美国AI前沿的赌注：烧钱竞赛与赢家通吃

these arguments are broadly sensible but the bet on the US AI frontier is simpler: they can outspend the competition, the competition eventually taps out, and they do eat the economic light cone.

Indeed I think the bear argument is self-refuting. If all labs need to 10x their spending annually to remain competitive, *but* the frontier gets the lion's share of revenues (at higher margins too!), for the normal reason that AI market is a superstar market and the best intelligence/second is the best value for the money (case in point: me spending on Fable despite many models being theoretically "better bang for the buck"), then eventually - and very soon – the runners-up just can't justify new funding rounds, and so don't have the capital to keep up, neither capex nor opex. This has already gutted several American startups; I am endlessly impressed by the chutzpah of Chinese labs that stare at this reality and do not blink. No idea how they fare in the long term, China is a magical land; but Wenfeng is pretty insistent that at the end «2 big and 2 small» is all that should survive there, too. And in the West, it might just be 2 big. Moreover Anthropic is already narrowly profitable, or close to that anyway. Consider: frontier AI inference is priced not against the competition in AI but against the hour of white collar labor, and not even against the compensation but increasingly against the revenue this hour buys your employee. How much is that worth? Say, $2.2M/year at Meta. That's $251/hour, at 24/365 utilization. Meta tech staff actually reportedly clocks 2,235 hours, which is nearly $1000/hour. That's a lotta tokens. That's actually over a billion output tokens of DeepSeek-V4 at $0.87/1M, and V4 is served with like 80% margins, and it'd be 95%+ on Vera Rubin. Hopefully you can see that there's a lot of distance to cover yet. I also don't get what spikiness matters; AI is an amplifier of the capability of your remaining human staff, and we see that better staff becomes dramatically more productive, in a way that – perhaps to some limit – allows you to keep scaling your returns by buying more and more frontier tokens, even granting Anthropic their $1T revenue by EoY 2027. Who knows.

I think Wenfeng is correct to say that the main realistic reason to be bearish on this scenario is political – some "lab" just won't be allowed to East India Company the rest of the economy, it'll be torn apart and brought low by the state, or by the coalition of its incumbent complements. But by default, ignoring fearsome sociopolitical dialectics, Dwarkesh makes sense; much as it pains me to say so.