Open source AI models are eating token volume across the enterprise, but frontier labs like Anthropic are still capturing the money. On Vercel's AI gateway, DeepSeek has surged into the lead for token volume over the past week, processing just over a third of everything passing through the platform. Anthropic still accounts for more than half of overall AI spend on the same platform, a share that has slipped only slightly over the past month.
The gap between usage and revenue is now the defining shape of the AI model market. Decagon CEO Jesse Zhang laid out the argument in a post on Monday titled "Everyone is wrong about open source AI in the enterprise," contending that frontier and open source models are not competitors but two phases of the same life cycle.
“Everyone is wrong about open source AI in the enterprise”— Jesse Zhang, Decagon CEO
In Zhang's telling, expensive state-of-the-art models are used to prove out new use cases. Once a workload is understood well enough, teams migrate it to a cheaper open source alternative. New use cases arrive fast enough that frontier spend never really falls.
Key facts
- 01DeepSeek now processes just over a third of tokens flowing through Vercel's AI gateway, jumping into the lead over the past week.
- 02Anthropic still accounts for more than half of overall AI spend on Vercel, with its share only slightly down over the past month.
- 03On OpenRouter, DeepSeek V4 Flash handles 5.3 trillion tokens weekly versus just over 2 trillion for Anthropic's Opus 4.8.
- 04Opus 4.8 costs roughly 23x more per token than V4 Flash — $1.37 per million tokens compared to 6 cents.
- 05Nvidia's newly released Nemotron is not yet reflected in the dashboards but is expected to leap to the front of open source usage.
The public dashboards support the shape of the argument even if they don't confirm every detail. Z.ai, the lab behind the GLM-5.2 model, jumped into fourth place on Vercel over the same week that DeepSeek took the top spot. Much of Anthropic's revenue share is being propped up by its own rising prices, but the ranking has held.
OpenRouter tells a similar story on a larger, less enterprise-weighted user base. DeepSeek V4 Flash is now processing 5.3 trillion tokens weekly, while Anthropic's Opus 4.8 handles just over 2 trillion — a wide gap on volume that flips entirely once price enters the equation.
“The frontier labs will keep owning discovery. Open source will increasingly own production.”— Jesse Zhang, Decagon CEO
Opus 4.8 costs roughly 23x more per token than V4 Flash: $1.37 per million tokens versus 6 cents. OpenRouter doesn't publish a spending leaderboard, but at that ratio Opus almost certainly captures the majority of dollars flowing through the platform even while trailing badly on token count. The two-tier structure — cheap open source doing the bulk work, expensive frontier models doing the harder work — is now visible in the raw data.
Nvidia's Nemotron isn't yet reflected in either dashboard, but it is poised to leap toward the front of open source usage. The model's adaptability and Nvidia's distribution advantages give it a shot at pressuring DeepSeek's position within a quarter or two.
The alternative reading is less flattering to open source. Some enterprise use cases are difficult enough that they can't be moved to a cheaper model without breaking, and the frontier labs are simply holding the workloads that matter most. That view would suggest Anthropic's revenue share is a moat, not a lagging indicator, and that the migration to open source will stall well short of the top of the stack.
Either way, the coffee-beans-to-Starbucks scenario — where foundation labs become commodity inputs while application layers capture the margin — has not played out on the frontier tier. Vertical AI companies have indeed switched to lighter models where they can, and the economics of GPT-wrapper startups have held up. But token for token, Anthropic and its peers are still selling the premium good at premium prices, and buyers are still paying.
The pattern also complicates the standard bear case on Anthropic's revenue durability. If open source were straightforwardly cannibalizing frontier spend, the Vercel and OpenRouter data would show it by now. Instead, both platforms show frontier labs holding revenue share while the base of AI-addressable tasks expands underneath them. The frontier tier isn't shrinking — it's being continuously refilled with new work that hasn't been productionized yet.
The read for anyone modeling the AI market is that model-price compression and frontier-lab revenue can coexist for longer than the bears assume. As long as the set of AI-addressable tasks keeps growing faster than mature workloads migrate down the stack, Anthropic and OpenAI can lose token share every quarter and still grow revenue. The risk is not open source itself but the moment new use-case creation slows — at that point the two-tier economy collapses into a one-tier commodity market, and premium pricing goes with it.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
