Handshake AI’s data business is nearing $1bn annualised revenue, Latent Space says

Handshake AI is building a large business around the expert data and simulated environments used to train and evaluate AI systems. Latent Space says the company reached nearly $1bn in gross annualised revenue in roughly a year, a striking sign of the market forming around what frontier labs need beyond more compute.

Watch Desk analysis

What happened

Latent Space’s deep dive with Handshake chief executive Garrett Lord and colleagues covers training environments for biology, finance and alignment. The episode describes work ranging from scientific-image tasks to a simulated investment bank, where an AI agent builds financial models and a banker grades its spreadsheet.

One example shows why these environments are more than a supply of labelled answers: an early grader rewarded an agent for finding a target figure, even though a real investment banker would check much more. The discussion also covers how labs evaluate datasets, where rubrics fall short of expert judgement, and why some enterprises want private benchmarks. Read and listen to the Latent Space deep dive.

Why it matters

Training data is often discussed as a pile of examples. This account points to a more involved business: designing realistic tasks, deciding what counts as a good answer, and testing whether a model can do the work rather than merely satisfy a convenient grader. That matters wherever AI systems are being judged on specialist work, from finance to biology.

The reported revenue figure also suggests this market has grown quickly. It is gross annualised revenue as described by Latent Space, not a disclosed audited result, but it gives a sense of the commercial scale being claimed.

Our read

The interesting product here is not simply “more data”. It is a better test of whether an AI can handle the messy standards experts use outside a demo. That is a useful business to watch, though a good benchmark still needs to measure the work people actually care about. A model should not get a gold star for finding one number in a spreadsheet and miss the rest of the bank.

What to watch

• Whether Handshake’s reported growth translates into sustained business with frontier labs.
• How expert judgement is built into graders for specialist tasks.
• Whether private enterprise benchmarks reveal capabilities that public tests miss.
• How buyers distinguish convincing benchmark scores from useful performance in practice. Activity teaser: Handshake AI’s data business has reached nearly $1bn in gross annualised revenue in roughly a year, Latent Space says. Its deep dive explores what frontier labs buy: not just examples, but simulated tasks and expert evaluation across areas including biology, finance and alignment. One banking example shows how a weak grader can reward an agent for finding a target figure while missing the broader judgement a banker would apply. That is a reminder that better AI testing depends on asking better questions. Should companies spend more on expert-built evaluations, or is scalable automated grading good enough?

Discussion spark: Should AI companies spend more on expert-built training environments and evaluations, or is scalable automated grading good enough?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.