📰 Key Highlights

Decagon CEO Jesse Zhang recently posted a counterintuitive take: the more mature enterprise AI deployments become, the more they tend to switch to lightweight open-source models — but overall spending on frontier models has barely declined. He believes frontier and open-source models aren’t competitors but two stages of the same lifecycle. Expensive frontier models validate new use cases, and once a scenario matures, it’s handed off to cheaper open-source replacements. At the same time, new application scenarios keep emerging, keeping demand for frontier models at a high level.

The data backs this up. Vercel’s AI Gateway shows that over the past week, DeepSeek jumped to first place in token usage, handling roughly a third of the platform’s traffic; Z.ai’s GLM-5.2 also climbed to fourth. Yet in terms of overall AI spend share, Anthropic still accounts for more than half of total platform spend. Recent price hikes have slightly trimmed its share, but the dip isn’t significant. OpenRouter’s data tells the same story: DeepSeek V4 Flash processes 5.3 trillion tokens per week, while Opus 4.8 processes about 2 trillion — but Opus 4.8 is priced at $1.37 per million tokens, roughly 23 times V4 Flash’s $0.06. That means frontier models still dominate on the revenue side. On top of that, Nvidia’s Nemotron is about to enter the arena with strong ecosystem connections.

Zhang’s conclusion: “Frontier labs will continue to dominate the exploration phase, while open-source will increasingly own production deployments.” That explains why the rise of open-source hasn’t really hurt the revenue of top frontier model providers like Anthropic — yet.

💬 JudyAI Lab Perspective

This piece reveals a market structure that’s happening right now but isn’t very intuitive: open-source models are grabbing the traffic, while frontier models are still making money. They aren’t zero-sum competitors — they’re collaborators across a lifecycle.

The Decagon CEO’s “two-stage theory” is something AI builders should take seriously. Frontier models handle concept validation: high cost, but used to map out the feasibility and boundaries of new scenarios. Once a scenario is validated and demand stabilizes, you switch to cheaper open-source options like DeepSeek. Vercel Gateway data confirms this logic: DeepSeek V4 Flash handles a third of the platform’s traffic, but Anthropic still accounts for more than half of overall spend. The two aren’t contradictory — this is rational layering in mature enterprises, with frontier models handling exploration and open-source handling scale. We’re seeing this layered thinking gradually become the standard for sizable AI applications, not just a special setup reserved for big players. The constant emergence of new scenarios is the real reason demand for frontier models stays high.

Next time you’re evaluating AI model selection, ask yourself one question first: is this scenario still being figured out, or has it already proven out? The answer changes the model selection logic completely.

📅 Source Info

🔗 Further Reading