Handshake AI reaches $1b in gross annualized revenue selling training data

Frontier AI labs are spending real money on the specialists who label their training data. Latent Space published an interview on 9 November 2026 with Garrett Lord and the Handshake AI team. The podcast reports that Handshake went from launch to $1 billion in gross annualized revenue in one year. Handshake AI is the training-data arm of Handshake, the US college job board.

The host, swyx, also cites earlier reporting from The Information on the same figure. Sacra, a private-market research firm, carries a separate note. It says Handshake's group annualized revenue hit $1.1b in April 2026. That is up from $550m before. At launch it was $5m to $10m. The AI business is the growth engine.

What the labs actually buy

The podcast walks through three product lines. The Handshake team names them one by one, with worked examples.

The finance line runs simulated banks. An agent reads emails and documents, then books journal entries. Collin Schweiker, who leads the finance line, describes one case. The agent spotted that a goodwill deduction was needed. It then failed to apply the rule in the generated code. That gap is the training signal Handshake sells.

The biology line asks medical experts to grade a vision model on tissue samples. Nick Magazine, the life sciences lead, says the model picked the correct slides for advanced vessel necrosis in brain tissue. It also picked one extra slide that did not show the feature. That false positive is the fact the labs want recorded.

The alignment line watches what an agent does with a sandboxed file system. Steve Kelling's team logs whether the agent reads the file or edits it. It also logs whether the agent then tries to clear the shell history.

The business under it

Sacra reports that after paying the experts, net annualized revenue is about $450m. OpenAI and Anthropic are the primary customers. Handshake's own site says it has paid out more than $500m to over 188,000 fellows. Pay runs from $17 to $150 an hour.

The shape that matters for a founder in AI is this. Handshake's old product has 750,000 users. A layer of them (PhDs, lawyers, radiologists) can sell their judgement for a rate the labs will pay. The job board is now an inside channel for recruiting labelers at scale.

What it means for someone selling AI tools

Model quality is still bought one expert at a time. Say a buyer is looking at an AI product that claims to do specialist work. That could be reading a loan file or grading a tissue sample. The buyer should ask who labelled the training data and how. Handshake's product is a map to those questions. The labs are buying it for a reason.

Source

What Frontier Labs Buy: A Data Deep Dive with Handshake AI, Garrett Lord and team, by Latent Space.