Sriram Krishnan on Open Source AI's Biggest Week Yet

What It Covers

Former White House AI policy advisor Sriram Krishnan joins a16z's Theo Jaffe and Sofia Puccini to analyze the surge of open-weight models—including Kimi K3, Qwen, and Grok—and what this shift means for Frontier Lab pricing, national security, distillation policy, and America's competitive position in global AI.

Key Questions Answered

• •Frontier Lab pricing pressure: Open-weight models like Kimi K3 reaching near-Frontier capability creates direct pricing pressure on Anthropic and OpenAI. Developers can route agents handling routine tasks—email scanning, calendar checks—to cheaper open models, eroding Frontier Lab gross margins. Expect token price cuts and extended model availability windows, as seen with Anthropic extending Claude's deprecation deadline.
• •Security paradox of over-restricted models: U.S. Frontier models like Claude refuse certain cybersecurity-related prompts, pushing security professionals toward Chinese open-weight models like Kimi K3 for legitimate vulnerability research. Krishnan frames this as a strategic liability: American defenders end up using foreign AI tools to audit their own code, creating an asymmetric security disadvantage.
• •Distillation policy imbalance: Chinese model developers can freely extract reasoning traces from American models, while U.S. startups face legal ambiguity about distilling from other American models. Krishnan endorses proposals from Sequoia's Dean Ball and Ben Thompson to enshrine distillation rights for American developers, leveling the competitive playing field against Chinese counterparts.
• •Moat shifts from intelligence to harness: As open-weight models commoditize mid-tier intelligence, Frontier Labs' defensible advantage moves toward product stickiness—tools like Claude Code, Codex, and CoWork. Developers should evaluate AI vendor lock-in not by model benchmarks alone, but by the surrounding toolchain, workflow integration, and switching costs those products create.
• •Open-weight economics benefit the full stack: When open-weight models deliver genuine user value, every infrastructure layer—NeoClouds like Baseten and Fireworks, chip providers, data centers—captures economics previously concentrated at Frontier Labs. Enterprises preferring on-premise deployment for compliance reasons become a direct revenue channel, making open-weight model providers viable businesses without requiring closed-model margins.