AI Agent Era’s Entry Anxiety.

This summer, Elon Musk is set to do something unprecedented in history: merge a large model company with a rocket manufacturing company for an IPO.

OpenAI’s most questionable move right now might be venturing into mobile phones. However, Sam Altman seems to think otherwise.

In Q1 of this year, OpenAI’s revenue and user growth fell short of expectations. Competitor Anthropic, with Claude Code, has attracted the most willing-to-pay users. Following this script, OpenAI should be contracting and focusing on proving its profitability ahead of an IPO by the end of this year or early next year.

Contrary to this, supply chain news suggests it is gearing up to challenge the world’s most mature, closed, and profitable consumer electronics category: the iPhone.

Reports indicate that OpenAI is accelerating the development of its first AI Agent phone, aiming for mass production by mid-2027, with a target of shipping 30 million units in the next two years.

Is it crazy?

Perhaps not. OpenAI seems to have recognized a more pressing issue: ChatGPT is intelligent, but it lacks physical capabilities.

It can answer questions but struggles to complete tasks. It resides within other systems—Apple’s, Microsoft’s, operating systems, browsers—thus lacking true authority.

The focus here is not on why OpenAI wants to build a phone, but rather how the company gradually realized that without its own terminal device, ChatGPT cannot truly thrive.

ChatGPT’s Success as Path Dependence

In April 2026, SpaceX secured an option to acquire Cursor for up to $60 billion later this year.

OpenAI initially believed in models—not phones, browsers, or specific apps. It believed in intelligence itself.

In its worldview, as long as the model is strong enough, the entry points, products, and business models will naturally progress.

This is not just rhetoric. In 2020, OpenAI published the influential Scaling Laws paper, establishing a relatively optimistic belief: as models, data, and computing power scale, intelligence will improve in predictable ways.

In other words, the priority is not to seize entry points first but to continue strengthening the model. With sufficient intelligence, the world will naturally yield.

This belief was validated on November 30, 2022.

On that day, ChatGPT launched. It had no flashy interface, no hardware, no pre-installed platform—just an input box on a webpage. Yet, it provided an unprecedented experience; you could type a sentence, and it would respond like a human.

The shock was not just that AI could converse, but that it did so without relying on any traditional entry points. No phone manufacturers pushed it, no operating systems prominently featured it; users found it themselves.

Within two months, it surpassed 100 million monthly active users, becoming the fastest-growing consumer application in human history.

OpenAI appeared to be right. Microsoft quickly deepened its investment, embedding its capabilities into Copilot, Office, and Bing; Apple also integrated ChatGPT into Apple Intelligence at the 2024 WWDC.

At this point, OpenAI stood at the center of the era: the strongest model, the largest user base, and the deepest collaborations.

However, this is where the problems began.

ChatGPT’s success was so dazzling that it easily led OpenAI to believe that the model itself was the entry point. It didn’t need to own a phone or control an operating system—if the intelligence was impressive enough, users would come on their own.

The real cracks began to emerge from here.

Claude Code Redefines Revenue Rules

The first crack came from Anthropic.

In May 2025, it launched Claude Code. There were no flashy demos or explosive launch events. This product simply integrated into developers’ terminals, codebases, and Git workflows, helping engineers get their work done.

Six months after launch, Claude Code’s annualized revenue reached $1 billion; within a year, it exceeded $2.5 billion. By April 2026, Anthropic’s total annualized revenue surpassed $30 billion.

Meanwhile, OpenAI reported monthly revenues of $2 billion, annualized at about $24 billion.

Anthropic achieved higher revenue with far fewer users than ChatGPT. This is where OpenAI should truly be concerned.

The reason is simple—it penetrated a segment of users most willing to pay.

The question is, why did OpenAI lag behind?

Not because it couldn’t see the potential of Agents. It was the dazzling success of ChatGPT that led OpenAI to continue along its original inertia: building stronger models, expanding the user base, and seeking the next universal entry point.

Over the past two years, OpenAI has launched many 0 to 1 attempts—GPT Store, Sora, Operator, Deep Research—all stemming from this mindset. They collectively point to one judgment: as long as the model is strong enough, new products, new entry points, and new business models will emerge naturally.

However, Anthropic chose a different path. It did not first create a super entry point for everyone but instead embedded Claude Code into developer workflows, repeatedly refining one thing—ensuring AI could complete tasks.

This is where OpenAI was slow. It wasn’t that it didn’t create new products; it failed to immediately capitalize on a high-paying scenario, scaling it from 1 to 100.

Sora is a typical example. It shocked the audience at launch, but video generation consumed massive computing resources, and user retention and business models were unclear. Later, OpenAI shut down Sora, which in some ways was a pruning—realizing that creating an impressive AI demo and penetrating a high-paying workflow are two different things.

Model capabilities can create highlights, but commercial efficiency comes from consistent delivery of results.

At this point, OpenAI finally recognized: Agents are not an added feature but the core of the next phase of AI commercialization. ChatGPT cannot just prove its intelligence; it must demonstrate its ability to complete tasks for users.

But when it truly began to take on tasks, it encountered not the ceiling of model capability but the ceiling of authority.

How to Monetize 900 Million Users

OpenAI is certainly also in pursuit. In May 2025, it launched Codex in direct response to Claude Code. By April 2026, Codex achieved a weekly active user count of 3 million.

However, in the coding arena, OpenAI faces an uphill battle to regain ground—Anthropic has already established a stronghold in the coding Agent mindset, leaving newcomers to play catch-up.

This is why OpenAI has begun reallocating resources: shifting focus from projects that easily create highlights but struggle to penetrate commercial loops towards Agents, enterprise markets, and more foundational research.

What it truly needs to look at is the larger card in its hand—900 million weekly active users.

These users are not programmers and won’t pay for code. Yet, each of them has needs: writing emails, creating proposals, researching, booking travel, shopping, organizing files.

If ChatGPT can evolve from a “talking” entry point to a “doing” entry point, that would represent OpenAI’s true commercial capability.

Imagine a scenario: you want to buy a flight ticket, telling ChatGPT your time, budget, and preferences, and it helps you check flights, compare prices, and look at hotels, ultimately providing you with a confirmation button.

At that moment, a part of the value of travel booking platforms would be bypassed. Price comparisons, ad placements, commissions, and user decision influences would all be redistributed. Buying insurance, paying credit cards, and settling utility bills follow the same logic. As long as the Agent can complete tasks, OpenAI has a chance to earn a share of every transaction commission and every advertising influence.

This is where the true value of 900 million users lies—ChatGPT evolves from merely answering questions to taking over tasks and transaction entry points.

However, once AI starts performing tasks, it is no longer just a model in a chat box. It needs to know your location, see what’s happening on your screen, and access your files, calendar, emails, and payments.

The question then shifts from “Is the model strong enough?” to “Who has the authority?”

And authority is precisely what OpenAI lacks.

ChatGPT Lives in Others’ Houses

OpenAI initially believed that collaboration could solve the entry problem. Apple provided it with the iPhone, and Microsoft offered Office, Windows, and enterprise clients. At the time, this seemed a victory for OpenAI’s model belief.

However, with the arrival of the Agent era, the problem changed.

With Apple, ChatGPT is an external expert being called upon. It can answer questions but cannot truly take over screens, cameras, notifications, payments, and files—these permissions Apple will not relinquish. Otherwise, the iPhone’s “soul” would no longer belong to Apple.

The same goes for Microsoft. In the past, OpenAI provided models while Microsoft integrated AI into Office and other entry points. But as OpenAI began developing Codex and enterprise Agents, it encroached on Microsoft’s territory—Agents naturally need to enter workflows, write code, handle files, and complete tasks for employees, which are core areas of Microsoft’s sovereignty.

Thus, while OpenAI and Microsoft’s relationship did not immediately fracture, the boundaries have changed. In April 2026, both parties adjusted their agreement, converting Microsoft’s exclusive authorization into a non-exclusive one, allowing OpenAI to serve clients on any cloud.

The meaning of this is clear: OpenAI does not want to be merely a supplier within the Microsoft ecosystem. It aims to face clients directly, deliver Agents independently, and secure its own entry points.

At this juncture, its relationship with Apple and Microsoft has become delicate. Because Agents require not just a display position but a default entry point, system permissions, and the first smart terminal users interact with daily.

These are things Apple and Microsoft will not provide, nor can they.

Ultimately, ChatGPT is powerful, yet it continues to reside in others’ houses—Apple’s house, Microsoft’s house, browsers’ houses, operating systems’ houses. It can be called upon, integrated, and serve as a good supplier, but it cannot decide when to appear or what permissions it can access.

A mobile phone, however, is the most aligned with its resource endowment. The 900 million weekly active users are already willing to entrust their questions to ChatGPT—transitioning this mindset to a device is shorter than building an operating system or a browser from scratch.

What it aims to create is not just another iPhone filled with apps but a phone dedicated to Agents—a body that allows ChatGPT to see, access, and execute tasks.

This is also why, in May 2025, OpenAI spent about $6.5 billion acquiring Jony Ive’s hardware company. He was the industrial designer of the original iPhone and one of Steve Jobs’ most important associates. OpenAI seeks him not just to create beautiful hardware but to redefine personal devices in the AI era.

Returning to the initial question: why would a large model company want to build a phone?

What OpenAI desires is not just a phone but sovereignty.

It aims to find a default entry point for ChatGPT. However, this endeavor will inherently position OpenAI against Apple. In the past, Apple could treat ChatGPT as a supplier; if OpenAI truly aims to create a phone for the AI era, it will no longer be a supplier but a competitor to Apple in personal entry points.

Looking back over the past few years, OpenAI’s story has actually undergone a reversal.

It once believed that as long as the model was strong enough, the world would reorganize itself around intelligence. The explosion of ChatGPT indeed validated this belief—it attracted hundreds of millions of users without hardware or pre-installation, relying solely on a web input box.

However, with the arrival of the Agent era, OpenAI realized it still lacked the most critical element: sovereignty.

ChatGPT’s success is both a victory and a form of path dependence. It led OpenAI to believe for too long that the model itself was the answer. Only after Claude Code achieved $2.5 billion in annualized revenue and both Apple and Microsoft hesitated to relinquish system permissions did OpenAI realize that no matter how strong the model, it still needed entry points, authority, and tasks.

Thus, OpenAI’s venture into mobile phone development is not merely about creating a device; it is about giving ChatGPT its first physical embodiment.