Physix FrontierShanghai WAIC 2026 Press Conference Highlights Overlooked by Media Focus
The name StepFun is not unfamiliar within the AI circle. Its founder was formerly a tech executive at ByteDance, and core team members come from top companies like Google, Microsoft, and Baidu, with a solid technical foundation. But what truly alerted me wasn't its technical parameters, but its layout strategy in the AIOS (Artificial Intelligence Operating System) field—it seems to be walking a completely different path from Apple, Huawei, and Xiaomi.
Background: Why AIOS Has Become a Battleground
Over the past two years, large models have moved from labs to applications, but most landing scenarios got stuck at the underlying architecture of the "operating system." Traditional mobile OSs (Android, iOS) were designed for touch interaction and inherently do not support the real-time inference needs of AI agents. For AI to truly become an "agent," scheduling rights at the bottom layer need to be open, and memory, compute, and sensor data need to be accessed by AI models within milliseconds. This is a comprehensive reconstruction from the kernel layer to the application layer.
Current mainstream players fall into two categories: One is hardware manufacturers, like Apple with Apple Intelligence, Huawei with HarmonyOS native AI, and Xiaomi with HyperOS + Large Models. They possess hardware entry points and user stickiness but cautiously maintain their ecological walls. The other is internet giants, like Google with Gemini Nano and ByteDance with Doubao. They excel in data and algorithms but lack terminal control. Both paths have pros and cons, but the commonality is—they both try to define AIOS with their own rules.
StepFun's Choice: Don't Build an OS, Be the "Catfish in the OS"
StepFun's "world's first AI agent phone" actually doesn't involve making phones themselves, but cooperating with a second-tier phone brand to embed their self-developed AI agent framework into the system's bottom layer. From technical documentation, their approach is unique: instead of modifying the Android kernel, they decouple AI model inference requests from system services via a lightweight virtual machine layer. This allows users to let AI agents take over native apps like SMS, Calendar, and Camera without changing their operating habits.
It sounds like a "plug-in," but its true value lies in proving that third-party AI platforms can invade traditional OSs to achieve "painless upgrades." For phone manufacturers anxious about the AI era, this is equivalent to a low-cost breakthrough path. If StepFun's framework matures, any small-to-medium manufacturer can leverage it to quickly obtain AI experiences close to top-tier manufacturers without needing to self-develop hundred-billion-parameter large models.
Can It Really Be a Catfish? Depends on Three Key Points
The first is data sovereignty. StepFun's framework needs access to vast amounts of local user data (contacts, location, usage habits) to deliver "agent-level" proactive services. Whether phone manufacturers are willing to hand over data control to a third party is a huge trust barrier. No matter how beautiful the technology, if the negotiation table seats OPPO or vivo, who have already self-developed AI, StepFun might not even get through the door.
The second is ecological stickiness. The winning hand of AIOS has never been single-point capability, but the synergy of the application ecosystem. Apple's Apple Intelligence can integrate into dozens of native apps because it has over a decade of developer ecosystem; Huawei's HarmonyOS AI relies on the hardware matrix of the Internet of Everything. StepFun currently cooperates with only one phone manufacturer, and app adaptation mainly relies on copy-pasting, lacking native innovation supported by developers.
The third is commercial sustainability. Deploying AI agent frameworks is extremely costly. Every inference consumes cloud compute or edge-side chip resources, and phone manufacturers are nearly obsessively sensitive to hardware costs. StepFun's business model currently mainly charges annual licensing fees, but the marginal cost of AI services in the 5G era is rising. If phone sales fall short of expectations, this company might be crushed by the "free upfront funding" it planted.
Trend Prediction: There Won't Be Just One Catfish, But the Real AIOS Is Born in 2028
Despite many uncertainties, StepFun's emergence is still an important signal. It breaks the narrative that only giants can define AIOS, showing small-to-medium manufacturers the possibility of "hitching a ride." If its framework secures orders from three or more top phone brands within half a year, the entire market will be reactivated—manufacturers will be forced to open bottom-layer interfaces, even pressuring Google and Apple to accelerate opening AI capabilities.
But I lean towards a calmer judgment: StepFun is likely just the "first generation catfish." The true chess players of AIOS are not a single tech company, but multi-dimensional complexes that simultaneously possess chip design, system bottom layers, application ecosystems, and data closed loops. The role of this catfish is to stir the oxygen in the pond, forcing all contestants to swim faster. That AIOS which truly changes the game rules will likely appear around the World Artificial Intelligence Conference in 2028, in a form none of us expected.
Original Link: https://www.tmtpost.com/8060634.html
