along multiple dimensions:
- Geographic: AI capabilities are concentrated in a few countries (US, China, UK, France, Canada). Developing nations may become consumers rather than producers of AI technology.
- Corporate: The compute and data requirements for frontier models create enormous barriers to entry, potentially leading to oligopolistic market structures.
- Labor: Workers whose skills complement AI (prompt engineering, AI system design, oversight) may see wage increases, while those whose skills substitute for AI face displacement.
- Educational: Access to AI tools and training varies widely. Students at well-resourced institutions can use AI effectively while others cannot.
Countervailing forces include open-source models (LLaMA, Mistral, DeepSeek), local deployment tools (Ollama, llama.cpp), efficient small models, and falling hardware costs. Whether these forces are sufficient to prevent AI from widening inequality is one of the defining questions of the next decade.
30.6 Policy and Governance
Governments worldwide are beginning to regulate AI:
- EU AI Act (2024): The first comprehensive AI regulation. Classifies AI applications by risk level (unacceptable, high, limited, minimal) and imposes requirements including transparency, conformity assessment, and registration for high-risk systems.
- US approach: Executive orders on AI safety, voluntary commitments from major AI labs, and agency-level guidance (FDA for medical AI, SEC for financial AI) rather than comprehensive legislation.
- China: Regulations on generative AI, deepfakes, and algorithmic recommendations, combined with strategic investment in AI capabilities.
- Intellectual property: Unresolved questions about whether AI-generated content can be copyrighted and whether training on copyrighted works constitutes fair use. Multiple ongoing lawsuits (New York Times vs. OpenAI, Getty Images vs. Stability AI) will set precedent.
Regulate too little and AI causes harm (bias, job displacement, misinformation). Regulate too much and you push AI development to jurisdictions with fewer safeguards while slowing beneficial innovation. Getting this balance right is one of the hardest policy challenges of our time, and it requires input from the AI practitioners who understand what the technology can and cannot do: people like you.
30.7 What This Means for Your Career
Regardless of whether you are a student, a working professional, or a researcher, here are the practical implications:
- Invest in AI literacy: Understanding AI capabilities and limitations is becoming as essential as computer literacy became in the 1990s. You are already doing this by reading this book.
- Build complementary skills: AI amplifies domain expertise. A doctor who understands ML is more valuable than either a doctor or an ML engineer alone.
- Stay adaptable: The specific tools and techniques that are hot today may be obsolete tomorrow. Focus on fundamentals (statistics, optimization, systems thinking) that transfer across paradigm shifts.
- Contribute to the conversation: AI policy decisions are being made now. They will affect everyone. Engage with the discussion rather than leaving it to people who may not understand the technology.
30.8 Exercises
- Read the “GPTs are GPTs” paper (Eloundou et al. 2023). Which occupation categories are most and least exposed to LLM capabilities? Where does your own current or planned career fall on the exposure spectrum?
- Estimate the total cost of fine-tuning a 7B parameter model on 100,000 training examples using a cloud GPU provider (e.g., Lambda, RunPod, or AWS). Compare this with the cost of using a frontier model API (e.g., GPT-4, Claude) for the same number of inference calls.
- Debate: Should frontier AI model weights be open-sourced? Make the strongest possible case for both sides, considering innovation, safety, competition, and access.
- Research one country's approach to AI regulation (EU, US, China, UK, or another). Summarize its key provisions and evaluate whether it balances innovation with safety effectively.
- Track the cost per million tokens of a frontier AI API over the next six months. Plot the trajectory. What does this suggest about the future economics of AI applications?