Chamath vs. Baker: The Open Source AI Fight That Missed the Point
Both are right about their champion. Neither addressed that China is winning.
On Sunday morning, Chamath Palihapitiya posted a one-liner: "The most important company in AI is an American open source company. No federal permission slip required." He was referring, implicitly, to Meta and its Llama models — the argument he'd made on the All-In Podcast the day before: banning open source would force every US company onto closed AI at "50 to 100x the cost," tanking the stock market.
Gavin Baker fired back within hours. "Nvidia is the leading open source AI company by a wide margin. The silliness of this post is compounded by the fact that the letter did not suggest that closed source AI companies should open source their models." He attached a HuggingFace chart showing NVIDIA's repo count towering over every other organization.
They're both right. They're also both wrong — about what "winning" means and who's actually winning.
Two different scoreboards
Chamath and Baker are measuring completely different games.
Chamath's game: enterprise model market share. Meta's Llama owns this. Among enterprises using open source LLMs, Llama holds 70% market share (Menlo Ventures, Dec 2025). Mistral is at 15%, Chinese models at 10% combined. If you run a Fortune 500 company and you're picking an open source foundation model, you probably pick Llama.
But that lead is eroding. Total open source enterprise adoption fell from 19% to 11% of all enterprise AI usage over the past year. Llama hasn't had a major release since Llama 4 in April 2025. On OpenRouter, the main open model API router, Llama models "pretty much died out" by late 2025 — their token share collapsed as users shifted to Chinese models.
Baker's game: institutional open source output. NVIDIA isn't just a model company; it's an open source platform company. As of October 2025, NVIDIA was the #1 institutional contributor on HuggingFace with roughly 384 repos — ahead of Alibaba, Google, Meta, and every other organization. It has 23 models on public leaderboards, holding the #1 spot on LiveCodeBench (coding), AIME (math), and safety benchmarks. It releases robotics foundation models (GR00T), physical AI models (Cosmos), biomedical models (Clara), and the entire CUDA-adjacent software stack. HuggingFace CEO Clem Delangue called NVIDIA "the king of American open source AI" — the leading US counterweight to Chinese open weights.
The elephant neither mentioned: China
Here's the uncomfortable fact that Chamath and Baker both sidestepped. Chinese open source models now lead on actual usage:
- 41% of HuggingFace downloads are Chinese models (Spring 2026). They passed the US.
- ~30% of OpenRouter inference traffic is Chinese — DeepSeek, Qwen, Kimi, GLM.
- Kimi K3 (Moonshot AI) just debuted at #3 on DeepSWE, the hardest software engineering benchmark, at 69% — behind only GPT-5.6 Sol (73%) and Claude Fable 5 (70%). It's the first open-weight model to achieve frontier-level performance.
- SemiAnalysis called it bluntly: "Kimi K3 sits above Gemini on every composite benchmark. Google should feel incredibly embarrassed."
- On a 16-task blind benchmark, Kimi K3 scored 90.49 overall — beating Claude Opus 4.8 (88.76), Sonnet 5 (87.43), and GPT-5.6 Sol (87.30). At $1.43 for all 16 tasks, it cost less than Fable ($1.49).
- SemiAnalysis host Max on the K3 emergency episode: "I'm shocked how markets are still so inefficient and we haven't had a single American company that's at least on par with the fifth best Chinese company."
The real scorecard
| Lens | Winner | Reality Check |
|---|---|---|
| Enterprise LLM adoption | Meta (Chamath) | 70% share, but total OS enterprise pie shrinking from 19% → 11% |
| Open source breadth / infrastructure | NVIDIA (Baker) | #1 HF contributor, 23 models, but not a model company per se |
| Actual inference usage | China | 41% HF downloads, 30% OpenRouter traffic |
| Frontier capability | China (Kimi K3) | First open-weight model at frontier; beats Opus, Gemini |
| Price per token | China | $0.45/1M tokens open-source vs $3.50 OpenAI vs ~$10 Claude |
What this means
The open source AI debate isn't really Chamath vs. Baker. It's America vs. China.
For investors, the implications are threefold:
First, NVIDIA's moat is deeper than the Street prices. The "give away the models, own the ecosystem that runs on your chips" strategy means even Chinese open source dominance flows through NVIDIA silicon. Every Kimi K3 inference run, every DeepSeek download, every Qwen fine-tune — they all run on GPUs NVIDIA sold.
Second, Meta's open source lead is fragile. No major Llama release in 15 months. Enterprise share declining. Inference traffic collapsing. The model that defined "American open source AI" is at risk of becoming legacy infrastructure while Chinese labs iterate monthly.
Third, the US regulatory response is the wildcard. Chamath's argument (banning open source would tank the market) and Baker's counter (the letter didn't propose banning open source) both miss the real scenario: what happens if the government bans Chinese open source models? Enterprises that have standardized on Qwen or DeepSeek would face the same cost shock Chamath warned about — just from a different trigger.
The two American champions are strong but fighting different battles. The uncomfortable question neither addressed on Sunday: is American open source AI leadership already lost?
More research at bargo.ai/research.