Open source AI surges in volume, but Anthropic still dominates spending

A new theory from Decagon CEO Jesse Zhang suggests the AI economy isn't a zero-sum battle between frontier labs and open source. In a post titled "Everyone is wrong about open source AI in the enterprise," Zhang argues they operate in complementary phases: frontier models (like Anthropic's Opus) prove out new use cases, which eventually mature and migrate to lighter, cheaper open source alternatives. This lifecycle means that as mature applications shift, new ones keep emerging, keeping overall frontier spending steady. Zhang's thesis finds support in Vercel and OpenRouter data. While DeepSeek's V4 Flash now processes over a third of tokens on Vercel's platform and leads OpenRouter with 5.3 trillion tokens weekly, Anthropic's Opus 4.8 handles just 2 trillion but at 23x the cost ($1.37 vs $0.06 per million tokens). This pricing power means Anthropic still captures the majority of total AI spend, despite losing volume share.

The phenomenon isn't isolated to Anthropic. Nvidia's upcoming Nemotron model, leveraging Nvidia's infrastructure ties and adaptability, is poised to further reshape the landscape. Yet the core dynamic persists: frontier labs command premium pricing by owning the discovery phase, while open source models dominate cost-sensitive production. Zhang's lifecycle model explains why both segments can thrive simultaneously. It also implies that frontier providers like OpenAI, Anthropic, and Google DeepMind won't be easily disrupted — their value lies in early-stage, high-difficulty problem solving, not just raw token throughput. As long as the overall market for AI-addressable tasks keeps expanding, the two-tier economy of models appears stable. Frontier labs may sell the innovation, and open source sells the scale, but both have a clear role in this rapidly growing ecosystem.