Whatnot Just Bought An AI Company To Solve A Problem Amazon Never Had To Deal With

If you've never used Whatnot, here's the quick version: it's like QVC but livestreamed, and instead of one channel, there are thousands of sellers running auctions at the same time, all day, every day. Someone's selling Pokémon cards. Someone else is doing vintage sneakers. Another person's mid-auction on a golf club.

Now imagine trying to recommend the "right" show to someone browsing at 9 pm on a Tuesday, when half those auctions will be over in ten minutes, and new ones are popping up constantly.

That's the problem. Whatnot just spent money to fix.

The actual deal

Whatnot announced this week that it's acquiring Shaped, a machine learning company that builds real-time recommendation and search systems. Shaped's whole thing was helping businesses combine customer data with AI models to make search and discovery feel personal they'd worked with companies like Outdoorsy and QVC before this.

Shaped's founder, Tullie Murrell, is joining Whatnot along with about a dozen of his engineers and researchers. Murrell's going to run a new group inside Whatnot called Applied AI Research. He came from Meta before starting Shaped, so this isn't someone's first rodeo with recommendation systems at scale.

Why regular e-commerce recommendations don't work here

Think about how Amazon recommends products to you. Your browsing history, past purchases, and stuff other people bought together, it's a relatively stable catalog. A product that exists today will probably still exist tomorrow. The system has time to learn.

Whatnot doesn't get that luxury. Emmanuel Fuentes, who runs Data and AI there, put it simply: "Inventory changes by the second, shows start and end continuously, and buyer intent shifts throughout a show."

An auction that's live right now might be completely gone in five minutes. A seller could start a new stream at any moment. What someone wants to watch at the beginning of a show is different from what they want twenty minutes in, once they've seen what's selling. It's a moving target in basically every direction at once.

They've actually been chipping away at this for years

This isn't a "we suddenly noticed a problem and bought a company" situation. Fuentes said Whatnot has spent the last six years working on recommendation speed, specifically, they got it down from about a day's delay to just minutes. That's already a pretty wild improvement if you think about it. A full day used to pass before the system adjusted to what you actually wanted.

Bringing in Shaped's tech is meant to shrink that gap even further, closer to actual real-time. And the scale they're working with is genuinely huge. Whatnot says it processes more than 500,000 hours of live video and millions of interactions every single week, all feeding back into the recommendation engine.

This acquisition isn't happening in a vacuum. Whatnot has been growing fast sellers just crossed 1 billion total orders, the company raised $225 million last year at an $11 billion valuation, and they added 20 million buyers in the past year alone.

They've also been throwing open the doors on categories. More than 35 new ones launched last year (art, golf, vinyl records, random stuff like that), and over 45 more in just the first half of this year. Every time you add a category, you're adding a whole new kind of shopper with different habits, which makes the recommendation problem even messier.

And honestly, this is part of a wider trend eBay and Poshmark have both been building out AI features for search and personalization, too. Everyone in resale and secondhand commerce is racing toward the same idea: help people find the thing they actually want in a pile of stuff that's constantly shifting.

Whatnot just decided the fastest way to get better at that was to buy the team that already knew how to do it.