Third-party inference providers are going to war. opened GLM 5.3's weights on Aug 28. Five weeks later, 32 hosts are serving it on OpenRouter and the price floor has collapsed. Output price fell 89% at the extreme, from $4.40 ( to $0.49. Cheapest today is $1.32, still 70% under list. Open-weight labs are already reacting. moved GLM 5.3 itself off MIT. Alibaba shipped Qwen3.7-Max API-only with no weights. MiniMax changed its license to block commercial use without written authorization. One thing stays constant: someone has to own the hardware. Today that's whoever can finance it, which means a handful of giants. The GPU underneath is the primitive that lasts. That's why we're building a GPU vault. Own a share of a real GPU node, get paid from the rent it earns.