Comment by paulbgd
14 hours ago
As I was reading the start of your argument, I thought you were gonna call the models a depreciating asset! Totally agree about GPUs too, but literally everything they’re spending money on has to be rebuilt to stay competitive. They have to go for the moonshot of training a full new model when better tech comes, they have to upgrade GPUs to keep their data centers efficient.
Technically, the model is a depreciating asset too. Just consider the difference between a model you need a B200 cluster to run vs one you can run on a Raspberry Pi. One's going to have a moat around it that gives it value and the other isn't. It's a hyperbolic argument to be sure but the nature of "enthusiast" hardware is that we're currently running, say, ~27B parameter models on hardware for a few thousand. What's that going to look like in 2 years?
Anthropic/OpenAI really need to train ever-bigger models to keep their moat. But that assumes there isn't a law of diminishing returns and also that a compressed model isn't sufficient for what many people need.
You mihgt say that the training is a barrier. And it is, kind of. Notice how it's Chinese companies coming out with open-source models like DeepSeek and Qwen? That's no accident. As soon as DeepSeek came out I knew what was going on: China is going to make sure no single Western company "owns" AI. It's in their national interest for that not to happen.
I wouldn't be surprised if the rush-to-IPO is motivated, at least in part, by getting ahead of Chinese AI commoditization.
How many failed foundation model training run cycles do you think these companies can tank before the bubble pops and deepseek/etc. catch up to frontier quality?
If Ant, OAI, etc. aren't able to make 20-30% improvements on Opus 4.6 in 2026, does the music stop playing altogether? It seems like they'd lose their ability to charge >10% gross margin on inference in a span of 3-6 months.