Key points
- Levie calls the US ban on non-US citizens using Anthropic's frontier model a new precedent that would gut Anthropic's API business.
- He predicts Germany, France, Japan and the UK may build their own models as countries learn models can be pulled at any time.
- Levie says SaaS is not dead because agents need enterprise data and workflows, citing Box's growth.
AI export controls will accelerate open and sovereign models
The US government's move to bar non-US citizens from Anthropic's frontier model marks a new precedent, and Levie argues it will push countries and developers toward open weights and sovereign AI. He expects intelligence to become a multiplayer market rather than a race among one or two US labs. That does not make SaaS obsolete. Agents need the same enterprise data, guardrails and workflows, so document and system-of-record platforms become more valuable even as some software gets compressed.
Takeaways
- Export control sets a new AI precedent. The US barred non-US citizens from using Anthropic's frontier model, the first time a frontier model was legally held back. Levie says labs cannot police customer nationality and the rule would obliterate Anthropic's API business.
- Pullback risk pushes sovereign AI. Once governments and developers see a model can be withdrawn by one government, Germany, France, Japan and the UK may build their own. China and Mistral already produce strong open models.
- Open weights likely stay near frontier. Levie frames the gap as months, not years, and is optimistic because China and Nvidia have economic and strategic reasons to keep open models competitive. Mistral has similar incentives. He is about 80 percent confident on the export and open-model debate.
- AI market is becoming multifaceted. Cursor routes across models while also offering its own fast model. Baseten and Fireworks win by post-training open models. Mistral's opportunity is much better, and Levie says intelligence will not come from one or two models.
- SaaS is not dead, data platforms gain. Agents still need the same enterprise data and guardrails, so systems of record and document platforms become more important. Snowflake had a blockbuster quarter and Box accelerated growth as customers upgraded to its enterprise agent plan.
- Safety rhetoric built the review regime. Levie traces the current government review environment to warnings from Anthropic and other safety researchers. He suspects some safety advocates quietly welcome a process where government approves model releases after red teaming, an outcome he calls bad.
- AI job dislocation is temporary. Graduates face short-term disruption, but AI startups are hiring engineers to manage coding agents. Life sciences and industrial firms will also hire more software engineers. Box can barely fill its sales headcount.
Box's agent playbook
- Treat the agent as a new user alongside people and applications.
- Give agents access to the same enterprise documents and contracts people use.