Ben Horowitz: The Fight Over Open Source AI

What It Covers

Ben Horowitz of Andreessen Horowitz joins the a16z podcast to argue that open source AI models are essential for national security, market competition, and safety. He addresses a16z's co-signing of NVIDIA's Open Weights letter, the risks of AI monopolization, distillation legality, and why the AI market remains under 3% penetrated.

Key Questions Answered

• •Open Source Safety Precedent: Historical data supports open models being safer than proprietary ones — Linux and the open internet had significantly fewer security vulnerabilities than Windows. When the entire developer community can inspect and patch a system, safety problems get solved faster than when a single monopoly controls the codebase with no external accountability.
• •AI Market Penetration: Despite widespread coverage, the AI market is likely less than 3% penetrated, meaning all players — open and closed source — will continue growing simultaneously. Builders should not treat current market dynamics as stable equilibrium; token usage at enterprises is growing roughly tenfold annually, signaling sustained expansion across every category.
• •Platform Risk for Application Builders: Startups building on proprietary models like Anthropic face a documented risk: when an application category gains traction, the foundation lab enters that category directly, charges application providers full price, and subsidizes its own competing product. Building on open source models eliminates this platform dependency and removes the threat of being undercut or cut off.
• •Distillation Legality and Competitive Value: AI model outputs are not covered by copyright law, making distillation a terms-of-service issue at most, not a legal one. Application companies — including Cursor — have used open source and distillation to bootstrap products affordably. Builders should treat distillation as a legitimate cost-reduction tool while monitoring individual platform terms carefully.
• •Open Source Business Model Alternative: Selling model inference at high margins does not work for open source labs, but a Palantir-style services and deployment model does. Smaller, well-post-trained open models frequently outperform large proprietary models on specific enterprise tasks while costing less. Builders targeting B2B use cases should evaluate fine-tuned open models before defaulting to frontier closed alternatives.