Rauch says open models hit 78.4% of Vercel gateway tokens
Moonshot AI and DeepSeek ranked third and fourth on spend, and Rauch says adding Z.ai pushes the group past OpenAI.
Guillermo Rauch says today may be a record day for open models on Vercel AI Gateway, with open weights taking 78.4% of token volume against 21.6% for closed models.
While spend 馃挷 usually tells a different story, #3 and #4 today are Moonshot AI & DeepSeek.
Tokens are one measure. Money is another, and Rauch is careful about the difference. He notes that spend usually tells a different story, but on this day Moonshot AI and DeepSeek landed at third and fourth by spend. Add Z.ai to the pile and he says their combined spend passes OpenAI, which sat at number two.
He also flags what that spend figure is not. It is the cost of running inference for those models across providers, mostly in the US, not revenue going to the open weight labs themselves. Somebody is hosting those models and taking the margin, and it is not necessarily the people who trained them.
The business read
One gateway on one day is not the market. Vercel AI Gateway routes traffic for a particular slice of developers, weighted toward people already building on Vercel, and a single day can move on one large customer shipping one large job. Rauch himself hedges it as a record that may have happened, not a trend he is declaring.
What makes it worth two minutes anyway is where the number sits. If cheap open weights are genuinely eating into what founders send to closed frontier models, gateway traffic is the first place it shows up, before anyone publishes a revenue chart. Token share moving ahead of spend share is exactly what a price-driven migration looks like from the outside, high volume going somewhere cheap while the expensive calls stay closed.
The gap between the two numbers is the part to sit with. Open models can be 78.4% of the tokens and still not be where most of the money goes, because the closed models are getting used for the work people will pay a premium for. For a solo founder running a product with real inference costs, that is the actual decision, which jobs are cheap and high volume and can go open, and which ones you still pay up for. Nobody in this thread put a line on that yet.