On Monday, Decagon CEO Jesse Zhang revealed a provocative new idea, posted underneath the title “Everyone is wrong about open source AI in the enterprise.” The put up grapples with probably the most fascinating contradictions of in the present day’s AI economic system: Extra mature AI deployments are switching to lighter fashions, he says, even at his personal firm. However the general spend on costly state-of-the-art fashions has barely budged.
It’s a brand new means to consider the connection between frontier and open supply fashions. In Zhang’s telling, they aren’t opponents, and open supply fashions’ success isn’t coming on the expense of frontier labs. As an alternative, they’re two phases of the identical life cycle, with costly frontier fashions getting used to show out use instances that may be handed alongside to cheaper open supply alternate options as they mature.
As extra mature use instances switch to lighter models, new use instances hold arising — and the general spend on frontier fashions barely goes down.
Zhang doesn’t give a lot information to assist the purpose, however the information isn’t arduous to search out. Vercel’s AI gateway dashboard reveals that, in simply the previous week, DeepSeek has surged into the lead for token volumes, now processing simply over a 3rd of the tokens passing via the corporate’s infrastructure. Z.ai — the lab behind the favored GLM-5.2 mannequin — jumped into a decent fourth place over the identical interval.
However should you scroll all the way down to general token spend, you’ll see Anthropic nonetheless accounts for greater than half of the general AI spend on the platform. On condition that a lot of the latest change comes from Anthropic’s personal rising costs, the share has dropped barely over the previous month, however not considerably.
OpenRouter tells an identical story, capturing a a lot bigger (however barely much less enterprise-y) phase of the market. DeepSeek V4 Flash is the primary winner on general utilization, processing 5.3 trillion tokens weekly. The most well-liked frontier mannequin, Opus 4.8, handles simply over 2 trillion. OpenRouter doesn’t rank fashions by whole spend, however it registers the common token value for Opus 4.8 as roughly 23x larger than V4 Flash ($1.37 per million tokens, in comparison with simply 6 cents), which might imply Opus was nonetheless most likely capturing the lion’s share of spending.
These figures don’t even seize the most recent arrival, Nvidia’s Nemotron, which is poised to leap to the front of the pack by advantage of Nvidia’s robust connections and the mannequin’s personal excessive adaptability.
These figures don’t absolutely show Zhang’s level concerning the AI life cycles, however they do present frontier labs like Anthropic aren’t struggling an excessive amount of from the rise of open supply — at the very least not but. One clarification is that the market of AI-addressable duties is rising so quick that the highest fashions are in a position to keep their place simply by dominating early-stage deployments. As Zhang places it, “The frontier labs will hold proudly owning discovery. Open supply will more and more personal manufacturing.” One other clarification could be that, whilst purchasers transfer to open supply, many use instances are so troublesome that they’ll’t be solely changed with cheaper alternate options.
Both means, this two-tiered economic system of fashions might turn out to be a comparatively secure characteristic of the AI economic system.
As just lately as final September, I used to be writing concerning the chance that basis labs would find yourself selling coffee beans to Starbucks — that’s, serving as commodity inputs whereas the appliance layer reaped the advantages. Some elements of that prediction got here true: Vertical AI performs switched to lighter fashions, for one, and the economics of “GPT wrapper” startups have remained largely secure.
However we’re additionally seeing that, token for token, frontier suppliers have been in a position to maintain on to probably the most fascinating a part of {the marketplace} — the premium token worth. And that doesn’t appear more likely to change any time quickly.
Whenever you buy via hyperlinks in our articles, we may earn a small commission. This doesn’t have an effect on our editorial independence.
