Will Open Source AI Overtake Closed Models? Ft. Olama, Fireworks and Open Router

Jeff Morgan (Ollama), Dmytro Dzhulgakov (Fireworks), and Alex Atallah (OpenRouter) debate whether open-source inference will reach parity with closed models in five years.

• Open-source models are currently ~20–30% of inference tokens vs. closed-source, per Alex Atallah’s estimate.
• DeepSeek succeeded partly because Fireworks and others had to self-host it — DeepSeek’s own servers crashed and blocked payments at launch.
• DeepSeek was the first strong open-source reasoning model with visible chain-of-thought, which closed models like o1 did not expose.
• Fine-tuning may weaken as a differentiator as powerful RL-trained foundation models reduce the marginal benefit of customization.
• Enterprise demand for full model ownership — including fine-tuned weights — is a structural driver of open-source adoption.
• All three panelists predicted roughly 50/50 open vs. closed inference share in five years, with open-source fragmenting across model families rather than one dominant model.
• Atallah argued decentralized inference providers are the wildcard: without them, closed source likely stays above 50%; one provider he cited earns $360k/day in incentives but sustainability is unproven.

2025-05-12 · Watch on YouTube