Last week’s US government restrictions on Anthropic’s model exports are just the latest sign that the game has changed. AI industry watcher Andrew Curran writes that traditional benchmark scores no longer measure real intelligence. What matters now is whether a model can deeply infer user intent, autonomously reason through complex tasks, and feel alive in conversation. That’s the real moat.
https://twitter.com/AndrewCurran_/status/2066332670817456584
And the leading models are already training their successors, kicking off an exponential acceleration. Back in 2023, Anthropic’s own fundraising documents predicted that the company that trains the best 2025–2026 model would build a lead so large that rivals could never catch up.
That prediction has become reality. The golden window for any organization to build its own top-tier AI was just three years — 2023 to 2026 — and it’s now shut. In that brief period, only Elon Musk managed to push to the frontier in 26 months. Almost no other country made a serious attempt. Europe, bogged down in regulatory quicksand, missed the moment entirely.
Curran forecasts that by 2030, the compute needed for training will surge a thousandfold, and both compute and talent will concentrate overwhelmingly in the US and China. Top-tier computing power will become a national security asset, like uranium. The US will guard it with export licenses and political moves. Any national AI that relies on someone else’s underlying tech is really just leasing a branded copy of a US or Chinese model. If that supply gets cut off, a country’s entire infrastructure — healthcare, education, military — could collapse in hours, like after an air strike.
And there’s another twist: once open-source models approach the frontier later this year, they may face coordinated global crackdowns from governments terrified of losing control.