Deepseek is a Problem

The video opens with a stark framing: the US will either win everything in AI or be completely outmaneuvered, with no middle ground. The speaker notes that 40% of the US stock market is concentrated in seven tech companies heavily tied to AI outcomes, making this a high-stakes national issue.

The core problem identified is that the business model for open-source AI in the United States is fundamentally broken. When a US lab invests heavily in building and training an open-source model, competitors can immediately take that model and serve it to customers at higher margins since they didn't bear the R&D costs. This creates a structural disincentive for US companies to invest in open-source AI.

In contrast, China's model works because the CCP actively picks winners and subsidizes companies, making it strategically advantageous to give away high-quality AI for free. This undercuts the margins of leading US closed-source labs and positions Chinese models as a cost-effective alternative for US enterprises making AI decisions right now — a particularly urgent moment given slow enterprise adoption cycles.

The speaker surveys the US open-source landscape: Meta has retreated from its open-source commitments, OpenAI treats open-source as a side project for goodwill rather than a core strategy, Anthropic has no open-source strategy at all, and Google's Gemma models are designed for local use rather than enterprise-scale deployment. Nvidia emerges as the most promising player, having committed $26 billion to open-source AI with a uniquely aligned business model — since the companies serving open-source models are also buying Nvidia chips, Nvidia profits regardless.

The geopolitical risks of US enterprises building on Chinese open-source models are outlined in detail. China could optimize its models for its own chips, potentially shifting global chip procurement away from Nvidia. Chinese models could also embed subtle cultural biases that persist even after attempts to remove censorship, influencing American cultural discourse. Additionally, reliance on Chinese AI infrastructure could disrupt the revenue flywheels that US closed-source labs like Anthropic and OpenAI depend on to fund their race toward AGI.

The speaker acknowledges the counterargument: if a US lab achieves AGI and triggers a hard takeoff of recursive self-improvement, none of the open-source competition may matter. Anthropic's flywheel — selling coding models to enterprises, generating revenue and data, and using both to build the next generation — is highlighted as a powerful example. However, the speaker argues that the interim period before AGI is achieved is precisely when Chinese open-source dominance could do the most damage by disrupting the conditions necessary for US labs to reach AGI at all.

Several solutions are proposed: federal grants or compute quotas for open-source AI as a public good, treating open-source AI as national infrastructure with tax credits and sovereign procurement guarantees, encouraging AMD and Intel to adopt Nvidia's hardware-funded open-source model, building smaller vertical models for specific industries like legal, biotech, and defense, and establishing AI standards to reduce duplicated effort among open-source startups. The video concludes by noting that Deepseek's latest model release validates all of these concerns.