AI Trends

AI Toys’ Double 11 Balance Sheet: Surging Sales, Falling Average Prices, and Return Rates Above 30%

36Kr · AIUpdated Nov 20 21:45

AI summary

From January to July 2026, AI toy sales on Chinese e-commerce platforms rose 158.7% year over year, while revenue increased 72.7% and ASP fell 33.3%. Return rates for some products lacking playability and usability reached 30%–50%. The article argues that as conversational modules become standardized, product premiums depend increasingly on practical functionality, user relationships, and established R&D and production capacity. The industry has also begun exploring an ongoing service model combining hardware with subscriptions. Although embodied capabilities may improve retention, they will raise software and hardware costs and introduce regulatory requirements concerning fee transparency and the protection of minors.

Why read this

AI score 71

A comparison of sales, average prices, and return rates reveals that competition among AI toys is shifting from conversational features toward use cases, retention, and ongoing services.

Topics

Tags

Body · AI translation

From January to July 2026, sales volume of AI toys on Chinese e-commerce platforms rose 158.7% year on year, while revenue increased by only 72.7% and the average selling price fell 33.3%. More were sold, and they became cheaper. This does not mean the sector is deteriorating; rather, the ability to “talk” is losing value: core modules can be brought down to 50 yuan, conversational modules cost just 30 to 50 yuan, and national subsidies combined with consumer vouchers mean anyone can make a plush toy speak. Once conversation becomes a standard feature in the supply chain, premiums are no longer driven by technology, but by use cases, relationships, and retention. The coming of age of AI toys is not learning to speak, but beginning to pay the price for being kept around.

This year’s Double 11 shopping carts had a lively new presence: AI. JD.com sold more than 50 million AI products, AI-related categories on Taobao’s mobile app received over 4 billion views, and even computing-power tokens were put on store shelves at a discount. AI toys were one of the most tangible categories: on the first day of sales on Tmall, sales of AI companion robots increased more than sixfold year on year; as of 6 p.m. on October 30, JD.com’s AI toy transaction value was up 150% year on year.

But the excitement is for spectators; the books are for industry participants.

The year 2024 was dubbed the first year of AI toy commercialization. Two years later, the industry appears to have completed a full cycle. At CES 2026 early this year, more than 30 AI companion companies exhibited, including at least 20 Chinese manufacturers. Many were startups attending for the first time with products that had yet to launch. On the capital side, a total of 96 investment institutions entered the field in 2024 and 2025, while more than 50 related funding rounds worth over 20 billion yuan have taken place since the beginning of 2025. No one wants to miss the chance to bet on the “AI version of Pop Mart.”

▲ Image/Robot Technology and Applications

Yet the Double 11 results were far from evenly distributed: only a handful of companies truly broke through, while persistently high return rates told the other side of the story. A counterintuitive conclusion must be put on the table first—the least valuable thing in this boom is precisely the two letters “AI.”

The arithmetic of volume and price: The more that is sold, the lower the unit price

First, consider three figures together.

Sandalwood monitoring data show that from January to July 2026, sales volume of AI toys on Chinese e-commerce platforms grew 158.7% year on year, revenue increased 72.7%, and ASP fell 33.3%. Products priced below 500 yuan accounted for most sales, the 100–200 yuan bracket formed the core market, and the 300–500 yuan segment grew fastest. Sales volume expanded at more than twice the rate of revenue; the gap between the two is the price that was squeezed out.

Why did prices collapse? There are three layers of reasons, each more concrete than the last.

The first is lower hardware costs. Industry gross margins are around 70%–80%, exceeding 90% in some cases. The cost of a core module can be reduced to less than 50 yuan, while a display costs about 80 yuan. A conversational module can be procured for 30 to 50 yuan, installed in a figurine costing 100 to 200 yuan, and sold for 500 yuan. One leading manufacturer told the media frankly that the ex-factory price was about 100 yuan. Last year, his was the only company to have pivoted from plush toys to AI toys; this year, more than a dozen companies have done so. Once “being able to converse” becomes a standardized module that can enter mass production within weeks in the Pearl River Delta, it is no longer a selling point but a factory-default feature.

The second is the stacking of subsidies. A pair of AI glasses listed at more than 2,700 yuan costs less than 2,400 yuan after consumer vouchers and national subsidies are applied, and less than 2,000 yuan after an additional trade-in subsidy. The combination of platform subsidies, national subsidies, and interest-free installment plans has directly lowered the threshold for “trying something new” by another notch. Part of the decline in average prices is therefore being paid on users’ behalf by government policy and platforms; it is not entirely the result of a homogenized price war. These two accounts must be calculated separately, or the intensity of competition will be misjudged.

The third is an explosion in supply. More than 30 companion-product companies at CES, triple-digit supply growth in new platform categories, and 96 institutions entering the field over two years have all crowded into the same window of opportunity. The line in Pinduoduo’s merchant recruitment notice—“AI capabilities can support a 300% premium, allowing merchants to bid farewell to 9.9-yuan products with free shipping”—sounds compelling, but it is a recruitment pitch, not a transaction result.

Then consider who is being filtered out. For products that merely stuff conversational modules into plush shells and lack playability and usability, industry reports put return rates at 30%–50%. By contrast, Mengyou Intelligence’s Ropet desktop pet has kept its global return rate below 4%. Both are AI toys, but one is returned while the other is kept—showing that high return rates are not the inevitable fate of the category, but the result of differences between products.

History has already offered an answer once before. In 2014, children’s robots powered by NLP technology shipped millions of units per month. In 2017, Alibaba, Xiaomi, and Baidu swept across the market with low-cost smart speakers, nearly wiping out children’s robots. What killed them in the previous cycle was not technological regression, but that “being able to talk” was too easy to replicate and too easy to replace with a cheaper device. The concern in this cycle rhymes with the last.

So the real significance of platforms collectively “legitimizing” AI toys during Double 11 is not that they issued an admission ticket, but that they changed the yardstick. Transaction value in Taobao and Tmall’s AI toy category exceeded 100 million yuan, up 300% from the previous period, with branded products already accounting for more than 80%; the category has branched into such segments as emotional companionship, children’s learning companions, and outdoor educational activities. As of 6 p.m. on October 30, JD.com’s AI toy transaction value had grown 150%. Pinduoduo created a dedicated subcategory and opened it to merchant applications.

▲ Image/Ku Wan Chao

Platforms reserve space for products that can be used repeatedly, while users reserve their budgets for products worth spending time with over the long term. Retention, repeat purchases, payment for services, and profitability have become the sole criteria for the next round.

The product divide: useful, engaging, and backed by a solid foundation

Every player that has truly gained a foothold has an anchor. The contest is not about who has connected to which model, but about what they have accumulated beyond the model. Broken down, there are three paths.

The first path is to sell “utility.” Mingxiaotang, which built its business on early-learning and storytelling devices, has put AI into learning tools: managing Pomodoro timers and check-in plans, adding pronunciation correction, location tracking, and calling. Its reason for existing is rooted squarely in the two scenarios of doing homework and going to school. Whale Robot, which began making educational robots in 2002, is using AI to remake the product lines and distribution channels for ages 3–22 that it has built up over more than a decade. Dr. Luka’s AI camera-based learning device lets children photograph a leaf outdoors, while FoloToy AI Passport puts language learning into a passport-shaped device. All follow the same logic—AI is an addition, not a subtraction.

This path is the easiest to underestimate. According to Mojing Market Intelligence data for January–May 2026, “functionality” was consumers’ top consideration when making a purchase, accounting for 31.7%. Programming education, early learning, and language learning together accounted for more than 60%, while emotional companionship represented only 19.8%. Public discourse is talking about emotional value, but at the checkout counter, consumers are paying for essential educational needs. Buzz and actual spending have never been the same thing.

The second path is to sell “staying power.” Fuzozo takes the approach of nurturing a relationship from scratch. Its proprietary MEM multimodal emotion model, EchoChain biomimetic memory system, and GrowMe personal development system enable it to remember where the last conversation ended, keep a diary, and develop a personality increasingly like yours as interactions accumulate. Women make up 80% of its users, a sizable share of whom buy it for themselves, while active users spend nearly 50 minutes with it per day on average. Talking Tom Cat takes the opposite route: the “Talking Tom Cat” IP acquired at great expense in its early years gets users to recognize the character first and its functions second. Its AI children’s companion, portable AI companion, and AI cuddle companion are now divided among three scenarios—home, outdoors, and bedroom. Powered by Qwen, the products incorporate emotion recognition, preference memory, location tracking, and one-touch emergency calling.

▲ Image/Fuzozo’s official Weibo account

One builds a moat with “a relationship nurtured over time,” while the other defends its position with “a recognizable character.” The paths are opposite, but the destination is the same: turning a one-off piece of hardware into a relationship users will want to reopen tomorrow.

The third path is to sell “an established foundation.” UBTECH’s main business is not toys, yet it has packed its proprietary agents’ personality settings, virtual travel, outfit changes, CP intercom, and always-online 4G connectivity across all scenarios into the palm-sized Mengzai and Youzai. This is the transfer of capabilities downmarket. Shifeng Culture has moved in the opposite direction. Drawing on its flexible production lines in Shantou and AI R&D center in Hefei, it has shortened the development cycle from the industry average of six months to as little as two months, showcasing full-chain ODM capabilities spanning product definition, ID design, hardware development, AI interaction, mass production, and global delivery. This is the upgrading of production capacity.

Even more memorable is Shifeng Culture Vice President Shi Jie’s assessment: the industry’s fundamental contradiction lies in a mismatch between supply and demand and pseudo-innovation that merely changes the shell. There are two typical mistakes—using AI logic to pile specifications into toys while ignoring genuine needs, and applying the traditional toy industry’s one-off sales model to AI products, making it “difficult to bear the ongoing cost of subsequent cloud services.” That sentence exposed the economics discussed in the next section earlier and more incisively than any industry report.

▲ Image/UBTECH Mengzai’s official website

Three paths, one conclusion: functionality answers “why buy,” relationships answer “why not return it,” and an established foundation answers “why us.” Anyone can make something talk; none of these three can be obtained simply by purchasing modules.

How the tech giants are positioning themselves: not making toys, just guarding the gateway

Looking behind the scenes at the companies supporting these teams reveals a conclusion that is often stated backward: the tech giants are not “only lending their brains without entering the field.” They simply have no intention of turning themselves into toy companies.

JD.com has gone the deepest. It established the Chameleon Business Unit in January 2026 to commercialize the JoyAI App, JoyInside, and digital humans. JoyInside has been integrated into products from more than 40 leading hardware brands and has reached partnerships with over 200 brands, driving a 120% increase in the number of end-device interactions. By November 11, cumulative interactions had exceeded 70 million. JD’s private-label brand J.Zao developed its own AI plush toy as an initial prototype, and it sold out immediately upon launch. Providing the intelligence, supplying the products, and developing its own prototype—all three are happening at once. What this provides is not the luck of betting on one particular blockbuster, but the confidence of not having to pick any single winner.

Alibaba is not after shelf space. At the Qwen Smart Hardware Exhibition early in the year, Alibaba Cloud released a multimodal interaction development kit integrating Qwen, Wanxiang, and Bailing, while Shifeng Culture, UBTECH, Talking Tom Cat, and others presented the results of their collaborations. The QwenNote Eva desktop robot shown at the Apsara Conference uses DingTalk as its collaboration platform and can assign tasks, listen and take notes, issue reminders, and change its expressions according to context—it is designed for an office desk, not a toy cabinet. Xu Dong, general manager of Alibaba Cloud’s Qwen large-model business, laid out the calculation plainly: “Not many products in a single category can sell more than one million units, and product formats are highly fragmented, so we work very closely with solution providers and module and chip companies.” Fragmented categories are not worth building complete devices for; what is worthwhile is giving Qwen another gateway into the physical world that does not depend on a phone.

▲ Image/QwenNote Eva’s official WeChat account

ByteDance is the most straightforward. The “attention-grabbing bag” that went viral in 2024 was merely an internal Mid-Autumn Festival gift and was not for sale. Volcano Engine directly stated that it would not produce AI toy products in the future, but would work with partners to build an ecosystem around AIoT terminals. It does not make money from selling toys; it profits from the volume, data, and model calls generated as every toy equipped with Doubao extends its reach. According to the reference article from The Paper that you provided, shipments of AIoT products integrated with Doubao had exceeded one million units by June 2025 and were expected to surpass ten million by year-end. During this year’s Double 11 shopping festival, Token transaction volume on Tmall rose 728% month on month, while sales of large-model Token packages on JD.com increased more than tenfold month on month. The contours of the model-call business are already clear.

The three companies may have different postures, but their calculations are the same: making toys is a one-off transaction, while building a gateway means owning a doorway people pass through every day. The tech giants have covered the most expensive part of the model costs for smaller manufacturers while also keeping control of the ceiling; the smaller companies earn hardware margins, while the giants collect model-call fees and gateway tolls.

The Embodied-AI Ledger: Once Toys Can Walk, Prices Will Rise Again

Industry reports generally point in three directions: interaction is shifting from passive question-and-answer exchanges to proactive perception and embodied action; the audience is expanding from children to young people living alone and older adults; and the business model is moving from hardware margins to “low-cost hardware plus content/service subscriptions.” These three statements all raise the same question: Where is the boundary of a toy, and how much does it cost to cross it?

A signal has already arrived from outside the arena. At Nvidia’s GTC in Berlin on October 21, Jensen Huang brought four Microducks onto the stage. They could walk, sit, perform somersaults, fall over, and get back up—movements learned by falling tens of thousands of times inside a virtual physics engine. But one detail mattered more than the spectacle: priced at US$399, or about RMB 2,700, the product is positioned as an open-source experimental platform, not a consumer electronics product.

Place that price alongside the sub-RMB 500 market base described in the first section, and the most sobering sentence in this article emerges: embodiment is a solution to retention, but it is also the enemy of affordability. CLS has already noted that most products priced below RMB 500 lack cameras and multimodal capabilities and have been criticized as “large models in a new shell.” Products with genuine “embodied AI” architectures are constrained by software and hardware costs and generally cost more than RMB 1,000, while their market acceptance remains to be seen. On one side, ASP has fallen 33.3% year on year; on the other, embodiment is pulling prices toward the thousand-yuan mark. The two curves are moving in opposite directions. Whoever first delivers embodied AI at a price in the hundreds of yuan without compromising reliability will truly secure a ticket to the next stage.

Subscriptions did not emerge from thin air as a business model; they are the answer forced by costs. One industry practitioner cited a US children’s companion-robot company that sold only hardware while providing continuous conversations, only to be dragged under by its ever-growing Token consumption. Hence BubblePal costs RMB 499 with a RMB 99 annual fee, UBTECH’s Meng UU charges RMB 99 for an annual pass, and Fu Zai provides 24 minutes of free chat time per day, with annual memberships ranging from RMB 99 to RMB 299. Yet consumers still believe that “a toy should be a one-time purchase.” Their lack of clear expectations about subsequent costs is itself a form of information asymmetry; proposals have already called for sales pages to be required to state “the annual cost range required for the full experience.” Hardware is the admission fee, while services are a long-term lease—for this business to work, the lease must first be shown clearly to customers.

The toughest part of this industry’s “coming-of-age ceremony,” however, lies not in technology but in regulation. The Interim Measures for the Administration of Anthropomorphic Artificial Intelligence Interaction Services, which take effect on July 15, 2026, explicitly require providers to establish a minors mode, usage-time limits, anti-addiction reminders, and mechanisms for informing parents, while strictly prohibiting AI from replacing or distorting real social relationships. A category truly comes of age only when it is required to take responsibility for its “boundaries.”

There is therefore no need to rush into romanticizing embodied AI. The leap from “able to talk” to “able to walk” is not merely technological; it also crosses cost sheets, charging models, and regulatory red lines. Once the boundaries of toys are broken, they may indeed become digital pets that know you in the physical world and grow up alongside you. But what will sustain that vision is not the word “family”; it is the retention curve and subscription-renewal rate.

Being able to talk is the admission ticket; staying is what earns a seat.

This year, more AI toys were sold, at lower prices, and they were also returned more quickly. Sales volumes rose, unit prices fell, platforms changed their category classifications, and users changed the yardsticks by which they judged them. Coming of age has never meant merely learning to speak; it means beginning to pay for three things: retention, continuous model calls, and boundaries and rules.

The next Pop Mart sought by investors will not be born on a shelf of products that “can hold a conversation.” Traditional art toys sell the “usefulness of the useless”—they do not need to answer back to be cherished, and their value peaks the instant a blind box is opened. The hurdle AI toys must clear is precisely the opposite: getting someone to want to turn one on again tomorrow.

When the tide recedes, the one left standing will not be the best talker, but the one that is hardest to replace.

Sandalwood: Zitan Data e-commerce monitoring, January–July 2026

CLS: China’s AI toy sales grew 158.7% in 2026

Elephant IPO: Elephant Research Institute releases the 2026 AI Toy Industry Research Report

Hebei News: Explosive growth! Views of AI-related categories on Mobile Taobao exceed 4 billion

Robot Technique and Application: CES 2026 opens, bringing together 18 Chinese companion-robot companies

WhalesBot: The “Full Ecosystem, New Journey” 2025 WhalesBot product launch was successfully held

IP Cultural and Creative Trends: An AI toy boom? More than 30 Chinese AI toys and companion robots appear at CES 2026

Tech Planet: Selling out on e-commerce platforms at RMB 99 with a 50% return rate—how profitable is the AI toy business attracting tech giants?

Operations Research Society: Using AI to “replicate” Labubu and raking in RMB 120 million in the first year—the new RMB 10 billion trend has exploded

Sina News: In the first half of JD.com’s 11.11 shopping festival, 150 million service orders were placed and more than 20 million AI products entered households across China

This article is from the WeChat public account “New Entropy” (ID: baoliaohui), written by Sisi and published by 36Kr with authorization.

Source

36Kr · AI