For two years the story was one line. Inference keeps getting cheaper. Tokens fell by thousands of times and nobody expected that to turn. Look at the hardware under those tokens though. H100 rental is up about 40% in six months. On-demand capacity is nearly sold out. The wait for a cluster runs 12 to 18 months. What’s eating it is inference and agents, firing far more calls per task than anyone planned for. Cheap tokens, scarce compute. Send a request to the wrong model and you pay twice. Once at the model’s price. Again in capacity that half the market is queuing for. Having the strongest model isn’t the edge anymore. Not wasting scarce compute is. Each request on the cheapest model that can handle it, nothing frontier-priced that didn’t need to be. That’s the whole reason the auction exists.

i challenged people at TOKEN2049 to name every sponsor on my suitcase for $100 would you have won?