On Monday, Decagon CEO Jesse Zhang published a provocative theory arguing that open-source AI models are not actually competing with frontier labs like Anthropic — they are serving a different phase of the same lifecycle. Zhang’s post, titled “Everyone is wrong about open source AI in the enterprise,” suggests that as enterprise deployments mature, they migrate from expensive frontier models to cheaper open-source alternatives. But because new use cases keep arising, overall spending on premium models barely declines.
What the data shows about token volume vs. spending
Zhang’s theory is supported by publicly available data from two major AI infrastructure platforms. Vercel’s AI gateway dashboard shows that in the past week, DeepSeek surged into the lead for token volumes, now processing just over a third of all tokens passing through the company’s infrastructure. Z.ai, the lab behind the GLM-5.2 model, jumped into fourth place over the same period.
Also read: SK Hynix files for U.S. IPO, giving American investors a direct bet on the AI memory boom
However, when measuring total token spend on Vercel, Anthropic still accounts for more than half of all AI spending on the platform. While that share has dropped slightly over the past month — partly due to Anthropic’s own rising prices — it has not declined significantly.
OpenRouter, which captures a larger but less enterprise-focused segment of the market, tells a similar story. DeepSeek V4Flash is the clear winner on overall usage, processing 5.3 trillion tokens weekly. The most popular frontier model, Opus 4.8, handles just over 2 trillion. But the average token cost for Opus 4.8 is roughly 23 times higher than V4Flash — $1.37 per million tokens compared to just 6 cents — meaning Opus still likely captures the lion’s share of spending.
Also read: AI-driven layoffs in tech surpass 120,000 in 2026: A running list of major cuts
Two explanations for the stability of frontier spending
Zhang offers one explanation: frontier labs will continue to own the discovery phase of AI adoption, while open-source models will increasingly own production. As he puts it, “The frontier labs will keep owning discovery. Open source will increasingly own production.”
Another possibility is that many enterprise use cases are simply too complex to be fully replaced by cheaper alternatives. Either way, the data suggests that a two-tiered economy of AI models may become a relatively stable feature of the market.
As recently as last September, this publication explored the possibility that foundation labs would end up selling “coffee beans to Starbucks” — serving as commodity inputs while the application layer reaped the benefits. Some parts of that prediction came true: vertical AI plays switched to lighter models, and the economics of “GPT wrapper” startups have remained mostly stable. But frontier providers have held on to the most desirable part of the marketplace: premium token pricing.
Nvidia’s upcoming Nemotron model, which is likely to leap to the front of the pack due to Nvidia’s strong industry connections and the model’s extreme adaptability, could further complicate the picture. But for now, Anthropic and other frontier labs appear insulated from the open-source surge — at least until the market for AI-addressable tasks stops growing at its current pace.
Frequently Asked Questions
Are open-source AI models taking revenue away from frontier labs like Anthropic?
Not yet. While open-source models like DeepSeek now process a majority of tokens on platforms like Vercel, frontier models still account for over half of total AI spending due to much higher per-token pricing.
What is the ‘two-phase lifecycle’ theory for AI models?
Decagon CEO Jesse Zhang proposes that frontier models are used to prove out new use cases, which are then migrated to cheaper open-source models as they mature. This means overall spending on frontier models remains stable as new use cases continuously emerge.
How much more expensive are frontier models compared to open-source alternatives?
On OpenRouter, Anthropic’s Opus 4.8 costs roughly 23 times more per million tokens than DeepSeek V4Flash — $1.37 compared to $0.06.
Be the first to comment