great shoutout from @chamath on the all in pod today. "you need to go to a vendor that you trust, could be Nebius" perfect recap of why, if your data actually matters, an open source inference path like @nebiustf is the way to go his point, in plain english: you paste a protein design, a deal strategy, a process that is the company into a shared chat api. The vendor says “zero data retention", but that usually means best efforts, not a hard guarantee. then someone hits like / thumbs-up, or the terms allow training on “insights,” not names, or the next model version just… knows the shape of the answer you thought was private (like in the Navier Stokes case) the scary leak ends up being the solution pattern showing up somewhere you don’t control. so the move isn’t “never use ai”, it’s stop putting crown jewel work on a shared frontier api stand it up on your terms: open weight models, inference on hardware that’s you can control, a vendor that provisions your stack, the lane Chamath pointed at (aws class, @nebiustf, @FireworksAI_HQ, etc.) Boards are late to this. “We signed zdr” is starting to look thin next to “do we actually control where this runs.” That’s when the CIO who took the easy api deal becomes the problem