as I read it, it's an attempt at reference to economy of scale under the thesis "AI silicon will keep getting cheaper because more and more people will produce it" as the main underpinning for how to reduce their unit economics. which, y'know, great! that's exactly what people like to hear about manufacturing and such! lovely! it's only expensive because it's the start! oh, the woe of the inventor, the hard and expensive path of the start!
except that doesn't hold up in any reasonable manner.
they're not using J Random GPU, they're using top-end purpose-focused shit that's come into existing literally as co-evolution feedback from the fucking industry that is using it. even some hypothetical path where we do just suddenly have a glut of cheap model-training silicon everywhere, imo it's far far far more likely to be an esp32 situation than a "yeah this gtx17900 cost me like 20 bucks" situation. even the "consumer high end" of "sure your phone has a gpu in it" is still very suboptimal for doing the kind of shit they're doing (even if you could probably make a great cursed project out of a cluster of phones doing model training or whatever)
falls into the same vein of shit as "a few thousand days" imo - something that's a great soundbite, easily digestible market speak, but if you actually look at the substance it's comprehensive nonsense
this is fantastic, more of this kind of thing is definitely good