Avi Goldfarb | Rethinking Business as Technology in the Age of AI

Summary

Avi Goldfarb argues that businesses can understand technological change by identifying what a technology makes cheaper. Computing lowered the cost of arithmetic; the internet lowered the costs of searching, copying, tracking, and verifying information. AI’s core economic effect, he says, is lowering the cost of statistical prediction—filling in missing information.

This framing helps explain why AI applies beyond conventional forecasting. Translation, diagnosis, writing, image generation, and coding can all be understood as prediction tasks. As prediction becomes cheaper, companies may use it to automate or improve activities that were not previously treated as prediction problems.

Business Frameworks and Strategic Ideas

• Technology-cost framework: Ask what a technology makes faster, better, or cheaper. The resulting cost reduction—not just the technology itself—can reveal how it may reshape business.
• AI as prediction: Treat AI as a tool for generating likely outputs from data. For example, a model creating an image from a prompt draws on patterns in its training data rather than simply retrieving a matching image.
• General-purpose technology (GPT) feedback loop: A GPT can create a cycle in which:
  1. An upstream industry improves the technology.
  2. Downstream businesses invent new applications.
  3. Those applications create demand for further upstream improvements.

Goldfarb identifies this cycle as a potential source of unusually large productivity gains. It also means GPTs are hard to confirm in real time: rigorous evidence may require decades of diffusion and innovation—potentially around 40 years.

• Competition and market power: The feedback loop depends on downstream innovators believing that upstream suppliers will keep improving the technology and will not use their position to squeeze or exclude them. If they lack that confidence, downstream investment and innovation may stall.

Competition: Opportunities and Risks

Reasons for Optimism

• There are multiple AI suppliers, including established and newer companies, open and closed models, and specialized application developers in areas such as banking, financial services, graphic design, and healthcare.
• AI has meaningful marginal costs—especially compute and energy—in addition to the fixed cost of developing models. If marginal costs remain significant, the market may be less prone to “winner-take-all” outcomes than some software-platform markets.
• Open models can make it harder for leaders to preserve an advantage: Goldfarb says lagging models may be only a few months behind leaders.
• Data has diminishing returns for a particular use case. In statistical terms, prediction quality improves roughly with the square root of the amount of data, so each additional unit of data has a smaller effect as the dataset grows.

Reasons for Concern

• Endogenous sunk costs: As leaders invest heavily and advance the technology, catching up can become increasingly expensive. Goldfarb notes that firms may need to spend billions to reach the frontier, creating a natural oligopoly even if a monopoly does not emerge.
• New applications can reset scale advantages: Discovering new use cases may require a higher minimum efficient scale, temporarily giving leading firms market power.
• Feedback-driven data advantages: Faster feedback can create a reinforcing loop—better predictions attract more usage and data, which can improve predictions further. The strength of this effect remains contested.
• Regulation can entrench incumbents: Rules focused on harms other than competition may raise administrative burdens or restrict entry. Goldfarb cites evidence suggesting that GDPR and privacy regulation helped dominant firms in Europe.
• Tying and bundling: Linking AI to other dominant services—such as office software, cloud services, chips, or user interfaces—could reduce adoption and increase industry rents.

Practical Implications for Business

• For AI deployment: Look beyond obvious forecasting applications. Consider which workflows involve generating a plausible missing input, recommendation, diagnosis, draft, or design.
• For technology suppliers: Assess both fixed investment and ongoing compute and energy costs; those costs shape the economics of scaling and competition.
• For businesses building on AI platforms: Consider whether a supplier’s incentives, pricing, or bundled services could limit future product choices or downstream innovation.
• For policymakers and companies assessing the market: Distinguish competition concerns from other regulatory goals. Rules that unintentionally raise entry costs can protect incumbents.
• For assessing AI as a GPT: Treat the claim as a hypothesis to monitor. AI’s broader productivity impact depends on complementary downstream invention, not merely on model capability.

Metrics and Timelines Mentioned

• AI advances have been especially noticeable since 2012.
• Prediction performance improves approximately with the square root of sample size, implying diminishing data returns for a given use case.
• Open-model leaders may be only a few months ahead of followers.
• Catching up to frontier firms may require billions of dollars in investment.
• Establishing whether a technology is a general-purpose technology may take roughly 40 years of diffusion and feedback.

Presenter and sources mentioned: Avi Goldfarb; Nord and Bresnahan; Hal Varian and Carl Shapiro; Katherine Tucker; coauthors Jay and Joshua (as heard in the subtitles); Dario Amodei; Diane Coyle; Bresnahan and Trajtenberg; Florent (surname unclear in the subtitles); Sutton; and Peltzman and Wall (names unclear in the subtitles).