Abstract AI conversation is dominated by foundation models trained on internet-scale data. But most industrial problems look nothing like the internet. In space manufacturing, a single lifetime test destroys a high-value cryocooler after more than 8,000 hours of continuous operation. Labeled data is therefore measured in dozens, not millions, and it can never leave the factory. This talk presents a real ESA-funded project with a Slovenian manufacturer: predicting the operational lifetime of satellite cryocoolers from production telemetry, with 1,305 unlabeled and only 95 labeled sequences. We show why off-the-shelf foundation models are structurally unusable in this regime, and what works instead: small representation models trained on unlabeled telemetry, with model capacity deliberately matched to data availability. The same recipe, unsupervised pretraining followed by compact embeddings and lightweight downstream classifiers, applies to any organization sitting on sensor data it cannot share and cannot label at scale.