$6.4B for a Microphone with a Subscription

#The $6.4 Billion Question Nobody Asked

In May 2025, OpenAI acquired Jony Ive's startup io Products for $6.4 billion. The pitch: build a screenless, always-listening AI companion that would replace your phone. A device so elegant Sam Altman called it "the coolest piece of technology the world will have ever seen."

Ten months later, the device is delayed to at least February 2027. The team can't agree on the device's "personality." They don't have enough compute to run it. And the privacy model is, charitably, nonexistent.

Meanwhile, ChatGPT works on every phone, tablet, laptop, and browser already shipping. The cloud is right there.

#The AI Hardware Graveyard Already Has Occupants

Let's check in on the alumni:

Humane AI Pin. Raised $230 million. Shipped fewer than 10,000 units. Almost all were returned. Sold its assets for $116 million to HP, bricking every device it had shipped. A $114 million lesson in product-market fit.

Rabbit R1. Sold 100,000 units on hype alone. Mass returns followed. By early 2026, reports of unpaid employee salaries and financial distress suggested the company's runway was done. RabbitOS 2 pivoted from "autonomous agent" to "AI agent assistant," which is another way of saying "we don't know what this is for either."

Both products confused viral demos with demand. Both shipped hardware to solve problems that software on existing hardware already handles. Both are dead or dying.

OpenAI looked at this wreckage and said: "We should do that, but spend 30x more money."

#Three Problems, Zero Good Answers

According to reporting from Windows Central, the device faces three fundamental issues:

Compute. OpenAI already struggles to get enough compute for ChatGPT. The idea that they'll dedicate inference capacity to a hardware product shipping 40 to 50 million units in year one is ambitious math. The device is supposed to process lighter tasks on-device and offload complex inference to the cloud. Which means it's a cloud product in a hardware case. Which raises the question: why not just use the cloud?

Privacy. The device is always on. Cameras and microphones are always active. It continuously captures your location, activity, communication, and environmental context. Critics predict "AI-generated Glassholing", the same social backlash that killed Google Glass, but now with inference attached.

In a post-GDPR, post-Cambridge Analytica world, shipping a pocket surveillance machine with no screen for users to see what it's recording is a bold privacy strategy. "Bold" meaning reckless.

Personality. The team is reportedly struggling to define what the device's AI companion should "feel like." The goal: a helpful friend that's a computer but not a "weird AI girlfriend." The fact that this distinction requires active engineering effort tells you everything about where we are.

#The Cloud Works. Use It.

Here is the uncomfortable truth OpenAI doesn't want to confront:

ChatGPT already runs on 400 million weekly active users' devices. iPhones. Androids. Laptops. Browsers. Voice mode works. Vision works. Agents work (or will, when they work at all). The software is the product. The interface is whatever glass rectangle the user already owns.

Building a separate device to access the same inference, but with worse battery life, no screen, no app ecosystem, and a microphone that never turns off, is not innovation. It is a solution in search of a problem.

The pitch for edge AI hardware was that latency and privacy would be better with on-device processing. But OpenAI's own architecture offloads complex inference to the cloud. So you get the privacy risk of an always-listening device AND the latency of cloud inference AND the cost of hardware logistics. The worst of every world.

#The Phone Gambit Is Even Wilder

If the earbuds project wasn't enough, OpenAI is also reportedly building a full smartphone where AI agents replace apps entirely. Qualcomm and MediaTek are designing the chip. Analyst Ming-Chi Kuo projects 300 to 400 million annual shipments by 2028.

Three hundred million units. That exceeds iPhone volumes.

For a company that has never manufactured a physical product, never managed supply chains, never dealt with carrier relationships, never handled hardware returns, never navigated FCC certification, and whose core competency is training language models. The confidence is almost admirable.

The phone would replace apps with agents. Instead of opening Maps, you tell the AI where to go. Instead of a messaging app, you tell the AI who to contact. This requires every service on the internet to build agent-compatible APIs, every user to trust an intermediary with their entire digital life, and the AI to never hallucinate a restaurant, misroute a message, or book the wrong flight.

We can't get ChatGPT to consistently count the number of R's in "strawberry," but sure, let's route all human communication through it.

#The Real Play: Data, Not Devices

If you're cynical (and you should be), the hardware play makes more sense as a data acquisition strategy than a product strategy.

An always-listening, always-watching device that captures location, activity, communication, and environmental context isn't a phone replacement. It's a training data pipeline. Every conversation, every interaction, every ambient sound feeds the model. The device doesn't need to be profitable if it generates the richest multimodal training dataset ever assembled.

This isn't a phone. It's a microphone with a subscription.

#What Would Actually Work

If OpenAI genuinely wanted to improve AI hardware interaction, the path is obvious:

Make the API better. Let device manufacturers integrate GPT natively. Apple is doing this with its own models. Samsung could do it with GPT. The hardware already exists.

Invest in on-device models. Ship smaller models that run locally with zero cloud dependency. Actually solve the latency and privacy problems instead of creating new ones.

Build the agent protocol layer. If agents are the future, build the open standard that lets agents talk to services. Don't force users to buy new hardware to access it.

Every one of these paths costs less, ships faster, and reaches more users than a hardware product. None of them require buying a design studio for $6.4 billion.

#The Verdict

OpenAI is a company that runs one of the most successful cloud products in history, looking at a graveyard of failed AI hardware, watching its own device get delayed by compute, privacy, and identity crises, and deciding to double down anyway.

The cloud called. It works. The phone in your pocket has a browser, a microphone, and a camera. ChatGPT is already on it.

Build better software. Ship better models. Leave the hardware to the companies that know how to ship 300 million units of anything.

Or don't. The AI hardware graveyard has plenty of room.