The practical takeaway here is straightforward: Garry Tan wants the US to replicate, on its own turf, the training playbook that made DeepSeek a wake-up call. Distillation—training a smaller model to mimic the outputs of a larger frontier one—is already a proven technique for compressing capability into deployable, open-weight packages. Tan's argument is that American labs should be doing this with American frontier models, not leaving that space to Chinese developers.
Right now, the open-weight AI landscape has a gap. The most capable openly available models have increasingly come from Chinese labs, while US frontier development has concentrated inside closed, API-only systems from OpenAI, Anthropic, and Google. That asymmetry matters for builders who need models they can run locally, fine-tune freely, or deploy in air-gapped environments.
Tan's proposal would create a domestic supply chain for open-weight models: frontier labs like OpenAI or Anthropic provide the teacher model, and smaller American labs do the distillation work to produce capable, redistributable weights. This isn't a technical stretch—distillation at scale is well understood—it's more of a coordination and policy question about whether frontier labs will permit or support that kind of downstream use.
For builders, this matters because open-weight model availability directly affects what you can ship. Closed API models mean ongoing costs, data-sharing concerns, and dependency on a vendor's uptime and pricing decisions. A richer ecosystem of US-origin open-weight models would give teams more leverage and more options for sensitive or cost-sensitive deployments.
Watch this space for policy movement. Tan's position at YC gives him a direct line to both startups and Washington conversations about AI competitiveness. If frontier labs are nudged—through incentives or policy—to license their models for distillation by domestic open-weight labs, the downstream effect on the builder ecosystem could be significant within 12–18 months.