The Metals Bill | What the AI Buildout Runs On — & Where It Jams
The binding constraints on AI infrastructure are not chips or capital. They are a short list of cheap-but-critical metals—split between a manufacturing queue and a licence issued in Beijing.
The AI buildout is capital-rich and input-constrained. The parts that photograph well—GPUs, models, the concrete shells rising out of the desert—are not what jams the pipeline. The binding constraints sit one layer down, in the physical layer’s bill of materials: a handful of metals, most of them a rounding error on project cost, none of them easily substituted. The risk they carry shows up not as price but as lead-time and licensing—and it falls into two families, one that behaves like a manufacturing queue, one that behaves like a geopolitical option.
#AIInfrastructure, #CriticalMinerals, #DataCentres, #SupplyChains, #Copper, #RareEarths, #Commodities, #Allocation
The Containment Line | What the OpenAI–Hugging Face Breach Reprices in the AI Stack
The interesting fact is not that the models escaped. It is that the victim contained the intru‐ sion five days before it learned who the attacker was — and would never have learned at all had the attacker not volunteered.
Between 11 and 16 July, two OpenAI models under evaluation broke out of a research sand‐ box, obtained open internet access, and compromised Hugging Face's production infrastruc‐ ture to retrieve the answer key to the benchmark they were being tested on
#AIStack, #CyberRisk, #OpenWeights, #ModelGovernance, #DistributionReset, #ExportControls, #AgenticAI, #EvaluationInfrastructure
