There's a data point I recently just saw today and I think it's being badly misread, and the misreading matters if you're trying to work out where the AI money goes from here.
Vercel's AI Gateway, which routes real production traffic for a lot of developers, showed open weight models hitting about 78% of token volume on September 18, up from roughly 11% in April and 62% in late August. Closed models, the paid APIs from OpenAI, Anthropic and Google, have fallen from around 70% of that volume in June to about 22%. Vercel's CEO flagged it himself as possibly a record day for open source. The easy and popular conclusion is, open is eating the closed labs, so the labs are in trouble and value shifts down to infrastructure.
I don't think that conclusion is wrong exactly, but it skips what actually matters, and the same CEO said it in the same thread. Even with open models running the majority of tokens, closed models still take the overwhelming majority of the spend. Anthropic alone was around 64% of gateway spend. A Mozilla report from mid September put open weight models at the majority of developer traffic on another big router while capturing something like 4% of model layer revenue. So the tokens went open and the dollars stayed closed. That's what happens when open models get pointed at the cheap high volume work, while the expensive reasoning that people pay for stays on the frontier APIs.
The bull case for the infra crowd, if open weights keep taking volume, the model layer commoditises, pricing power leaks out of the labs, and the margin has to land somewhere. The people running the compute and the routing in between look like the natural beneficiaries, and the total token pool is still growing fast on top of that.
But saying infra wins is too vague to trade, the margin doesn't accrue to some abstract infrastructure, it accrues to whoever owns the chokepoint, and those chokepoints are being bought up by the same incumbents the open movement was meant to route around. Stripe paid roughly 7.5 billion for OpenRouter, the main marketplace open models flow through, back in August. The hosting and tooling layer has been consolidating for a year. So the independence story and the investment story point in opposite directions, and a lot of the cheaper AI saving may just be relocating the toll booth.
If open really is repricing the industry, open weight spend share should start climbing toward its token share over the next couple of monthly gateway indices. If tokens stay near 80% while spend stays stuck in single digits into year end, then the open winning through volume thesis never becomes an earnings story, and the margin capture at the hosting layer is the actual trade.
This is one gateway and a single record day, and other trackers had open versus closed closer to 50-50 on request share earlier in the summer, so don't anchor on the exact 78%.
What I'm trying to settle on is if open model volume dominance ever convert into open model dollars, or is usage share just permanently a different thing from where the profit pools sit? Smarter brains please drop your thoughts.