how to (actually) broker ai data

somewhere around last week my timeline began to flood with ai brokers and look, it's true, the money in this is absurd. but I've spent an unhealthy number of my waking hours doing it, and there are a few misconceptions that the timeline is propogating.

we've now hit a $12M MRR run rate ($144M ARR). there will people commenting that run rate is a deceiving metric and I agree! but the number is going up in a way that is frankly a bit ridiculous.

anyway I figured I'd just write down the whole thing, start to finish (for my future grandkids) on how it actually works, the grubby bits included, so here goes...

first, what the labs are actually hungry for

the labs have all hoovered up the same internet, it's the same crawl, the same github, the same reddit, so what they're scrabbling around for now is the stuff that is behind a locked door, like a chip team's real design and verification work, or wet lab and clinical data, or a factory where people wear head cams while they do the actual job, or the frankly bizarre internal workflows of a roofing contractor or a tax firm or a freight company, which sounds really weird until you realise things that are intuitive to us, isn't for the models.

and the appetite is shifting quite fast, I'd say on basically every call this month lab's asked about finance, legal, semiconductors, biodata, manufacturing, aerospace or some weirdly specific workflow data, or RL environments with proper graders, and almost nobody's SPECIFICALLY asked for generic marketing agency text or a big static dump of non-esoteric code, so if your pitch is "we've got ten years of emails, what's that worth" I'm sorry to be the one to tell you but the answer is less than you'd like. with that said, companies are still paying BIG BUCKS for businesses internal slack, salesforce, quickbooks, etc...

(psss fill out grepped.ai/form and get a quote in 24 hours)

getting the data in the door (the form is doing more work than you'd think)

every company that comes to us fills in an inventory before we touch anything, and it asks really dull questions, what the data is, roughly how much of it there is, what format, where it came from (and if any of it's scraped or someone else's, exactly where), who owns it and that one boring form throws out about half the pipeline on its own because the answers people give are, let's say, uhhhh, creative.

just from the last few days we've had a 40GB zip labelled “HIGH QUALITY CONVERSATIONS” that turns out to be one man’s entire WhatsApp history, including the videos & voice notes, dataset of “real human emotion” which is 13 hours of a smart-doorbell feed pointed at his neighbour’s front path, somebody earnestly offering us voter rolls with full names and addresses and a good dozen submissions which I cannot even mention.

after that we are onto...

checking what's actually in the folder (the frozen 2 funny bit)

this is my favourite story in the industry and I'm not going to say who it was, but there's a broker who took in a dataset, priced it more or less on how many terabytes it was, sent it along to a lab, and somewhere deep down in the folder structure, nested in like nine layers of sensible looking directory names, were roughly four thousand copies of Frozen 2, which the seller had stuffed in there purely to make the folder look chunky, and the broker never opened it but the lab obviously did, and I'd be amazed if that broker ever gets a reply from that lab again.

so before anything moves we're asking whether it actually matches what they said on the for. on a serious note, we take data QC extremely seriously. which labs clock straight away and is, I reckon, a big part of why they keep coming back to us!

de-identification, the bit where amateurs get people hurt

we've got our own (extremely intelligent) team in-house doing it and we work it through third-party de-identification people as well, so it's two lots of eyes, atleast.

and some stuff you just have to walk away from. anything where the whole value of it is who the people are, because once you've scrubbed that out there's nothing left worth buying, and on the other side of it, if a seller tells us they don't want something shared, it doesn't get "mostly redacted", it just never leaves them.

packaging it up so a (tired) researcher can say yes

the data buyer reading about a corpus has a hundred more to read in their inbox, so we don't make it hard for them. labs tell us which industries they're hungry for, and when something lands in one of those sectors we route it to them first. sometimes that means a first look, sometimes a deal where they're the only eyes on that sector if they want it. everyone else can get our catalogue of corpora. we're flexible on exclusive versus non-exclusive, and we work hand in hand with each lab until they know everything they need about what they're buying, because that's the part we care about most.

whatever the route, every package covers the same ground: what it is (in detail), what it's good for, how big it is and whether that number's been checked or just claimed, the format and how it gets delivered, where it came from and who holds the rights, where it stands on personal data, and the known limitations.

to give you an idea, one of the things hot rn now are RL environments for physical chip design, graded by actual signoff tools, there are plenty more "hot" things which you can learn about by joining our whatsapp group (https://chat.whatsapp.com/Ixey4ObAtjc8hAz8BXeikr) where we tell our referrers exactly what the labs are paying serious money for, usually before anyone else knows it’s in demand.

routing, which is the bit you genuinely can't copy

through peculiar forces of fate, I know alot of ai people. we work with the big labs, with the neolabs, and with a load of smaller specialist labs you've honestly never heard of, and a couple you'll never hear of at all, because they work in really private, sensitive sectors and alongside government.

routing is just knowing which is which before you pitch, plus who actually signs, what their legal people will moan about and roughly what they paid for the last thing like it.

btw if anyone at the labs are reading this, we have a previous sales database that alot of the other labs found helpful :)

what a deal looks like

the average one lands somewhere around $200K to $300K, exclusivity is usually two years (you can go shorter but it'll cost you on price), the licence can be one-off or ongoing, the lab sets the scope of use and it's written into the agreement, and timing-wise it's roughly a week for de-identification after the data's extracted, about five days for acceptance, about a week for the lab to pay, so call it a fortnight from upload to money on average, give or take

the money, and where I get a bit frustrated

if you introduce a company to us and it closes, you get 10% of the deal, capped at $100K a deal, so on an average one that's twenty or thirty grand for what might have been a single intro, renewals pay you again, you get paid within 30 days of the lab paying us (per instalment if it's paid in instalments), your credit on that company lasts twelve months and covers every deal they do with us in that window, you can cold email and run ads and use our name, and US and non-US companies of any size all count.

and I'm a bit loud about this because one of the biggest players out there pays referrers a 7.5% fallback, keeps the right to pay you whenever it feels like, and if it decides to terminate you, your unpaid fees just evaporate, most of the data marketplaces don't do referrals at all, and the ones brave enough to publish what they take off sellers are sitting at 15 to 30 percent, whereas ours is written down, every scout gets the terms on paper, and there's no "up to" and no "at our discretion" anywhere in it.

so should you just go and do it yourself

honestly, maybe, at some point, or you can broker & learn through someone who handles the tough parts. up to you!

anyways, if you've got a company with data sat around, or you know someone who does, chuck it in here and we'll have a look:

grepped.ai/form

**lock in for 3 months and your future self will be forever grateful :)