Cheaper tokens do not mean less compute, they mean more of it. @GavinSBaker is exactly right here because an open weight token uses roughly the same compute and energy as a closed model token when the models are a similar size.This chart from one AI gateway shows how fast the shift is happening. Open weight models were a minority of token volume in June, and by mid-September they made up 78.4% of it. While the spend data tells a very different story. Closed models like Claude Opus 4.8 and GPT 5.6 Sol still take the biggest share of the dollars, even though they are losing share of the tokens. That gap is the whole point. Developers are moving their volume to cheaper models, but they are not using less AI because of it but using a lot more. Ramp's data in Gavin's chart shows the same pattern. Token volumes rose about 7x from January to September, while AI spend actually fell over the same period. This means the model layer is likely to see its margins squeezed as prices keep falling. Gavin also points out that open weight tokens are some of the highest margin tokens for the cloud providers that serve them. Every one of those tokens still has to run on GPUs, memory and networking gear. Cheaper models make AI more affordable, and that pulls even more demand into the data center. That is why I think the infrastructure layer keeps winning no matter which model wins. The companies I am watching most closely are $NVDA, $AMD, $AVGO and $MU. Bullish on AI infrastructure. If you want to see how I’m investing across compute, memory, and networking, check out my full Milk Road Pro portfolio below. link.milkroad.com/d8cj6v