📰 Key Summary

AfterQuery, a startup specializing in AI training data, is reportedly valued at $3.2 billion after a new funding round — just five months after announcing a $30M Series A at a $300M valuation back in April. That’s a valuation jump of more than 10x in under half a year. According to Y Combinator partner Gustaf Alströmer, this is the fastest a startup has ever gone from founding to unicorn in the accelerator’s history. AfterQuery’s two co-founders, now 22 and 23 years old, were part of Y Combinator’s Winter 2025 batch just 18 months ago. Back in April, the San Francisco startup said its annualized revenue had already hit $100 million, working with several major AI labs — customers named include Nvidia, Legora, and Korean AI lab Motif Technologies. AfterQuery is part of a new wave of startups following in the footsteps of Mercor and Scale, hiring doctors, lawyers, and other domain experts to help train models. But unlike traditional efforts focused on getting models to answer questions correctly, AfterQuery trains models and AI agents to learn how professionals actually do their work — what the company calls “encoding the patterns, decisions, and reasoning of the world’s top practitioners into models.” News of the funding round was first reported by Forbes; AfterQuery has not yet commented.

💬 JudyAI Lab Take

AfterQuery went from a $300M Series A valuation to a $3.2B unicorn in just five months, with co-founders who are only 22 and 23 — a pace even YC itself says is unprecedented.

This shows how the capital logic around AI training data is shifting. AfterQuery isn’t doing the traditional “is the answer right or wrong” labeling work — instead, it recruits doctors, lawyers, and other professionals to encode their actual problem-solving thought processes and decision logic directly into models, training AI agents to learn “how to do the work” rather than just “how to answer.” This lines up with the trajectory of companies like Mercor and Scale, and it signals that what major AI labs need most right now isn’t more raw data volume — it’s high-quality data that demonstrates genuine expert judgment in action. For AI builders, that’s a signal worth noting: if your product needs a model to handle domain-expert-level tasks, generic pretraining or RLHF alone won’t cut it anymore. “Process demonstration” data may be the key differentiator for the next stage.

If your product involves vertical, domain-specific tasks, it’s worth asking: is your AI mimicking the answers, or mimicking the process an expert goes through to get there?

📅 Source Info

🔗 Further Reading