Yup.. imo, the labs’ compute capacity for any given period of time should be viewed as fixed, and has to be allocated to either R&D or inference (COGS). Model safety/alignment is a sub-segment of R&D. So all else equal, if they allocate more compute from within the R&D bucket to safety/alignment (e.g. shifting R&D spend away from areas like pushing the frontier on RSI).. Or from the inference bucket to R&D for safety/alignment.. it’s probably net negative for near term revenues. But over longer period of time, safety/alignment should be a capability representing a source of incremental demand for compute.. Zuck articulated it well that safety/alignment really is just a standard product feature. And seems pretty clear the labs don’t have enough inference (or R&D) capacity as it is. So they’re going to need even more net new compute.