digest
Short Summary
Main recommendation: The U.S. should support both open weight and closed weight AI models and avoid restricting open weight models, so American companies can compete with China's cheaper open models. Baseten's CEO argues that a strong open weight ecosystem gives enterprises ownership and long-term control over their intelligence, and that this requires policy that does not limit competition. He also says the U.S. needs more talented teams and labs to match China's competitive open model ecosystem.
Key Points
- Open weight models are downloadable, modifiable trained model files that anyone can run on their own, while closed models stay inside one company's servers.
- Over 230 companies, including Baseten, signed a letter organized around NVIDIA's position asking government not to restrict open weight models, which the letter argues are key to U.S. competition.
- The CEO says the U.S. can match China's cheaper open weight models; the real gap is that China has five or six labs competing on open models, while the U.S. does not yet have that ecosystem.
- He argues closed and open models both matter, since closed models serve certain customers while open models give enterprises ownership and long-term control.
- On compute, he sees demand growing roughly tenfold per quarter and dismisses concerns about circular financing among hyperscalers.
- He says the company has raised funds four times in 18 months, with a latest valuation reportedly around $13 billion, and would go public if needed to fund a lasting company.
Details Worth Noting
- Concrete instructions: The video gives no step-by-step instructions for users. The policy-level actions the CEO recommends for government are: commit to a competitive ecosystem where all model types can thrive, do not restrict competition, and set up American companies and labs to produce open source models. He also says U.S. competition should be across all fronts, with the goal of being best across every model, not only closed weight.
- The host framed the choice as either supporting cheap, open models to compete with China or building a closed, higher-priced market; the CEO said that framing may be wrong and that both are needed.
- The CEO cited NVIDIA's work and Thinking Machines as examples of good, talented, dense teams, but said U.S. open weight models have not yet reached the frontier level of intelligence.
- Baseten runs open weight models for businesses seeking to avoid the cost of frontier closed models, and it has recently started hosting proprietary closed-weight models as well.
- Compute demand is described as exceeding supply, and the CEO says the key is ensuring enough compute is available for that demand.
Bottom Line
The debate is less about picking open or closed models than about whether U.S. policy and talent can build a competitive open weight ecosystem that matches China's on price and capability.
