Terminal Battle: StepFun Adds Features, but OpenAI's Subtraction is the Real Fight

Last week in Shenzhen, I met an old friend working in smart hardware supply chains. He had just finished a three-day meeting with the StepFun team, and his first words were: "They want to make phones, but they can't even negotiate down the ODM mold fees."

This isn't a joke. Foundation large model companies making terminals sounds like a dimensional strike, but the pitfalls in actual implementation are more numerous than algorithm parameters. StepFun released STEPX in Shanghai, simultaneously launching an AI-native terminal brand. This move resembles those "AI phone" PPTs from 2023—screens embedded in glasses frames, large models stuffed into earbuds—but the core question is: Why do users need another terminal?

Addition vs Subtraction: Two Routes, One Big Bet

Dimension	StepFun (STEPX)	OpenAI (AI Agent Phone)
Product Form	Multi-device matrix (glasses, earbuds, brooches)	Single-point breakthrough (phone)
Core Selling Point	Multimodal interaction, calling large models anytime	Natural language as the sole interaction entry
Implementation Difficulty	Extremely high (hardware supply chain, OS adaptation, fragmented scenarios)	High (but phone form factor is mature, low user cognitive barrier)
Team Execution	Strong algorithms, weak hardware (need to find ODM)	Full-stack self-development of hardware+software+chip (but cost out of control)
Business Model	Sell hardware + subscription—but can gross margin cover R&D?	Hardware loss + API revenue share—betting on scale effects

I checked the transcript of StepFun's launch event; they emphasized "STEPX is an AI-native terminal," meaning interaction is completely driven by large models, without traditional App logic. Sounds cool, but the Ray-Ban Meta smart glasses launched in 2023 had less than 10% DAU—users are willing to take photos and post to social media, but not to keep talking to their glasses. Scenario dependency is too strong, and the assumption for AI phones is "users are willing to voice-chat with AI anytime, anywhere"—this requires complete reshaping of behavioral habits, which tech breakthroughs alone can't solve.

What is the core metric? User daily average usage time. If STEPX devices can't get users to use them for more than 30 minutes a day, they're toys. And if OpenAI's phone can compress voice interaction latency to under 300 milliseconds and make users feel "faster than opening an App," it might truly replace existing interactions.

Team Execution is the Watershed

I know StepFun's team background well; it's a star team from Kai-Fu Lee's lineage, with model capabilities in China's first tier. But making terminals faces them with triple hells:

1. Hardware Supply Chain: ODMs can take orders, but QC, yield rates, and return rates—every line is a pitfall. When Xiaomi made phones, their supply chain team exceeded 200 people. What about StepFun?

2. OS and Ecosystem: An AI-native OS means self-developing underlying interaction frameworks, totally different from Android/iOS. Currently, only Apple's Siri Shortcuts globally achieves this (and it often disconnects). On what basis does StepFun think they can?

3. Sales Channels: Online sales don't move, offline stores are too expensive to set up. Huawei and OPPO have tens of thousands of stores. As a new brand, StepFun's users can barely pronounce "STEPX."

In contrast, OpenAI's approach to phones is more "subtraction": Don't reinvent the wheel. Use the existing phone form factor but replace the OS with AI-driven logic. They negotiated licensing with Apple and Samsung but finally chose self-development—because "only controlling hardware ourselves guarantees latency and privacy." Their core bet is achieving voice interaction latency under 50 milliseconds, making users feel "no difference from talking to a real person." If achieved, it's disruptive; if not, it's just another Siri.

My Judgment: StepFun Made the Mistake Every Startup Makes

Doing addition, trying to mask unclear product definition with feature stacking. They gave users too many choices: glasses, earbuds, brooches... but what users need is "one device solving all AI problems." OpenAI only makes phones because the phone is the only device users carry every day, with a mature ecosystem. StepFun is dispersing resources; a multi-device matrix means multiple times the R&D investment, material costs, and QC risks, while revenue doesn't necessarily grow linearly.

Trend Prediction: By end of 2024, the AI terminal market will face its first wave of consolidation

• If OpenAI's phone can be mass-produced in Q1 2025 with latency <100ms, it will eat up the entire high-end AI interaction market, forcing Apple and Google to follow suit.
• StepFun's STEPX will likely become a "niche novelty item," similar to Smartisan phones back then—good reputation, but shipments possibly under 100,000 units. The only way out is finding B-side scenarios (like factory inspections, medical assistance) to close the loop in vertical scenarios.
• The real winners might be companies that don't make terminals but build AI interaction middleware, such as startups providing ultra-low-latency voice APIs, or toolchain companies helping hardware vendors quickly integrate large models.

Jensen Huang said the iPhone moment for AI hasn't arrived yet. But one thing is certain: The next iPhone won't come from a company that only understands algorithms. If StepFun doesn't find a hardware partner or a channel willing to burn cash for subsidies within 6 months, the buzz from this launch will quickly turn into stale news like the previous "AI smart speakers."

Original Link: https://www.leiphone.com/category/ai/t0fM6SFQ2W7BojX3.html