The Fable 5 shutdown demonstrated that cloud-hosted AI models can be disabled globally through a single regulatory action, shifting open-source AI from an ideological preference to a practical resilience strategy. Supporting infrastructure now exists at scale, including Meta’s Llama, Mistral, and Qwen models, with tools like llama.cpp and vLLM enabling inference across multiple hardware backends without Nvidia dependency.
While Nvidia retains dominance in model training, inference workloads are increasingly competitive, with AMD and Apple Silicon closing the gap at the edge. However, training foundation models still costs hundreds of millions of dollars, meaning true openness is limited to inference — the weights are free, but generating new ones remains accessible only to well-capitalized organizations.
The Fable 5 shutdown clarified something that had been theoretical: if a government can disable a cloud-hosted AI model for every user on the planet with a single letter, the only AI you control is the one running on hardware you own with weights you can actually download. Open-source AI stopped being an ideological position and became a resilience argument in about twelve hours.
The ecosystem to support that argument exists now in a way it didn’t two years ago. Meta’s Llama family, Mistral’s open-weight models, Alibaba’s Qwen. Hugging Face hosts over 500,000 models. llama.cpp runs on Nvidia CUDA, AMD ROCm, Apple Metal, and CPU fallback. vLLM handles multi-GPU serving without touching the Nvidia ecosystem if you want to avoid it. The tooling is not the bottleneck. An AMD-powered EVO-X2 mini PC delivers usable inference without an API key and without an Nvidia card. That was not true at the start of 2024.
The hardware consequence of most coverage misses is that when inference frameworks support multiple backends, GPU vendor lock-in weakens at the edge, even if Nvidia still dominates training. CUDA’s moat is in model training, where the largest clusters still run H100S and B200S, and nothing else is close. At the inference layer, where the actual production workload ultimately sits, the market is more competitive and becoming more so. Open models running on AMD or Apple silicon are how the CUDA advantage gets chipped away over time.
European governments watched the Fable 5 shutdown in real time, and some of them are now asking a specific question: which AI can we run that no foreign government can shut down? The Red Hat model, which gives the weights away and sells enterprise support around it, is the business structure that aligns most naturally with that requirement. Mistral is building exactly that.
The thing nobody wants to say: Meta, Alibaba, and Mistral release open weights from a position of capital that no startup can replicate. Training a 70-billion-parameter model from scratch costs hundreds of millions of dollars. What’s “open” is inference. Training remains a privilege for companies with the GPU clusters to afford it, which means there are still gatekeepers in the open-source AI economy; they just operate before the model is released rather than after. Downloading the weights is free. Generating new ones costs money that few organizations have.