Open Source Tops Token Volume, Anthropic Still Holds Half of Spend
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Decagon CEO Jesse Zhang published "Everyone is wrong about open source AI in the enterprise," arguing frontier and open source models occupy two phases of the same lifecycle — frontier labs own discovery while open source increasingly owns production.
- Vercel's AI gateway data shows DeepSeek now leads token volumes on the platform at just over a third of tokens processed, with Z.ai's GLM-5.2 jumping to fourth place in the past week.
- Anthropic still accounts for more than half of overall AI spend on Vercel, a share that has dropped only slightly over the past month despite Anthropic's own price increases.
- OpenRouter data shows DeepSeek V4 Flash processing 5.3 trillion tokens weekly versus 2 trillion for the top frontier model, Opus 4.8 — yet Opus costs $1.37 per million tokens versus just $0.06 for V4 Flash, roughly 23x more.
- Nvidia's Nemotron is the newest entrant and is poised to leap to the front of the open-source pack thanks to Nvidia's distribution strength and the model's adaptability.
- The author's earlier "coffee beans to Starbucks" prediction — that labs would become commodity inputs — has only partially come true: vertical AI has indeed switched to lighter models, but frontier providers have retained the premium token price.
- Overall frontier-model revenue has barely budged because the market of AI-addressable tasks is expanding fast enough that new frontier use cases replace the ones that migrate to open source.
Why it matters: Frontier labs like Anthropic are monetizing roughly half of the AI spend on major gateways while ceding raw token volume to cheaper open-source alternatives — meaning enterprise AI budgets still flow disproportionately to a few premium providers even as the usage base fragments, giving Anthropic a durable revenue moat against open-source commoditization for now.
