China's AI Phone Race Splits Three Ways: ZTE, StepFun, and Honor Diverge in Strategy
0xBroomberg
At the WAIC conference, three Chinese phone makers unveiled three distinct agentic-AI strategies — app layer, OS layer, and perception layer — signaling the industry has no consensus yet on what the next-generation smartphone looks like.
What layer is ZTE betting on?
ZTE's Nubia brand launched the NaviX Ultra, integrating ByteDance's Doubao assistant with voice wake and a dedicated AI button.
The pitch: multi-step cross-app tasks — price comparison, ordering, trip planning — handled by a single voice command, with core inference running on-device.
An earlier co-branded prototype, the M153, sold out all 30,000 units at ¥3,499; resale prices briefly doubled.
This means → ZTE is betting on "app-layer coordination": leave the OS untouched, let AI stitch existing apps together.
Why is StepFun rewriting the operating system?
StepFun built Step AOS, a purpose-built agentic OS that breaks phone functions into service endpoints AI can call directly.
In plain terms = other makers let AI "run errands" on top of the existing OS; StepFun built AI its own house.
The companion agent is called Amoo; payments still require manual user confirmation, and WeChat has not yet joined the launch ecosystem.
This reflects the OS-level approach: the biggest ambition, but the slowest path to adoption — ecosystem partners haven't caught up.
What does Honor's "robot phone" actually add?
Honor is focused on physical-world perception: a retractable four-axis titanium gimbal with a camera capable of 360-degree tracking.
It combines voice, gesture, and motion inputs; the AI model is co-developed with Alibaba.
In plain terms = ZTE teaches the phone to "understand what you say"; Honor teaches it to "see what you're doing."
Whose business model is under threat?
ZTE and ByteDance use MCP — model context protocol, an architecture that lets AI talk directly to app back-ends, bypassing the graphical interface.
This means → users no longer need to open apps one by one, so ad impressions drop, and internet platforms' traffic models take a hit.
Revenue-sharing terms between phone makers, model developers, and internet platforms must be renegotiated — a commercial friction all three strategies trigger.
Can phone chips handle the load?
Qualcomm China VP Xu Hao said at WAIC that moving from single-turn dialogue to autonomous agents raises token-processing demand by roughly an order of magnitude — potentially hundreds of thousands to millions of tokens.
Qualcomm's response splits three ways: CPU for task planning, low-power sensors for ambient processing, and NPU acceleration for models with 2–3 billion parameters.
In plain terms = phone AI used to "hear one line, reply one line"; now it must "watch everything, judge constantly" — the compute gap needs all three legs of the chip working at once.
What ultimately decides who wins?
Moore Threads SVP Dong Longfei called edge AI "a visible long-term market," showcasing the MTT AICUBE home AI unit (current compute 50 TOPS) and the MTT AIBOOK computing laptop.
His thesis: AI infrastructure will follow the historical pattern of migrating from centralized systems to personal devices.
This means → which technical route wins is not the real question — whether ecosystem partners sign on and revenue-sharing models land is the shared proving ground for all three paths.
Content is for reference only, not financial advice.