Ben Horowitz & Marc Andreessen Compare the AI Boom Vs. Internet Boom

Core Analogies: Internet vs. AI vs. Computing History

• Network vs. Computation: The AI boom is fundamentally different from the Internet boom because the Internet was a network connecting existing computers, whereas AI is a new type of "computer" or information processing system.
• Primary Analogy: The most accurate historical parallel for AI is the evolution of the microprocessor and early mainframe era, rather than the social network dynamics of the Internet boom.
• Deterministic vs. Probabilistic: Legacy computers (von Neumann machines) are deterministic, literal, and reliable; AI models are probabilistic, non-deterministic, and capable of varying outputs or arguing with users.
• Industry Dynamics: Internet startup dynamics were dominated by network effects and positive feedback loops (Metcalfe's Law), while AI dynamics are driven by information processing capabilities rather than user connectivity.
• Future Risks: The AI cycle will likely experience standard technology boom-and-bust volatility, including potential overbuilding of chips and power infrastructure.

The "God Model" vs. Distributed Computing Ecosystem

• Mainframe Fallacy: Early computing history saw a belief that only a few massive systems were needed (e.g., Thomas Watson Sr. predicting the world would need no more than five computers).
• Historical Trajectory: Computing power has evolved from expensive mainframes (costing millions, requiring white-lab-coat maintenance) to mini-computers ($500k), PCs ($2,500), and smartphones ($500), eventually reaching embedded chips costing pennies.
• Current State: Modern computing forms a massive pyramid: a few supercomputers/mother models at the top, followed by millions of PCs, smartphones, and billions of embedded systems in devices like cars (approx. 200 chips per new vehicle).
• Prediction for AI: The AI industry will not consolidate into a few "God models"; instead, it will develop an ecosystem of models in every conceivable shape, size, and capability to suit specific privacy, security, and data needs.

Adoption and Lock-in Dynamics

• Usage Barrier Reduction: AI is the easiest computer to use historically, as it interfaces via natural language (English) rather than complex commands or proprietary operating systems.
• Lock-in Mechanism Shift: Prior eras created vendor lock-in through high complexity and difficulty of use (e.g., training staff on IBM mainframes or the cost of switching from iOS), whereas AI's ease of use removes traditional friction.
• Open Question: It remains an open industry question whether users will be free to choose models based on specific task requirements or if they will be locked into specific "God model" providers despite the low barrier to entry.