Deepseek seems to prove there's no super secret sauce that makes these models irreproducible outside the US and that the companies here are suffering a bit from the glut of cash/credits leading them to burn tons of extra processing power that could have been optimized away.
A nice model, does not a billion dollar company make. The hard part of AI is not the model; Apple needs people to do the 80% rest-of-the-work; how do you make AI useful to the average person? How can we get inference on edge devices as cheap and efficient as possible? Models are boring. Everyone fully expects that we'll see an N% increase in intelligence every six months now. Yawn. The exciting thing now is: What are we using AI for?
Spending a couple billion dollars also doesn't make an actual billion dollar company. It's yet to be proven that all this spend on LLM training and running can actually get translated into an actual profit.
The US AI industry does have easier access to them but again Deepseek proves countries outside the US can get enough access to them to produce similarly powerful models. Unless the US really clamps down on exports of cards, which will be hard given they're made overseas and only designed in the US, it's not like they're a unique resource only US companies can access.
