Billions of dollars will be spent on company data by the end of this year, and probably tens of billions by the end of next, as labs push their models into autonomous white-collar work.
But a model can’t learn to do an office job from the public internet alone. Why not? Because the record of how that work actually gets done isn’t public!
How an NDA gets negotiated, how a fracking operator schedules its crews, how a memory chip goes from spec to tape-out... all of it sits idle inside companies.
There are workarounds. You can hire domain experts to write tasks and rubrics for “simulated workspaces”, where a model gets rewarded when its work meets the expert’s standard. But that is slow and expensive, and it only captures what the experts think to write down.
Or you can license a company’s data: years, sometimes decades, of real work sitting idle in systems like Slack, Salesforce and AutoCAD, with real-world outcomes to grade against.
The second route is messier, harder and fuzzier. But we think it’s also more scalable, more capital efficient and, when done right, better for hill-climbing.
The problem is that the market is badly inefficient on both the sell side and the buy side.
Nobody has a standard way to value a corpus, and every buyer we work with licenses data differently.
Data is the third leg of the stool, next to compute and algorithms, and its market is the one still running on Slack channels and Wolf of Wall Street-esque tactics.
So today we’re introducing Buyside: a single place where labs and RL environment teams can find, judge, negotiate and license data from real companies.
Detailed
Every dataset tells you what’s actually inside before you ask. You see the mix of emails, documents, tickets, code, CAD and recordings, how much there is, and how many years it covers. You can see what you’d be training on without a discovery call.
Priced
Most listings carry an indicative price, set with the owner, right next to the data. No more time wasted asking “price range?” on every corpus that seems alluring. Most prices aren’t set in stone, and neither are terms, so make an offer: we have a 30-minute response time on pricing offers.
One-minute samples
One click and you have a sample of the real data. No call, no back-and-forth. The first time, you accept our mutual NDA in the same window, and it covers every sample after that.
Synthetic data and expert-written tasks have taken models a long way. The next stretch needs the real thing: the record of how work actually gets done, and what happened as a result.
Buyside is built for that shift. We curate and clean company corpora and turn them into something a lab can actually evaluate, price and buy.
AI’s spread through the economy depends on this data changing hands efficiently.
We’ve been in beta with a small group of partners. Today we open the waitlist, and we’re letting new buyers in every hour.
The hills are real. Now there’s data to climb them.