The duopoly's moat is a token-cost question now

The AI duopoly's moat is no longer a model-quality question. It is a token-cost question, and it is being decided in public arithmetic this quarter.

The Anthropic reported run-rate, closing in on $70B+ ARR against a $100B path, is the cleanest single test of whether OpenAI and Anthropic can hold pricing against open-source models reportedly running 80 to 90 percent cheaper on the same workloads. If the frontier API pricing survives that compression, the moat is real. If it does not, $100B is a story rather than a number.

Sacks, on All-In, read Leopold Aschenbrenner's reported margin call as a leverage and momentum unwind rather than a break in AI capex fundamentals. He is the careful reader. The bet Aschenbrenner ran, $225M to as much as $45B at 3.5x leverage on the AI trade, tested the same thing every concentrated bet tests: how long can a position survive a drawdown before the lenders force the question. Citadel's reported buyout of the portfolio is what that unwinding looks like in practice.

Calacanis added the load-bearing data point. He reports real token-cost migration in his own portfolio: open-source models running 80 to 90 percent cheaper than frontier APIs, on workloads that used to be frontier-only. That is the mechanism the duopoly has to beat, not benchmarks.

Watch the ARR line. That is the falsifier.