Post

AI has lost the Mandate of Heaven I’m afraid. WSJ saying it’s a “when” the ai bubble bursts. SNL dabbing on their biggest names for being weird. FT says 50% of chips sold in 2026 and 2027 have nowhere to go. Hundreds of billions short on revenue for even 10% ROIC. A disaster.

It’s just a matter of time now, and mostly a matter of when the money runs out. There is no future for Anthropic or OpenAI at their current scale. Hyperscalers need $308bn of annual ai revenue - around the revenue of Microsoft - for 10% return on 2026 and 2027 capex. Could be more with cost of debt.

The level of financial misalignment is as bad as I’ve been warning, and you’re going to see headline after headline of stuff I’ve said in the newsletter for years. I say this with both pride and sadness because being right is going to be many layers of horrible.

calling my shot now that if any of these boxed GPUs sold at pennies on the dollar flow to computational chemistry researchers at scale, you're going to get breakthroughs in biochem that you can only dream of because convincing DOE or whoever to get grants for more compute is impossible

And in order to get actual actionable science, the data work to train on clean data sets is insane. Can't just dump everything in a hopper like they have and underpants gnomes the way to scientific discovery.

normally I don't engage with these but i'll make something clear vast majority of comp chem is simplifying down the most accurate computations we know of in order to make it tractable with existing compute resources; we *can* switch to more accuracy but then the computation takes forever

so if you can upgrade existing racks with even just a couple of these for cheap, the throughput you get allows you to just change a couple settings in your CHARMM/QChem setup and get results you'd salivate over, previously