OpenAI Is Making a Phone: Why the Hardware Endgame Is Still the Smartphone

A Road Long Paved

When supply chain leaks about OpenAI building a smartphone surfaced, the industry's first reaction was largely "a company crossing into unfamiliar territory." But if you had heard what OpenAI hardware lead Richard Ho said at an internal IEEE exchange session at Stanford two weeks ago, you'd realize — this isn't a pivot. It's an inevitability.

Richard Ho came from Google's TPU team and now heads OpenAI's hardware division. During that closed-door session, he never once mentioned the word "phone," yet every layer of logic he laid out pointed to the same conclusion: AI models ultimately need a hardware vessel of their own, and the smartphone is the only mature answer available today.

The Next Leap in Models Will Be Born from Hardware

The core topic of the exchange session was why OpenAI must build its own hardware.

Richard Ho's answer was remarkably direct: "You have to design hardware for where the model is going, not for where it is today."

The implications run far deeper than the words suggest. Over the past few years, the evolution of large models has been constrained by architectural bottlenecks of general-purpose GPUs. Whether it's the energy consumption problem on the training side or the latency problem on the inference side, neither can be fundamentally solved by simply stacking more compute. OpenAI clearly believes that the next capability leap won't come from algorithmic breakthroughs alone — it will require deep software-hardware co-design.

This mirrors the same logic behind Google building TPUs and Apple developing M-series chips — when your computational needs are sufficiently unique, general-purpose hardware is no longer the optimal solution.

But Richard Ho's ambitions clearly extend beyond training chips in data centers. Throughout the session, he repeatedly emphasized one concept: AI's value must ultimately be unleashed at the "edge." Models can't live in the cloud forever. They need to perceive the real world, respond in real time, and be ubiquitous.

Why a Phone and Not Some Other Device

Over the past two years, the AI hardware startup space has endured a brutal wave of elimination. The Humane AI Pin, Rabbit R1 — products that attempted to redefine the AI device form factor all met with a cold reception from the market.

Their failures revealed a harsh truth: Users don't need a new product category. They need their existing devices to become smarter.

Although Richard Ho didn't directly comment on these products during the session, he offered a key insight — the core challenge of AI hardware isn't "inventing something new" but "embedding intelligence into users' existing behavioral habits." This means that any AI hardware effort must start with a device users already can't live without.

And the smartphone is precisely that device.

From a supply chain perspective, the phone is the most mature consumer electronics category on the planet. Screens, cameras, microphones, speakers, sensors, communication modules — every input and output interface an AI model needs to perceive the world is already on a phone. More importantly, smartphones benefit from the muscle memory of billions of users worldwide.

For OpenAI, building a phone isn't about competing with Apple or Samsung for market share. It's about seizing a complete vertical integration opportunity — from chips to operating system to application layer. Only then can models truly reach the edge and achieve seamless cloud-to-device orchestration.

From Chips to Devices: OpenAI's Vertical Integration Ambition

Piecing together the information Richard Ho shared during the session, OpenAI's hardware roadmap can be roughly divided into three layers:

The first layer is chips. OpenAI is already advancing its custom chip initiative, aiming to tailor computing architectures specifically for its models' inference needs. Richard Ho's TPU background is the critical foundation for this effort.

The second layer is the system. With custom chips in hand, the natural next step is to build a complete software-hardware system around them. This includes on-device inference frameworks, power management, and cloud-model coordination mechanisms.

The third layer is the device. The phone is that device. It's not merely a "shell" — it's the delivery interface for the entire AI experience. The camera becomes the model's eyes, the microphone becomes its ears, and the screen becomes the window through which the model converses with users.

Stack these three layers together and you get a complete vertical integration system. This also explains why Richard Ho said "design hardware for where the model is going" — he's not designing a single chip or a single device, but an entire infrastructure for AI models to "live" in the physical world.

The Phone Is One Destination, Not the Starting Point

It's worth noting that for all model companies, the phone is only one destination — not the starting point.

The starting point is model capability itself. Only when models are powerful enough, efficient enough, and reliable enough does hardware become meaningful. Otherwise, you end up with another Humane AI Pin — an empty shell that promises the future but can't deliver the experience.

OpenAI's decision to push forward with a phone plan at this particular moment suggests they already have sufficient confidence in their models' on-device capabilities. GPT-4o's multimodal abilities, continuous optimization of inference efficiency, and the maturation of cloud-edge collaborative architecture — all of these are prerequisites for an AI phone that actually "works well."

From an industry landscape perspective, this move will also redefine the competitive dimension. Previously, model companies competed on parameter counts, benchmark scores, and API call volumes. But once OpenAI introduces its own phone, the competition shifts to who can deliver the most complete AI experience in the user's pocket.

Outlook: The "iPhone Moment" of the AI Era?

In 2007, Steve Jobs redefined the phone with the iPhone. The core of that moment wasn't how advanced the hardware was, but that software ecosystem and hardware design were being conceived as one unified vision for the first time.

The path OpenAI is taking today follows fundamentally the same logic: When the AI model becomes the new operating system, it needs hardware tailor-made for itself.

Of course, whether OpenAI's phone succeeds depends on too many variables — supply chain management capabilities, pricing strategy, market positioning, and most critically, whether the on-device model experience can truly surpass the existing iPhone + Siri or Pixel + Gemini combinations.

But regardless of the outcome, the direction Richard Ho revealed in that closed-door session is already irreversible: for AI companies, building hardware is not optional — it's mandatory. The phone isn't the ultimate form factor, but it's a necessary waypoint on the road to that ultimate form.

As Richard Ho put it — design hardware for where the model is going. And where the model is going is into everyone's pocket.