Handshake hits $1 billion in AI data sales in a year

Handshake's rapid climb to $1 billion in gross annualized revenue shows that frontier labs are willing to pay for more than scraped text. The company builds reinforcement learning environments, which are simulated tasks where an AI agent must act, and supplies expert data to grade its work. In a finance demo, an agent had to find a deal assignment in a bank's email and build a workbook; it failed when it carried a goodwill figure into Python and subtracted it, even though it could explain the tax treatment when asked directly. In biology, Handshake's VIALS benchmark asks models to interpret scientific images, such as identifying brain-tumor tissue with the most advanced vessel necrosis. The model picked the two correct slides but added a third, showing how evaluation goes beyond a simple score. These examples illustrate why labs hold out evaluation tasks and why expert judgment remains hard to codify.