Aaron Levie on Why Open AI Wins

Aaron Levie joins the conversation to discuss the recent open-weights letter signed by Jensen Huang and Box. He frames open-weight AI not as zero-sum competition with closed models, but as complementary innovation that drives overall AI progress and creates more use cases. Levie argues that the letter has two main points: open-weight models are crucial for AI innovation and diffusion, and the U.S. should invest more heavily in open-weight development rather than rely on Chinese models.

On the contentious topic of distillation, Levie takes a permissive stance, arguing it's difficult to establish ethical or legal distinctions between training models on public internet data versus training on AI model outputs. He contends that if models are trained on broadly available knowledge, there's no clear line that makes distillation illegitimate. However, he acknowledges that if he ran a frontier lab, he would attempt to prevent distillation through defensive mechanisms.

Addressing national security concerns about Chinese open-weight models, Levie argues that attempting to restrict China's AI development is futile and counterproductive. He emphasizes that China has the talent, resources, and strategic motivation to develop AI regardless of U.S. restrictions. The alternative scenario—forcing China to develop on different hardware stacks—could actually accelerate their independence and capability rather than slow their progress.

On the economics of open versus closed models, Levie contends that the true moneymaker in AI is inference, not model weights. As competition increases among frontier labs, token costs will converge toward infrastructure costs, making the traditional margins from closed models less defensible. He argues that even open-weight releases wouldn't significantly harm closed labs' revenue if they control the preferred inference infrastructure or post-training environments.

Regarding why frontier labs haven't embraced open-sourcing more aggressively, Levie identifies two categories of reasoning: those with longer time horizons may see open-source as economically viable long-term, while labs like Anthropic likely view it as a safety issue, believing centralized control is necessary for responsible AI deployment.

Levie discusses Anthropic's new Claude Opus 5 model positively, noting meaningful improvements over Opus 4.8 across general knowledge work and domain-specific tasks relevant to Box's enterprise customers. He identifies a problem with Claude's current approach where certain capabilities trigger safety measures and downgrade to weaker models, which he argues is untenable for the future of AI adoption.

On the impact of AI at Box, Levie reports that AI has not reduced hiring but expanded the engineering roadmap dramatically. Projects previously deemed too complex or too small to tackle are now feasible, allowing the company to be more ambitious while also maintaining human engineering talent. He emphasizes that the constraint is now financial rather than technical.

Final discussion centers on model routing as the emerging enterprise strategy, with Levie arguing that as models continuously improve and leapfrog each other, enterprises benefit from abstraction layers that enable model switching without lock-in. This applied AI layer creates value by handling workflow integration, data management, and vertical-specific implementation that pure horizontal models cannot.