On Monday, Decagon CEO Jesse Zhang published a provocative new theory, titled “Everyone is wrong about open source AI in business.” The post tackles the most interesting contradiction of today’s AI economy: Mature AI deployments are shifting to simpler models, he says, even in his own company. But spending on expensive high-end models has not changed.
It’s a new way of thinking about the relationship between boundary models and open source. According to Zhang, they are not competitors, and the success of open source models does not come at the expense of frontier labs. Rather, they are two phases of the same life cycle, with expensive frontier models used to prove use cases that can be transferred to cheaper open source alternatives as they mature.
As mature use cases shift to simpler models, new use cases continue to emerge – and spending on frontier models is not diminishing.
Zhang doesn’t provide much data to support the point, but the data isn’t hard to find. Vercel’s AI gateway dashboard shows that, in just the past week, DeepSeek has surged in token volumes, now processing just over a third of the tokens passing through the company’s infrastructure. Z.ai – the lab behind the popular GLM-5.2 model – jumped to a respectable fourth place at the same time.
But if you scroll down to the overall token spend, you’ll see Anthropic still accounts for more than half of the AI spend on the platform. Given that most of the recent change has come from Anthropic’s rising prices, the share has fallen slightly over the past month, but not by much.
OpenRouter tells a similar story, capturing a much larger (but slightly less enterprise) market share. DeepSeek V4 Flash is the main winner in all usage, processing 5.3 trillion tokens every week. The most popular borderline model, Opus 4.8, holds just over 2 trillion. OpenRouter doesn’t measure the models by the amount spent, but it registers an average token cost of Opus 4.8 about 23x higher than V4 Flash ($1.37 per million tokens, compared to just 6 cents), which means that Opus was still eating up the majority of the money spent.
Those figures don’t even take into account the new arrival, Nvidia’s Nemotron, which is poised to jump to the front of the pack thanks to Nvidia’s strong connectivity and greater model flexibility.
Those figures don’t fully prove Zhang’s point about AI life cycles, but they do show frontier labs like Anthropic aren’t too worried about the rise of open source — at least for now. Another explanation is that the AI-addressable job market is growing so fast that the top models are able to maintain their position by dominating first-tier shipments. As Zhang puts it, “Frontier labs will continue to own discovery. Open source will own production.” Another explanation may be that, as clients move to open source, many use cases are so complex that they cannot be completely replaced by cheaper alternatives.
Either way, this two-tier economic model may be a stable feature of the AI economy.
As recently as last September, I was writing about the possibility of foundation labs eventually selling coffee beans to Starbucks – that is, serving as inputs while the application layer reaps the benefits. Some parts of that prediction came true: Direct AI games were switched to simpler models, for one, and the startup economy of the “GPT wrapper” remained very stable.
But we also see that, token for token, the border providers have managed to hold on to the most desirable part of the market. premium token price. And that doesn’t seem likely to change anytime soon.
If you shop through links in our articles, we may earn a small commission. This does not affect our editorial independence.
