Over the past 18 months we visited 30+ toy manufacturers, solution providers and chip vendors. What follows is the real state of AI module adoption in children's toys in the first half of 2026. It is neither the "year one of the AI toy" that the media talks about, nor the "much noise, little rain" the pessimists describe. It is a market that is splitting apart, but along lines that are now clear.
1. The real division of labor among 3 types of player
Chip vendors: led by Espressif (ESP32 family), Beken (BK family), Rockchip and Allwinner. Their role is to sell chips and SDKs; they do not deal with toy brand owners directly. In 2026, AI-toy-related shipments from these chip vendors range from the millions to the tens of millions, but most of these chips were never designed for toys. They are general-purpose IoT chips that happened to be picked by toy makers.
Solution providers: led by Tuya Smart and Nablai. They add a layer of integration between the chip vendor and the brand owner, packaging the chip, PCB, firmware, voice and vision algorithms, cloud account system and app SDK into a module solution you can just pick up and use. Nablai's LX and TY families, aimed respectively at deep customization under the brand's own sovereignty and at going global fast, are typical examples of this kind of provider.
Brand owners: the toy manufacturers. They own the consumer-facing SKU design, marketing, sales and after-sales. To them, an AI module is one line on the BOM, and they do not need to build an AI team of their own.
Over the past year the boundaries between the three have grown clearer. Chip vendors compete on compute and power consumption; solution providers compete on integration depth and the account system; brand owners compete on IP and marketing. Each sticks to its own job and rarely crosses over.
2. The real differences between 2 technical paths
Path A: embedded AI module
A lightweight on-board model, on-device inference and cloud coordination. Representative products: the Nablai LX Series, and the ESP32-S3 paired with an on-device voice wake solution.
- Strengths: works on a weak network, fast response (on-device wake under 300 ms), data stays local
- Weaknesses: model capability is bounded by on-board compute, and complex conversation still has to go back to the cloud
Path B: general-purpose cloud API
The device only picks up audio and uploads it; all of the conversation is handed to a cloud LLM API.
- Strengths: strong models (GPT-4o, Claude and others can be plugged straight in) and fast iteration
- Weaknesses: always needs a connection, 1-3 second round-trip latency to the cloud, per-call billing that makes costs hard to predict, and high compliance risk around children's data
3. One commercial consensus
Whichever technical path they take, every brand owner seriously building AI toys in 2026 has reached the same consensus:hardware + subscription is the standard commercial model.
Hardware sells at a loss or at cost, and the profit is recovered through subscription (¥60-150 a year). What makes subscription work is not subscription features but subscription content: a story library, knowledge base and character settings that keep updating, so users have a reason to keep paying.
Vendors still building AI toys on a one-off hardware margin were mostly losing money or scraping by in 2026. Only those who close the subscription loop generate positive cash flow.
4. Where there is still no consensus
- Data ownership: whether user conversation data belongs to the brand owner or the solution provider is something the industry has not answered uniformly
- OTA frequency: how often firmware should be updated varies widely from one solution provider to another
- AI content moderation: approaches to sensitive topics (self-harm, violence and so on) are wildly uneven
5. Three observations for toy manufacturers
- Do not pay for AI; pay for user retentionevery AI module is at bottom a cost-reduction tool; whether users stay depends on your IP, your content and your subscription design
- Judge a solution provider on three thingswho owns the account system, the OTA channel, and the content update frequency
- Sell hardware overseas, sell subscriptions at homeoverseas users are more willing to pay for subscriptions, while the domestic market still runs mainly on one-off hardware sales
As the solution provider behind the LX and TY module families, what we see is this: 2026 is not the year one of the AI toy, it is the year the AI toy market splits. Brand owners who close the subscription loop will come out the other side of this round; those still watching from the sidelines will most likely miss it.