OAI and Ant are also at ~30-40% inference with the rest allocated for training It’s one of the most important assumptions in all of compute/tokenomics because it dictates how much compute is monetized and the monetization rate As these guys all scale from 1-2 gigs to 5 GW and to 10+ in coming years if you believe the ramp we’ll start to see natural deceleration in monetization rates or MUCH larger revenue bases Example: 5GW estimated YE for Ant and OAI, 40% inference = 2GW. If they’re at $100BN implied ARR means blended rate is $20BN GW ($100/5) but they’re monetizing at $50BN per GW ($100/2). Also crucially important to dictate economics as compute remains constrained. Eg if you pay $10B, $20B, even $50B for compute you need to charge MORE than that as a model/lab company selling tokens per GW to make money. Bake in the incremental costs of serving, as well. This is why when you go from 5 to 10GW, folks are saying Ant needs to show $100Bn then $200Bn otherwise definitionally you are monetizing inference at a lower rate UNLESS they allocate even less compute to it. But given the size and scale, if anything, imagine they try to monetize more