On Monday, Jesse Zhang, CEO of Decagon, published a provocative post titled “Everyone is wrong about open source AI in the enterprise.” He argues that while more mature AI deployments are shifting to lighter, open-source models, overall spending on expensive, state-of-the-art frontier models has barely declined.
The core argument
Zhang proposes that frontier models and open-source models are not direct competitors but rather two phases in the same lifecycle: expensive frontier models discover and validate new use cases, and as those use cases mature they can be migrated to cheaper open-source alternatives. As Zhang puts it, “The frontier labs will keep owning discovery. Open source will increasingly own production.”
What the metrics show
Zhang provided little original data, but available platform metrics support the broad pattern. Vercel’s AI gateway dashboard shows that in the past week DeepSeek surged in token volume and now processes just over one-third of the tokens passing through Vercel’s infrastructure. Over the same period, Z.ai — the lab behind the GLM-5.2 model — moved into fourth place.
However, when looking at total token spend on the platform, Anthropic still accounts for more than half of overall AI spending. That share has dipped slightly over the past month, in part because Anthropic has raised prices, but the change is not dramatic.
OpenRouter reports a similar but larger and somewhat less enterprise-focused slice of the market. OpenRouter’s numbers show Deepseek V4Flash processing 5.3 trillion tokens weekly, while the most-used frontier model, Opus 4.8, handles just over 2 trillion tokens. OpenRouter does not rank models by total spend, but it records an average token cost for Opus 4.8 of roughly 23 times that of V4Flash (about $1.37 per million tokens versus $0.06). That cost differential implies Opus is likely capturing the lion’s share of spending despite lower token volume.
These figures do not yet include Nvidia’s Nemotron, a recent entrant that is expected to climb quickly due to Nvidia’s strong relationships and the model’s adaptability.
Why Anthropic’s position hasn’t eroded
Two complementary explanations help account for why frontier providers like Anthropic haven’t lost market share:
- The market for AI-addressable tasks is expanding so rapidly that frontier models can maintain dominance by leading in early-stage discovery. Even as mature applications shift to cheaper models, new use cases continuously arise that demand frontier capabilities.
- Some use cases are sufficiently complex or sensitive that cheaper alternatives cannot fully replace frontier models yet.
Together these dynamics can produce a stable two-tier economy: open-source models dominate production volume and cost-sensitive deployments, while frontier labs continue to capture premium token revenue.
Context and implications
A year ago, some analysts speculated that foundation model labs might become commodity suppliers — the “coffee bean” analogy — while the application layer captured most of the value. Parts of that expectation have played out: certain vertical AI applications have indeed moved to lighter models, and many startups building application wrappers around large models have seen stable economics.
Yet current token-by-token economics show frontier providers retaining access to the most valuable part of the market: premium-priced tokens. Given the data and market dynamics, that situation appears unlikely to change dramatically in the near term.
Conclusion
Zhang’s lifecycle framing helps reconcile the simultaneous rise of open-source models and the sustained market power of frontier labs. Public metrics demonstrate substantial open-source token volume growth, but per-token price disparities mean frontier providers like Anthropic still capture a disproportionate share of spending. The two-tiered model economy—discovery by frontier labs, production by open source—may be a lasting feature of the AI market.
Data sources mentioned in the article: Vercel AI gateway dashboard; OpenRouter. Individuals and organizations referenced: Jesse Zhang (Decagon CEO), DeepSeek, Z.ai (GLM-5.2), Anthropic, Deepseek V4Flash, Opus 4.8, Nvidia Nemotron.
