Handshake, a company that has moved from launch to roughly $1 billion in annual scale within a year, is building a business around specialized data for frontier AI labs. In a Latent Space podcast deep dive, co-founder Garrett Lord and colleagues walk through what those labs actually buy, using examples from finance, biology and AI alignment. The episode is presented as the first in a series examining data providers that supply the training and evaluation material used to build leading models. The discussion points to a shift in how frontier models are improved. Rather than relying only on broad web-scale data, labs are paying for narrow, expert-curated datasets that test or teach specific skills, such as financial reasoning, biological knowledge and the behaviour that alignment work aims to shape. The show notes do not give contract values or customer names, so the scale of individual purchases and the identity of buyers cannot be confirmed from this source alone. For Australian businesses and professionals, this changes little day to day. Companies that simply use hosted AI models are not buying this data and face no new cost. Firms that fine-tune models or build internal evaluation sets should check two things: whether any training material contains personal information, which brings obligations under the Privacy Act 1988, and whether the data they source from vendors is accurately labelled and licensed. Worth watching, but no immediate action is required for most organisations.