$900B China Market Forces AI Live Shopping's U.S. Test

Live shopping was supposed to turn social video into instant retail, but the U.S. version still looks tiny next to China’s roughly $900 billion market. The new bet is that AI live shopping can fix the real bottleneck: not the stream, not the host, but discovery at the exact second a buyer is ready to act.

The gap is stark. In China, live shopping generated roughly $900 billion in sales in 2025, nearly the size of the entire U.S. eCommerce market, NIQ estimated in a Friday, July 17 press release cited by PYMNTS. In North America, 68% of consumers said they’ve never bought anything through social media.

That contrast makes Whatnot’s acquisition of Shaped more than another AI tuck-in. It’s a test of whether AI can turn a chaotic feed of live auctions, shifting inventory and split-second buying windows into a marketplace that actually knows what each viewer should see next.

AI Live Shopping Is Getting a Discovery Engine, Not Another Video Gimmick

Live shopping is familiar because it borrows from television home shopping: a host goes live, demonstrates a product, answers viewer questions and pushes a limited-time deal. The internet version compresses the funnel. Viewers can buy with one tap without leaving the stream.

The problem, as PYMNTS frames it, is not content. It’s matching. Thousands of streams can run at once. Inventory changes constantly. Auctions close in seconds. A shopper who joins late may miss the product that would have converted.

That is where AI live shopping changes the shape of the model.

• Old constraint: A viewer has to find the right stream manually or rely on broad recommendations.
• New bet: Real-time AI can route that viewer toward a stream where the right product is still available.
• Old failure point: The product disappears before the buyer arrives.
• New pressure point: The platform has to read behavior fast enough to keep up with live inventory.

XOOMAR analysis: this is less about making live shopping flashier and more about making it less wasteful. A live stream only has commercial value if the right buyer arrives while the right item is still in play.

China Has the Sales Scale. North America Has the Adoption Problem

The numbers show two different markets moving at different speeds.

Whatnot is already dealing with the kind of volume that makes manual discovery impractical. PYMNTS says the platform processes more than 500,000 hours of live video and millions of user interactions each week.

The company also expanded fast. It launched more than 35 new product categories in 2025 and more than 45 in the first half of 2026, with sellers surpassing 1 billion orders, according to the report.

That growth creates a harder recommendation problem. A buyer interested in collectibles, sneakers or apparel may behave differently depending on the moment inside a broadcast. Static recommendations can miss that timing.

XOOMAR analysis: China’s scale doesn’t mean the same model automatically works in North America. The source points to a clear adoption gap. If Whatnot can shrink the distance between interest and relevant stream, it gives live commerce a better chance to move beyond niche behavior.

Whatnot Bought Shaped Because Inventory Expires While the Viewer Is Watching

Whatnot acquired Shaped, a machine learning startup that builds real-time recommendation and search infrastructure, on Wednesday, July 15, TechCrunch reported, according to PYMNTS. Financial terms were not disclosed.

Tullie Murrell, Shaped’s founder and CEO, previously worked on recommendation systems at Meta before founding Shaped in 2021. He will join Whatnot with nearly a dozen engineers and researchers to lead a new applied AI research group.

The key point is speed. Live commerce does not behave like a normal catalog. Products appear, sell and vanish during the session.

“By combining Shaped’s technology with Whatnot’s existing systems, we can make recommendations faster, more responsive and more personalized,” said Emmanuel Fuentes, vice president of data and AI at Whatnot, per the report.

Shaped’s systems combine customer data with AI models and machine learning to support personalized search and discovery. PYMNTS says the combination could help Whatnot recommend relevant streams based on previous purchases and real-time behavior signals.

That is the core of AI live shopping: the storefront changes while the buyer is still watching.

From QVC Habit to One-Tap Checkout, the Impulse Is Familiar but the Timing Is Harsher

The behavioral idea is old. A host creates urgency, trust and product context in one package. The internet version adds comments, instant questions and one-tap purchasing.

The harsher part is timing. In traditional eCommerce, a product page waits. In live shopping, the moment can disappear. A viewer may enter after the desired item has sold or before it appears.

That makes recommendation quality more central than the host alone. The platform has to decide not just what a shopper likes, but when that interest is actionable.

PYMNTS also ties the story to broader AI-mediated shopping behavior. Its Intelligence report, “The Agentic Commerce Deep Dive: Payment Infrastructure and the Path to Trust for Merchants,” found that 48% of online shoppers used AI to research their most recent purchase. AI-referred traffic to U.S. retail sites grew 393% year over year in the first quarter of 2026.

XOOMAR analysis: those figures don’t prove live shopping will break through in the U.S. They do show consumers are becoming more comfortable letting AI sit between intent and purchase.

Payments and Trust Move Into the Same Live Moment

If live shopping scales, checkout becomes part of the entertainment layer. The source describes viewers buying with a single tap without leaving the stream. That puts payments, identity, data and trust under more pressure than a slower shopping journey.

This is where the fintech angle matters. Retailers and platforms want conversion. Consumers want relevance without feeling trapped in a sales machine. Payment providers need the transaction to clear cleanly inside a fast-moving session.

XOOMAR has covered the same trust pressure in other digital channels, including the stakes around the 80% digital shift putting the Regions Bank app on the line and CFO concerns that payment blind spots are costing trust. Live shopping compresses that problem into a few seconds.

The PYMNTS source does not address fraud, chargebacks, BNPL, influencer compensation or AI disclosure rules. Those remain open questions. But if one-tap buying inside live video grows, those issues won’t stay peripheral.

The Next Proof Is Whether Shaped Changes Buyer Behavior

The acquisition gives Whatnot a sharper AI story, but the test is operational. Can Shaped help Whatnot route buyers to relevant streams while items are still available? Can recommendations keep up as categories multiply? Can personalization lift live shopping without making the experience feel over-engineered?

The strongest version of this thesis is simple: AI live shopping works if it reduces missed moments. A buyer sees the right stream, at the right time, with the right product still available.

The weaker version is just as clear. If AI only adds more prompts to an already noisy feed, North America’s 68% nonbuyer figure will remain the number that matters.

The Bottom Line

• AI could solve live shopping’s discovery problem by matching viewers with products before auctions or deals disappear.
• Whatnot’s acquisition of Shaped signals that recommendation technology is becoming central to social commerce growth.
• The huge gap between China’s live shopping scale and North America’s low adoption shows how much room the format still has to grow.