A former Tesla engineer and a senior engineering manager at Phaidra, Raahul Singh and Vanč Levstik, described on the AI Engineer podcast how using large language models to search across industrial sensor networks in a data center led to hallucinations and omissions. In a facility with 64 GPUs, the model got answers correct 80% of the time, but at a scale of 460,000 GPUs across a gigawatt of power, accuracy fell to about 30%. The model invented equipment that did not exist and omitted real ones, a problem termed "semantic blindness." The engineers rejected the industry's prevailing wisdom and built a redesigned architecture called Fedra. Instead of asking the LLM to sift through 500,000 names, the system asks it to produce a plan, which a deterministic backend then executes on pre-indexed equipment hierarchies. The LLM only sees a compact summary of the factory's tree structure, exploiting the property that data centers grow wide, not deep, with organizational depth remaining at four to six levels.