China’s next AI winners may be hiding inside your devices

The most consequential AI company in China may not own the app consumers open every morning. It may be a company whose name never appears on the screen at all. Picture a car that can handle a request even when mobile coverage drops in a tunnel; a phone that understands a document without sending every interaction to a remote server; or a wearable, robot, or laptop that responds quickly because the intelligence is already inside the device. The products may carry the badge of Geely, Mazda, Samsung, or Xiaomi. Underneath, a smaller-model company may have supplied the model, runtime, chip adaptation, tools, and engineering that made it usable. China’s next AI business model may look less like a chatbot subscription and more like a Tier 1 automotive supplier: invisible, embedded, difficult to replace, and paid for through the hardware supply chain. The smallest model in the room may have the biggest commercial job For the past three years, artificial intelligence has been sold largely through spectacle: more parameters, more benchmark victories, more questions answered in a browser window. That is a useful way to measure a frontier model. It is not a useful way to put intelligence inside a car, phone, or wearable device. Hardware imposes unromantic constraints. A model needs to work with the chip that has already been selected. It has to fit within memory, power, and thermal limits. It needs to respond quickly enough that a driver will trust a voice interaction and reliably enough that a manufacturer will support it through a product cycle measured in years. That is why “smaller” is becoming a commercial advantage, not a consolation prize. ModelBest, the Beijing company behind the MiniCPM family, has built its identity around edge or device-side models. Its official site highlights applications across smartphones, AI PCs, intelligent vehicle cabins, robots, and wearables, emphasizing efficient local processing, lower latency, and offline functionality in supported applications. As AI leaves the browser and enters physical products, the winning model may be the one that does enough, reliably, on hardware people already own. From a model download to a production line ModelBest’s latest numbers help explain why that distinction matters. Downloads across the company’s MiniCPM family have surpassed 50 million as of September 2026, giving the Beijing startup one of China’s most visible open-source edge-AI ecosystems. More important commercially, its SuperMate on-device cockpit solution is moving from demonstration to deployment, with reported installations in mass-produced models like Changan Mazda’s EZ-6 and Geely’s Galaxy series. By mid-2026, ModelBest was said to be on track toward an annual target of more than 300,000 vehicle deployments, as partnerships expanded across Changan, Geely, and SAIC Volkswagen. The capital market is paying attention too: with a reported valuation above RMB 20 billion ($2.9 billion) and A-share IPO tutoring now underway, ModelBest is positioning itself as a candidate to become China’s first listed on-device large-model company. Those figures are not simply a growth story. A download and a vehicle deployment are different commercial events. A developer can download a model in an afternoon. A vehicle program is a long negotiation with hardware, software, safety, and manufacturing. The model must be adapted to a specific system-on-chip, tested against power and memory budgets, integrated, validated, and maintained after the car goes on sale. ModelBest’s own announcement of its upgraded SuperMate solution highlighted exactly those requirements: all-scenario agent services, multimodal interaction, and adaptation across multiple chip platforms. This is a different kind of AI work. The public model may still be open. The enduring value lies in the unglamorous layer around it: deployment, integration, tools, testing, and lifecycle support. Welcome to the AI Tier 1 Automotive industry language offers a useful analogy. A Tier 1 supplier does not merely sell a component; it delivers a system that meets the manufacturer’s specifications, works with other systems, survives validation, and can be serviced over time. It earns a place in the supply chain because replacing it is harder than downloading an alternative. That is the role ModelBest appears to be building. On-device AI suppliers increasingly resemble Tier 1 software providers in the automotive industry, managing chip adaptation, toolchains, vehicle-level validation, and ongoing maintenance. The model weights may evolve quickly. The integration does not disappear whenever a new benchmark is published. A cloud-model company can measure revenue in tokens or subscriptions. A device-side supplier may instead rely on design wins, per-device or per-vehicle arrangements, engineering services, upgrades, and ongoing support — often negotiated with an automaker, chip company, or operating-system partner rather than a consumer. It also changes who controls the customer relationship. The end user belongs first to the carmaker or phone maker. The device manufacturer controls the screen, the payment relationship, and the broader product promise. The model supplier must prove its value inside someone else’s product, often without public credit. That is a tougher position than running a popular consumer app. It can also be more defensible. China has an unusually good laboratory for this model China’s advantages are not limited to model talent. It has dense hardware supply chains, ambitious electric-vehicle makers, handset brands willing to differentiate through AI, a large domestic chip ecosystem, and a consumer base that expects software features to arrive quickly. The device-side race is already moving beyond vehicles. By mid-July 2026, seven on-device AI services for mobile phones, including services from Apple and Xiaomi, had completed regulatory filing. The exact commercial winners remain unclear, but the direction is unmistakable: AI capability is migrating from a destination in the cloud to a feature of the products consumers carry and drive. ModelBest markets MiniCPM for phones, PCs, intelligent cabins, robots, and wearables. The technical requirements overlap, but each device class has different procurement cycles, safety standards, user expectations, and value pools. China’s automakers are willing to ship new interaction models, phone makers need locally relevant AI features, and hardware companies have incentives to reduce cloud-inference costs. For an AI supplier, that is a rare industrial proving ground. Who captures the value? ModelBest’s opportunity is genuine. So is the uncertainty. A device-side model moves some inference work onto hardware that has already been purchased by the consumer. That can reduce the ongoing cloud-computing expense associated with every interaction. It can also offer lower latency and more resilient performance when connectivity is poor. Those are valuable features. But value creation does not guarantee value capture. The carmaker owns the vehicle. The phone maker owns the operating system and the consumer relationship. The chip company can influence what is technically possible. An open-source rival may offer comparable weights. If the model becomes a commodity, the supplier may be pressured into selling engineering hours rather than owning a recurring software position. While ModelBest’s public revenue, pricing method, and recurring-service share have not yet been disclosed, its future economics will depend on whether it can retain per-device value after the model is embedded and whether customers continue using its solution in the next generation of products. That is why the Tier 1 comparison is helpful. The best suppliers do not win because their components are impossible to copy. They win because the cost, risk, and inconvenience of replacing the integrated system are high. In AI, that stickiness will depend on reliability, chip compatibility, developer tooling, update discipline, and the trust earned through mass production. Lessons for leaders managing embedded intelligence Stop treating model size as a proxy for business value. The relevant question is whether the model can perform reliably within the constraints of the device and use case. Look below the app layer. Much of AI’s future value may accrue to companies that integrate intelligence into hardware rather than those that own a visible chatbot. Plan for lifecycle support. In cars and durable devices, a model is a long-term component; validation, updates, and maintenance matter as much as launch-day performance. Ask who owns the customer. Model suppliers, chip companies, operating-system providers, and original equipment manufacturers (OEMs) can all claim part of the same value chain. Partnership terms determine who actually captures it. Build switching costs through execution, not obscurity. The durable moat is excellent integration, tools, and service — not simply keeping weights unavailable. The bottom line China’s AI story is often told as a race to build the smartest model. The more consequential race may be to make intelligence shippable: small enough for a device, efficient enough for a chip, reliable enough for a vehicle, and integrated enough that an OEM will keep it for years. ModelBest is an early test of that thesis. If it succeeds, the next generation of AI winners may not be the companies consumers talk to. They may be the companies quietly sitting inside the products consumers trust. Ashley Dudarenok is a Chinese entrepreneur, author, and media personality, recognized by Thinkers50 as a “guru on fast‑evolving trends in China.” She is the founder of ChoZan, a China‑focused research and digital transformation consultancy.

Sectors

Insights

Company

China’s next AI winners may be hiding inside your devices