> ## Content Index
> Fetch the complete content index at: https://wire.fourthweb.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# YC's Garry Tan Pushes US Labs to Copy DeepSeek's Distillation Playbook
- URL: https://wire.fourthweb.ai/ycs-garry-tan-pushes-us-labs-to-copy-deepseeks-distillation-playbook/
- Published: 2026-09-13T19:32:10.000Z
- Updated: 2026-09-13T19:32:13.000Z
- Description: China's DeepSeek just showed the world how to punch above your weight class in AI, and now YC's president wants American open-source labs to learn the same trick from OpenAI and Anthropic.
- Author: Travis Wright
- Tags: AI Agent Economy, AI Infrastructure, OpenAI, Anthropic, Google AI, China AI

**China's** [**DeepSeek**](https://wire.fourthweb.ai/tag/china-ai/) **just showed the world how to punch above your weight class in AI, and now YC's president wants American open-source labs to learn the same trick from** [**OpenAI**](https://wire.fourthweb.ai/tag/openai/) **and** [**Anthropic**](https://wire.fourthweb.ai/tag/anthropic/)**.**

### The Summary

- [Garry Tan is pushing for US open-weight AI labs to use distillation techniques](https://techcrunch.com/2026/09/11/y-combinators-garry-tan-wants-u-s-open-weight-ai-labs-to-distill-frontier-models-too/?ref=wire.fourthweb.ai) on American frontier models, not just Chinese ones
- [The goal: build a robust ecosystem of American open-weight models](https://techcrunch.com/2026/09/11/y-combinators-garry-tan-wants-u-s-open-weight-ai-labs-to-distill-frontier-models-too/?ref=wire.fourthweb.ai) that can compete without Chinese dependencies
- Distillation lets smaller models learn from larger ones, making powerful AI accessible without massive [compute](https://wire.fourthweb.ai/tag/ai-infrastructure/) budgets

### The Signal

DeepSeek changed the conversation. When China's open-weight lab showed you could distill frontier-class reasoning into models that run on consumer hardware, it exposed a strategic vulnerability. [American open-source AI labs immediately started distilling DeepSeek's models](https://techcrunch.com/2026/09/11/y-combinators-garry-tan-wants-u-s-open-weight-ai-labs-to-distill-frontier-models-too/?ref=wire.fourthweb.ai), creating derivative models that inherited its capabilities. That means US developers were building on Chinese AI infrastructure, even if the resulting code was American.

Tan's proposal flips the script. Instead of American labs distilling Chinese models, [he wants them distilling from OpenAI, Anthropic, and other US frontier labs](https://techcrunch.com/2026/09/11/y-combinators-garry-tan-wants-u-s-open-weight-ai-labs-to-distill-frontier-models-too/?ref=wire.fourthweb.ai). The mechanics are the same. You feed a smaller model outputs from a larger one, training it to mimic the reasoning patterns without rebuilding from scratch. What changes is the dependency chain.

> "Giving the US a more robust set of open-weight options that aren't Chinese."

This matters because distillation is now the proven path to competitive open-weight models. Training frontier models from scratch costs hundreds of millions. Distilling from them costs a fraction of that. DeepSeek proved you could get 80-90% of frontier performance for 5% of the cost. Every serious open-weight lab is now chasing that efficiency curve.

The current dynamic puts American open developers in an awkward position:

- Distill from DeepSeek: get great models, accept Chinese AI in your stack
- Train from scratch: burn cash you probably don't have, lag behind
- Wait for American alternatives: watch competitors ship while you stand still

Tan's arguing for a third option that didn't exist until recently. [American frontier labs have the models worth distilling](https://techcrunch.com/2026/09/11/y-combinators-garry-tan-wants-u-s-open-weight-ai-labs-to-distill-frontier-models-too/?ref=wire.fourthweb.ai). They just haven't made them available for this kind of derivative training at scale. If they did, open-weight labs could build on GPT-4, Claude, or [Gemini](https://wire.fourthweb.ai/tag/google-ai/) the same way they're building on DeepSeek now.

The policy angle is obvious. If you're worried about AI supply chains running through Beijing, you need domestic alternatives that don't require choosing between sovereignty and competitiveness. Distillation makes that possible without subsiding hundred-million-dollar training runs for every startup.

### The Implication

Watch whether OpenAI and Anthropic bite. Letting open-weight labs distill your models means accepting that your competitive moat will compress faster. Your frontier advantage becomes everyone's baseline in months, not years. That's a tough sell for companies racing to AGI.

But the alternative might be worse. If American open-weight AI stays dependent on Chinese models, every regulation or restriction aimed at China creates collateral damage at home. Tan's betting that frontier labs would rather enable American distillation than watch the open ecosystem tilt permanently toward DeepSeek and its successors.

### Sources

[Hacker News Best](https://techcrunch.com/2026/09/11/y-combinators-garry-tan-wants-u-s-open-weight-ai-labs-to-distill-frontier-models-too/?ref=wire.fourthweb.ai) | [TechCrunch AI](https://techcrunch.com/2026/09/11/y-combinators-garry-tan-wants-u-s-open-weight-ai-labs-to-distill-frontier-models-too/?ref=wire.fourthweb.ai)