OpenAI Is Reportedly Building an AI Phone—and It

There's a weird tension at the heart of modern AI assistants: they're getting smarter, but they're still fundamentally trapped. ChatGPT can write code, analyze images, and hold conversations that feel increasingly human—but on your iPhone, it's still just another app in a folder, subject to Apple's rules about what it can and can't touch.

According to new supply chain reports from TF International Securities analyst Ming-Chi Kuo, OpenAI is trying to solve that problem the hard way: by building its own phone. The company is reportedly working with MediaTek and Qualcomm on custom mobile processors, while Luxshare Precision has won the exclusive manufacturing contract. Mass production is expected in 2028, with final chip specs locked in by late 2026 or early 2027.

This isn't vaporware. These are the moves companies make when they're serious about hardware.

The App Sandbox Problem

The core issue is architectural. Right now, even a powerful AI assistant on iOS exists inside what developers call a "sandbox"—a deliberately restrictive environment designed to keep apps from interfering with each other. That's good for security and privacy, but it's terrible for the kind of seamless, context-aware assistance that AI companies keep promising.

Say you want to order food, check your calendar, compare restaurant options across multiple apps, make a payment, and text your friend the details. That's five or six different apps, each with its own interface, each requiring manual navigation. An AI agent might understand exactly what you want, but the phone itself treats it like any other third-party app—limited, permission-gated, fundamentally separate from the system.

Kuo's argument is straightforward: "OpenAI needs full control over the operating system and hardware if it wants the agent to offer truly complete services." If AI is supposed to become the primary interface—not just a feature—then the current app-centric model starts to look outdated.

The vision is a phone where you don't really "open apps" anymore. You tell the device what you want done, and the AI figures out which tools, services, and APIs it needs to orchestrate. Apps might still exist, technically. You'd just stop thinking about them as separate destinations you need to visit manually.

Cloud + Edge = Battery Math

That vision has hard technical requirements, though. The processor needs to constantly maintain context (your location, schedule, recent conversations, app state) without annihilating battery life. Light tasks—checking your calendar, drafting a message—could run locally using smaller on-device models. More complex reasoning—comparing flight options across six booking sites, analyzing a contract—would get kicked to the cloud.

This is why MediaTek and Qualcomm matter. OpenAI isn't just building a normal phone and slapping ChatGPT on top. The chip itself would need to be designed around agentic behavior—power management, memory hierarchy, context switching between local and cloud processing.

Kuo offered an interesting comparison point: revenue from a single high-end AI training chip roughly equals the revenue from 30 to 40 AI phone processors. So per-unit margins are lower, but the scale is massive. OpenAI is reportedly targeting the global high-end phone market—around 300 to 400 million units annually. Even a small slice of that creates a significant new business line.

The Apple Talent Raid

OpenAI's hardware team has serious pedigree. According to The Information, the company hired more than 20 hardware experts from Apple last year alone. The team includes Tang Tan (25-year Apple veteran who worked on iPhone and Apple Watch) and Evans Hankey (former head of Apple's industrial design team). Product design is being handled by LoveFrom, the studio led by former Apple design chief Jony Ive.

Apple was apparently worried enough about the talent drain that it canceled an annual closed-door meeting originally planned for China, per the same report.

The rumored product lineup extends beyond phones. A smart speaker ($200-$300) is expected in February 2027. AI headphones code-named "Dime" or "Sweet Pea" are in development, featuring a metal cobblestone design and 2-nanometer chips. Smart glasses targeting 2028 would compete directly with Meta's Ray-Ban collaboration and Apple's rumored AR glasses. There's also a smart lamp prototype (launch uncertain) and some kind of AI pen or pocket device that Sam Altman has hinted at multiple times.

Altman once described smartphones as being "like Times Square—full of information bombardment and fragmented attention." He contrasted that with what OpenAI wants to build: "something more like a lakeside cottage where you can close the door when you need to focus."

That framing matters. This probably won't be marketed as "another smartphone." It'll be positioned as a calmer, more intentional, agent-first way to interact with technology.

China's Faster, Messier Approach

While OpenAI is building from scratch, China is already testing a more aggressive version of the AI phone concept. ByteDance partnered with ZTE to launch the Doubao phone (Nubia M153) at the end of last year. Engineering prototypes sold out immediately. The original price was 3,500 yuan (~$480), but resale prices reportedly hit 36,000 yuan (~$5,000) at one point.

Doubao's technical approach is fascinating and kind of chaotic. Instead of waiting for apps to build proper APIs for AI agents, the system uses what's called a GUI agent—it literally reads the screen and simulates manual actions, like a person would. This lets the AI compare prices across platforms, organize files, draft WeChat messages, book flights, order food—all by "seeing" the interface and tapping/swiping accordingly.

The upside is speed. People could actually use this thing in late 2024. The downside is security. WeChat, Alipay, Taobao, and multiple banking apps started blocking the Doubao phone for what are, honestly, pretty reasonable concerns. If an AI can bypass normal app boundaries and imitate user behavior, that's a nightmare scenario for payment systems and messaging platforms. It's useful, but it also punches holes through the entire permission model that keeps mobile apps relatively secure.

Doubao 2.0 is already in development (expected mid-Q2 2025), and ByteDance is reportedly in talks with other Android manufacturers—Honor, Vivo, and others. Blogger Digital Chat Station described it as "a wave of AI OS and Doubao AI phones approaching."

So you've got two paths: China moving fast by retrofitting Android phones with aggressive AI layers, and OpenAI building the whole stack from scratch. Both point to the same conclusion—AI at the app level isn't sufficient. If the agent is just another feature inside someone else's phone, it stays fundamentally limited.

The AGI Elephant in the Room

There's also a weird undercurrent around OpenAI's broader messaging. Sam Altman recently posted that after AGI arrives, "no one is going to work and the economy is going to collapse." Then he also posted that he's switching to polyphasic sleep because GPT-5.5 Codex is so good he can't afford to miss work time by sleeping normally.

The irony is... striking. The person building technology that might make work obsolete is also so excited by that technology that he wants to work more.

Many tech leaders are skeptical of the whole AGI framing anyway. Peter Steinberger (creator of Multibot) argued on a Y Combinator podcast that the industry should focus on specialized intelligence instead of general intelligence. Anthropic president Daniela Amodei called the AGI concept "outdated." Google DeepMind's CEO has argued that AGI can't be achieved without world models.

And OpenAI itself seems to be quietly backing away from the term. The company's 2018 charter mentioned AGI 12 times. The 2026 update mentions it twice. That's a significant tonal shift. The old charter was structured around a future AGI breakthrough. The new version focuses on "iterative deployment"—society adapting step by step as AI systems get stronger.

One line from the new principles: "The world needs to grapple with each successive level of AI capability." That's a very different framing from waiting for one giant AGI moment.

The 2026 charter also removed a notable commitment from 2018, where OpenAI promised that if another safety-focused lab got close to AGI first, OpenAI would stop competing and help them instead. That's... no longer in there.

So while OpenAI builds hardware to give its AI more control over your daily life, it's also quietly rewriting the philosophical commitments that were supposed to guide that work. Whether those two things are related is up to you to decide. But they're both happening, and they're both worth watching.

—Yuki Okonkwo, AI & Machine Learning Correspondent