AI no longer lives inside individual apps. It breaks through the glass wall.
A new kind of device is taking shape in the hands of Chinese technology firms — one where artificial intelligence no longer lives in distant servers but is woven into the very architecture of the phone itself. Companies like Stepfun and ByteDance are forging partnerships that blur the line between hardware maker and AI developer, treating chips, memory, and software as a single designed system rather than assembled parts. The smartphone, long a platform for applications, is becoming an autonomous agent — capable of orchestrating tasks across the full breadth of a user's digital life. Whether this transformation reaches ordinary consumers depends on one stubborn material reality: the cost of memory.
- Chinese firms Stepfun and ByteDance are racing to ship AI agent phones that embed artificial intelligence at the operating system level, dissolving the boundaries between apps and turning the phone into an orchestrating intelligence.
- The shift is already measurable — China shipped over 100 million AI-capable devices last year, and globally, nearly half of all smartphones shipped in 2026 are expected to carry generative AI capabilities.
- Five hardware segments — chips, memory, cooling, batteries, and radio frequency components — are being simultaneously redesigned to meet the thermal, power, and computational demands of always-on local AI processing.
- Memory costs are the critical chokepoint: even a compact AI model demands enough RAM to push wholesale phone prices above $400, effectively barring the mid-range and budget markets where most consumers shop.
- The industry is contracting overall — global shipments are projected to fall to their lowest recorded level — yet the pressure to restructure supply chains around AI is intensifying rather than easing.
The smartphone is ceasing to be a platform for apps and becoming something closer to an autonomous agent — and the companies building that future are restructuring hardware supply chains to match.
Stepfun, a Chinese AI startup, is preparing to launch its first AI agent phone through a deep partnership with manufacturer Huaqin Technology. This is not conventional outsourcing. The two firms are designing hardware, software, and AI models as a unified system. Stepfun has already seeded the ecosystem: its open-sourced Step-GUI models and a developer protocol that enables agent-capable phones to be deployed in minutes. ZTE's Nubia division launched a flagship phone running Stepfun's GUI Agent last October, and ByteDance is preparing to show its own Nubia-built AI agent phone at the 2026 World Artificial Intelligence Conference.
At the center of this shift is a new concept of what a phone does. These devices place AI at the operating system level — not inside individual apps, but above them, breaking through the invisible walls that have always kept applications separate. WeChat and several manufacturers have already launched voice-driven Agent-to-Agent features, with Honor phones leading implementation. Analysts describe these devices as "super carriers," systems where AI orchestrates everything.
The adoption numbers are striking. Counterpoint Research projects that 45 percent of all smartphones shipped globally in 2026 will carry generative AI capabilities, rising to 52 percent by 2027. OpenAI is targeting mass production of its own agent phone by mid-2027.
Yet a hard constraint shadows the optimism: memory. Running even a compact large language model on-device requires substantial RAM, pushing standard specifications well beyond today's 12 to 16 gigabytes. Counterpoint calculates that AI phones cannot currently be manufactured below $400 wholesale — a threshold that excludes the mid-range and budget segments where most phones are sold.
That bottleneck is reshaping five hardware categories at once. Neural processing units are becoming the primary computing engine. Batteries are adopting silicon-carbon and semi-solid-state chemistry to sustain AI's power demands. Cooling systems are incorporating graphene and vapor chambers to manage heat without adding bulk. Radio frequency components are being upgraded for WiFi 7 and advanced 5G, so on-device AI can coordinate fluidly with the cloud.
The partnerships forming across this landscape — Stepfun and Huaqin, ByteDance and Nubia, WeChat and device makers — are not isolated product launches. They are the early architecture of a new kind of ecosystem, one where the phone, the model, and the operating system are conceived as one. The moment memory costs fall enough to unlock mass-market pricing, the deeper restructuring of the industry will begin in earnest.
The smartphone is no longer just a phone. It's becoming the primary battleground where artificial intelligence moves from distant data centers into the device you carry in your pocket, and the companies racing to win that fight are reshaping entire hardware supply chains in the process.
Stepfun, a Chinese startup that has bet more aggressively than most on bringing AI agents directly to phones, is about to launch its first AI agent smartphone. The device will be manufactured by Huaqin Technology, a Shanghai-listed company, but this is not a typical outsourcing arrangement. The two firms are building something more integrated—a deep partnership where the hardware, software, and AI models are designed to work as a single system from the ground up. Stepfun has already been laying groundwork for this moment. Last year it open-sourced its Step-GUI series models, and released an open protocol that allows developers and phone makers to deploy an agent-capable phone in as little as ten minutes. In October, ZTE's Nubia division launched a flagship phone equipped with Stepfun's GUI Agent model, capable of understanding commands and invoking tools across different apps to complete tasks. Now ByteDance is preparing to exhibit its own AI agent phone, created jointly with Nubia, at the 2026 World Artificial Intelligence Conference.
What's happening is a fundamental shift in what a smartphone is supposed to do. These AI phones are no longer devices that run applications in isolation. They're becoming what analysts call "super carriers"—systems where artificial intelligence sits at the operating system level and orchestrates everything the phone does. WeChat and several manufacturers recently launched an Agent-to-Agent feature that lets users control messaging and other functions through voice commands to the phone's AI assistant, with Honor phones leading the implementation. The AI doesn't live inside individual apps anymore. It breaks through what the industry calls the "glass wall" that traditionally separates one application from another, reaching directly into the functions of multiple programs at once. This is a different kind of computing experience entirely.
The adoption curve is steep. China shipped more than 100 million AI phones, AI computers, and similar smart devices last year, and that number is expected to keep climbing. For the first time, sales of AI-equipped phones and computers are projected to surpass non-AI products this year. Globally, the picture is similar. Counterpoint Research estimates that 45 percent of all smartphones shipped worldwide in 2026 will have generative AI capabilities, up from 36 percent in 2025, and climbing to 52 percent by 2027. OpenAI is pushing toward mass production of its own agent phone by the first half of 2027.
But behind these optimistic projections sits a hard constraint: memory costs. Running large language models on a phone requires substantial amounts of RAM to store the model weights. Even a relatively compact seven-billion-parameter model like LLaMA needs at least 3.9 gigabytes of memory when compressed to its smallest practical form. This is pushing phone memory specifications upward from today's standard 12 or 16 gigabytes to much higher levels. The problem is that this added memory makes phones expensive. Counterpoint Research calculates that it's currently impossible to manufacture an AI phone for less than $400 wholesale—a price point that locks out the mid-range and budget markets where most phones are actually sold. Until memory becomes cheaper, AI phones will remain a premium product.
The memory shortage is part of a broader supply chain crisis. Global smartphone shipments in 2026 are expected to drop 13.9 percent year-over-year to 1.08 billion units, the lowest on record. Yet even as the overall market contracts, the shift toward AI is forcing manufacturers to rethink five critical hardware segments simultaneously. Chips are being redesigned around a new architecture—CPU, GPU, and NPU working together in what's called heterogeneous computing, with the neural processing unit becoming the primary engine for running AI models locally. Batteries are being upgraded with silicon-carbon anodes and semi-solid-state technology to handle the increased power draw of constant AI processing. Cooling systems are becoming more sophisticated, with graphene and vapor chamber heat spreaders needed to manage the thermal load without making phones thicker or heavier. Radio frequency components are being enhanced to support WiFi 7 and advanced 5G, ensuring that on-device AI can collaborate seamlessly with cloud systems in real time.
What emerges from all this is a picture of the smartphone industry at an inflection point. The integration between Stepfun and Huaqin, the collaboration between ByteDance and Nubia, the partnership between WeChat and phone makers—these aren't isolated product launches. They're signals that AI phones are no longer simple stacks of hardware components but deeply woven ecosystems where the operating system, the AI models, and the physical device are designed as one thing. The question that will determine whether this technology reaches ordinary consumers isn't whether AI phones can be built. It's whether memory will become affordable enough to make them accessible beyond the premium market. When that bottleneck breaks, the real reshaping of the industry begins.
Citas Notables
Memory remains the key factor determining whether generative AI can quickly enter the mid-to-low-priced phone market— Counterpoint Research
AI is upgrading from a function within the phone to a system-level orchestrator that permeates interaction— Sinolink Securities
La Conversación del Hearth Otra perspectiva de la historia
Why does it matter that Stepfun and Huaqin aren't just doing a typical manufacturer-client deal? Why does "deep integration" change anything?
Because it means the AI isn't bolted onto the phone as an afterthought. The system is designed from the silicon up to run these models efficiently. When you integrate at that level, you can optimize everything—the chip architecture, the memory layout, the power management—to work together. A traditional ODM relationship would mean Stepfun hands over specs and Huaqin builds it. This is different. They're solving problems together.
You mention that AI phones can't cost less than $400 right now. That seems like a huge barrier. How long before memory gets cheap enough?
The honest answer is nobody knows exactly. But the supply crisis is temporary. As memory production catches up and as engineers get better at compressing models—running them in 4-bit instead of 8-bit, for instance—the math changes. Once you can run a useful AI model in 8 or 10 gigabytes instead of 16, suddenly a $300 phone becomes possible. That's when this moves from luxury to mainstream.
The article mentions five hardware segments being reshaped. Is that just marketing language, or are these real, fundamental changes?
They're real. Take cooling. An AI phone generates more heat because the NPU is constantly working. You can't just make the heatsink bigger—the phone would be too thick. So you need graphene or vapor chambers, which are more expensive and harder to manufacture at scale. Every segment has that same problem. You're not just adding a feature. You're redesigning the entire device.
What does it mean that AI is moving from "a function within the phone" to "a system-level orchestrator"?
Right now, if you want Siri or Google Assistant to do something, it works within its own sandbox. It can't easily reach into your banking app or your email or your calendar and pull information across them. An AI agent phone breaks those walls down. The AI sits above the apps and can orchestrate them. That's a fundamentally different kind of power—and a different kind of responsibility, frankly.
You mention that global smartphone shipments are expected to drop to a record low this year. How does that square with all this growth in AI phones?
The overall market is shrinking because people are holding onto phones longer and because memory costs are making new phones expensive. But within that shrinking pie, AI phones are growing fast. It's a shift in composition, not absolute growth. The industry is consolidating around AI as the differentiator.