On Monday, Decagon CEO Jesse Zhang released a provocative new theory titled “Everyone is wrong about open source AI in the enterprise.” This post tackles one of the most interesting paradoxes of today’s AI economy. At his own company, he says as well, more mature AI deployments are switching to lighter-weight models. But overall spending on expensive, cutting-edge models has remained largely unchanged.
This is a new way to think about the relationship between frontier models and open source models. Zhang says they are not competitors, and the success of the open source model does not come at Frontier Labs’ expense. Instead, they are two phases of the same lifecycle, with expensive frontier models used to prove a use case and then succeeded by cheaper open source alternatives as they mature.
New use cases continue to emerge as more mature use cases switch to lightweight models, and overall spending on frontier models hardly decreases.
Zhang doesn’t provide much data to support this point, but it’s not hard to find. Vercel’s AI Gateway dashboard shows that in just the past week, DeepSeek has risen to the top in token volume and now processes just over a third of the tokens passing through the company’s infrastructure. Z.ai, which developed the popular GLM-5.2 model, jumped to a respectable fourth place in the same period.
However, if you scroll down to overall token spending, you’ll see that Anthropic still accounts for more than half of the overall AI spending on the platform. Given that much of the recent changes have been due to Anthropic’s own price increases, the stock price has fallen slightly over the last month, but not by much.
OpenRouter tells a similar story, capturing a much larger (but slightly less enterprise) segment of the market. DeepSeek V4 Flash is the main winner in overall usage, processing 5.3 trillion tokens each week. The most popular Frontier model, the Opus 4.8, handles just over 2 trillion. Although OpenRouter does not rank models by total spend, the average token cost for Opus 4.8 was recorded to be approximately 23 times higher than V4 Flash ($1.37 per million tokens, just 6 cents), meaning Opus likely still accounts for the majority of spend.
These numbers don’t even capture the latest arrival, Nvidia’s Nemotron, which is poised to jump to the front of the pack thanks to Nvidia’s strong connections and the great adaptability of the model itself.
While these numbers don’t completely prove Zhang’s point about the AI lifecycle, they do show that cutting-edge research institutions like Anthropic haven’t suffered too much from the rise of open source, at least not yet. One explanation is that the market for AI-addressable tasks is growing so rapidly that top models can maintain their position simply by dominating early adoption. Zhang said, “Frontier Laboratories will continue to own the discovery. Open source will increasingly own the production.” Another explanation could be that even if clients move to open source, many use cases are too difficult to completely replace with cheaper alternatives.
In any case, this two-tier model economy may become a relatively stable feature of the AI economy.
Back in September of last year, I wrote about the possibility that Foundation Research may end up selling coffee beans to Starbucks. This means that it acts as an input for the product while the application layer reaps the benefits. Some of those predictions came true. For example, the vertical AI play has switched to lightweight models, and the economic situation for “GPT wrapper” startups has largely stabilized.
But we also see that frontier providers are able to maintain premium token prices on a per-token basis, which is the most desirable part of the market. And that’s not likely to change any time soon.
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
