Every few weeks another open model claims frontier level results on a fraction of the compute. At what point is that a real dent in the Nvidia capex thesis, and not just single benchmark cope? Change my mind.

I keep trying to stress test my own long thesis on the AI capex names and I cannot decide if I am watching a moat crack or just reading marketing.

The pattern is hard to ignore now. In January 2025 DeepSeek reported training a frontier model for a fraction of the usual spend and Nvidia lost about 589 billion in market cap in a single day, the largest one day loss on record, on the fear that frontier capability might not need frontier spend anymore. Then GLM-5.2 landed as a 744 billion parameter open weight model that the company says was trained entirely on domestic Huawei Ascend chips, which is a claim I cannot independently verify but which points the same direction. And it is not just language models. On the embodied side, Robbyant put out a vision model called LingBot-Vision whose smallest 0.3B version they report matching Meta's 7B DINOv3 on depth, roughly a twentieth of the parameters.

If the same result on far less compute story is real and repeatable, the bear case writes itself. The capex line that justifies the multiples assumes compute stays scarce and pricing power holds. Every credible efficiency jump chips at that.

Here is why I have not sold anything. Every one of these numbers is self reported and cherry pickable. LingBot-Vision only edges DINOv3 on one depth benchmark and actually loses on KITTI and on ImageNet, so the smaller model is not universally better. DeepSeek's headline training cost was never independently confirmed. And the Jevons angle cuts the other way. Cheaper inference has so far expanded total compute demand rather than shrinking it, which would be good for the very names the bear case shorts.

So which is it. Are these efficiency claims a real crack in the capex moat, or single benchmark cope that Jevons quietly eats? Change my mind.