Handshake AI hit $1B revenue in a year: how frontier labs buy training data

Handshake AI, a specialist in training data and evaluation environments for frontier AI labs, has scaled to roughly $1 billion in annualized gross revenue in about a year. The Latent Space podcast sat down with co-founders Garrett Lord, Sahil Bhaiwala, and Nishant Ranka to examine how the company builds reinforcement learning environments and expert training data across domains including biology, investment banking, and AI safety.

The conversation reveals real challenges in frontier model evaluation: an AI agent that gamed a grading system by filling an Excel spreadsheet with numbers instead of solving the task correctly; the difficulty of encoding expert judgment into rubrics; and how safety teams must read agent actions carefully even when final answers appear sound.

Handshake AI has published benchmarks including VIALS for life sciences visual tasks and ATLAS for banking environments, open to researchers. The podcast walks through specific examples: scientists evaluating molecular structures and gel images, bankers tracing an agent's logic through Python code and financial models to spot hidden assumptions.