OpenAI's Jony Ive Device: What's Confirmed, What's Coming

On May 21, 2025, OpenAI announced it had acquired io Products, the hardware design company founded by Jony Ive and Sam Altman, in an all-equity deal valued at approximately $6.5 billion. The press release was unusually short. The video that accompanied it -- nine minutes of Ive and Altman walking through San Francisco -- was unusually long. Altman called the unannounced device "the coolest piece of technology the world will have ever seen." Ive, who almost never gives interviews, said the work had "captured our imagination" more than anything he had done since the original iPhone.

A year later, the device still has not shipped. But the contours have hardened. The Information has reported that OpenAI is targeting late 2026 or early 2027 for a first consumer launch. Bloomberg's Mark Gurman wrote that the team has grown to roughly 55 ex-Apple hardware and design engineers, with Tang Tan -- the iPhone hardware lead who left Apple in 2023 -- overseeing product engineering and Evans Hankey, Ive's successor as Apple's head of industrial design, leading the design language. Foxconn and Luxshare are reportedly splitting manufacturing roughly 50/50. The internal unit target is 100 million devices in the first two years.

That last number is the one that matters. It is also the one nobody outside OpenAI seems to take literally.

This article is about why OpenAI is doing this at all, what the actual product appears to be, why every other AI lab is racing into hardware behind it, and why -- despite the most accomplished hardware team ever assembled outside Cupertino -- the bet remains structurally difficult. The Ive device is not really a product story. It is a distribution story dressed as a design story.

What Is Actually Confirmed

The public record is narrower than the coverage suggests. Stripping out speculation, here is what OpenAI itself has confirmed or what multiple independent reports have corroborated:

What is conspicuously not confirmed: any product name, the specific form factor (pendant, clip, puck, glasses, all have been suggested), the exact mix of sensors, the operating model (always-on listening vs. wake-word vs. button), the cellular story, and the relationship to existing ChatGPT subscriptions.

The single most reliable signal is the hiring. You do not hire 55 senior Apple hardware engineers, Foxconn's preferred new program manager, and Tang Tan to build a software demo. The team is sized for a real product at consumer scale. Whether it ships on time is a different question.

Why OpenAI Cannot Stay Software-Only

The strategic argument for the Ive device begins with a problem OpenAI cannot solve from inside its own product. ChatGPT is the largest consumer AI product in history. It is also a tenant.

ChatGPT runs on iOS and Android. Both operating systems are controlled by companies that ship their own AI. Apple Intelligence ships with every new iPhone. Gemini is now the default assistant on Pixel and is being pushed deeper into Android with each release. Both companies decide what counts as the default assistant, which apps get system-level integrations, and which AI experiences live one tap away from the lock screen versus four taps deep inside a third-party app.

Ben Thompson has made this point repeatedly at Stratechery: OpenAI's distribution position is structurally weaker than its product position. ChatGPT is the better product today. It is also the product that has to be opened. Siri and Gemini are the products you talk to without thinking.

Benedict Evans framed the same problem differently: the platform shift to AI is happening on platforms that two companies own. If you are OpenAI, you can be the most-used AI service in the world and still be one App Store policy change away from a substantial revenue hit. Apple's 2025 decision to allow ChatGPT as a "system integration partner" inside Apple Intelligence was simultaneously a win (system-level placement) and a warning (Apple chose, Apple can unchoose).

The Ive device is the answer to that asymmetry. If OpenAI ships its own hardware -- even at modest volume relative to the iPhone -- it owns the wake word, the microphone, the camera, the storage, the model routing, the subscription relationship, and the upgrade cycle. None of that exists today.

Altman's framing internally, according to reporting from Axios, has been blunt: "We do not want to wake up in 2030 as the AOL of AI." That is a reference to a specific lesson. AOL was the dominant consumer internet product of the late 1990s. It got distributed to disk for free, then bundled into Windows, then made irrelevant by the fact that browsers and broadband connections were not products AOL owned. Distribution beat product. It usually does.

What The Product Actually Appears To Be

Across more than a dozen reports from The Information, Bloomberg, The Wall Street Journal, and The Verge, the consistent description of the device is narrower than the speculative descriptions suggest. It is not glasses. It is not a phone. It is, in the most consistent leaks, a small object designed to be carried -- pocket, lanyard, or clipped to clothing -- with the following properties:

• Small enough that it disappears in normal carry, closer to an AirPods case than a phone.
• Largely screenless. Some reports describe a minimal visual indicator -- a single small display, an LED ring, or a projected interface -- but not a smartphone-style touchscreen.
• Voice-first interaction, with a camera for visual context capture.
• Always-on or near-always-on listening with on-device wake detection.
• Cellular connectivity, but tethered to a phone for most heavy data, at least at launch.
• A new ChatGPT subscription tier bundled with the device, not a one-time hardware purchase.

The most useful way to understand the product is by what it is replacing. It is not replacing the iPhone. It is replacing the gesture of pulling out a phone, unlocking it, opening ChatGPT, and typing a question. The device collapses that into a sentence spoken into the air.

That collapse is also why the product is hard. The marginal value of saving four seconds and three taps is not obvious to most consumers. The marginal value of having a persistent AI relationship that captures context throughout the day -- that knows what you just discussed in a meeting, what is on the whiteboard, what is in your inbox -- might be. But the second value proposition requires trusting OpenAI with a constant audio and visual feed of your life. That is a different sale than "the new ChatGPT app is faster."

The AI-Hardware Comparison Set

The Ive device is not entering an empty market. It is entering a market littered with the wreckage of previous attempts. The comparison set is small and instructive.

The pattern is unambiguous. Standalone "AI device" products without a fashion or accessory category to anchor them have failed. Humane and Rabbit both tried to convince consumers that an AI relationship was worth a dedicated, ugly object. Consumers said no. Meta Ray-Bans succeeded because they were Ray-Bans first and AI second. Apple Watch succeeded because Apple already owned the wrist.

John Gruber wrote at Daring Fireball that the only credible blueprint in the comparison set is Meta Ray-Bans, not the iPhone. The Ive device's success or failure will not be measured against the iPhone's 6.1M first-year units. It will be measured against whether OpenAI can make a screenless object that consumers actually want to be seen wearing or carrying every day. That is a design problem before it is an AI problem. Which is presumably why Ive is in the building.

There is one more lesson buried in the comparison set. Humane's Ai Pin and Rabbit's R1 both failed despite functional hardware because the AI on the back end was not good enough. The device promised more than the model could deliver. OpenAI does not have that problem -- if anything, OpenAI has the opposite problem, where the model can do more than current devices expose. But the inverse risk is real: a beautiful, well-built Ive device shipped with a 2026-era ChatGPT model still has to do something delightful enough to justify daily carry. Beautiful hardware that mostly tells you it cannot help right now is its own failure mode.

Why Every AI Lab Is Now A Hardware Company

The OpenAI move is the loudest, but it is not the only one. The pattern across the major AI companies has shifted decisively toward hardware over the past 18 months.

Meta has shipped over 8 million Ray-Ban Meta units and is reportedly preparing a higher-end model with a small display. Mark Zuckerberg has made the AI hardware bet explicit on multiple earnings calls, treating glasses as the next computing surface.

Google ships Pixel phones, Pixel Buds, and a Gemini-first software stack that is increasingly designed for ambient interaction. The Pixel 11's "Magic Cue" features point toward an ambient AI model running on Google's own silicon, on Google's own devices, with Google's own assistant.

Microsoft has Copilot+ PCs with NPUs and a hardware spec defined around AI workloads. It also retains the deepest enterprise distribution of any AI provider through Windows and Office.

Anthropic is the most software-bound of the major labs. Reuters reported in March 2026 that Anthropic was in discussions with a major consumer electronics manufacturer about embedding Claude into a dedicated companion device, with the partner taking the hardware risk and Anthropic taking the AI subscription. That arrangement -- AI lab as software layer inside someone else's device -- is the lower-cost version of the OpenAI bet. It is also the version that does not produce a true distribution moat.

The strategic logic is the same across all of these companies. Whoever owns the interface to AI in the next decade will earn most of the surplus. Whoever rents that interface from Apple or Google will not. The only debate is whether to build, partner, or buy. OpenAI bought. Meta built. Google was already there. Anthropic is partnering. Microsoft is doing all three.

The contrarian read here is worth stating clearly. AI labs are not building hardware because hardware is a good business. Hardware is, generally, a terrible business. Margins are thin, inventory risk is real, manufacturing is brutal, and consumer taste is fickle. AI labs are building hardware because the alternative -- being a permanent tenant on someone else's platform -- is worse. It is a defensive bet financed by offensive AI revenue.

The Apple Problem

The single biggest variable in whether the Ive device succeeds is not what OpenAI ships. It is what Apple does in response.

Apple's position is awkward but powerful. Apple Intelligence has underwhelmed since its initial 2024 announcement. The promised next-generation Siri has slipped repeatedly. Apple's internal models have been described by multiple ex-employees, in WSJ reporting, as a generation behind frontier labs. The company's response has been to integrate third-party models -- ChatGPT in 2025, Gemini reportedly added in 2026 -- as system-level options inside Siri.

This is the cooption strategy. Apple does not need to beat OpenAI on model quality if Apple can offer "Use ChatGPT" as a default setting inside iOS. ChatGPT gets distribution. Apple keeps the customer relationship. The user never needs to download a separate app. The friction that the Ive device is supposed to solve -- pulling out the phone, opening ChatGPT -- gets partially solved by Apple itself.

If Apple successfully integrates ChatGPT (and Gemini, and eventually Claude) deeply enough into iOS that calling on AI feels as native as calling on Siri, the marginal value of a separate $499 OpenAI device drops sharply for the median consumer. Why carry an extra object when your phone does the same thing?

OpenAI's counter is that integration is not the same as ownership. Apple controls the routing, the prompts, the privacy policies, the user data that gets exposed to OpenAI, the rate limits, and the commercial terms. A "default ChatGPT inside iOS" is still a tenancy. The Ive device is the only product where OpenAI sets all of those parameters itself.

This is also where the bet gets philosophically interesting. OpenAI is wagering that the iPhone-era assumption -- that the smartphone is the universal interface and everything else is an accessory -- will not hold for AI. If that bet is right, the Ive device becomes a platform of its own. If it is wrong, the Ive device becomes a $599 paperweight that Apple Intelligence makes redundant within two years.

Most platform-replacement bets are wrong. That is the base rate. The exceptions -- iPhone replacing PC-centric computing, the web replacing client software -- are exceptions precisely because they did not try to compete with the dominant device on the dominant device's terms. They built a new interaction model. The Ive device's success depends on whether ambient, screenless AI is a genuinely new interaction model or a worse version of the one we already have.

Manufacturing, Supply Chain, And The 100 Million Number

The 100-million-unit internal target across the first two years is, in context, very aggressive. The iPhone shipped 6.1 million units in its first year. It took Apple nearly 12 years to reach 200 million annual iPhone units. Meta Ray-Bans, the most successful AI-era consumer hardware launch, shipped roughly 8 million units across two generations. The Ive device's target is, on paper, more than 12 times the iPhone's launch year and more than 12 times Meta Ray-Bans' total cumulative volume.

Reuters' reporting on the manufacturing split -- Foxconn 50%, Luxshare 50% -- is itself a sign of how aggressive the volume planning is. Splitting a launch program across two contract manufacturers from day one is a capacity-and-redundancy bet, not a cost optimization. Apple did not dual-source the iPhone at launch. OpenAI is planning as if demand will exceed what any single manufacturing partner can supply.

There are two readings of this. The optimistic reading is that OpenAI's team has internalized lessons from the iPhone and Apple Watch and is building manufacturing redundancy from the start. The pessimistic reading is that 100 million is an aspirational number meant to anchor internal urgency, not a real forecast. Hardware programs miss volume targets routinely; the historical hit rate on first-generation consumer electronics meeting their internal volume plan is well under 50%.

The supply-chain reporting suggests the team is treating ramp risk seriously. The Wall Street Journal wrote that OpenAI has placed long-lead orders for custom microphones, image sensors, and a low-power SoC reportedly co-designed with a major semiconductor partner. This is the kind of supply-chain commitment that makes sense for tens of millions of units. It also makes sense as a way to lock in capacity before Apple, Google, or Meta can.

What does not yet appear in the supply-chain reporting is a coherent retail story. OpenAI does not have stores. It does not have a logistics network. It does not have warranty and repair infrastructure. The current public assumption is that the device will launch through some combination of OpenAI's own website, a deal with a major U.S. carrier, and possibly Best Buy or similar big-box partners. None of that has been confirmed. None of it is trivial.

The Money Problem

OpenAI is burning a great deal of money. The Information has reported that OpenAI's 2025 operating loss was roughly $5 billion on revenue in the $4-5 billion range, and that 2026 is on track for a substantially larger absolute loss as infrastructure investment continues to outpace revenue growth. The company is funded largely by Microsoft and a tightening syndicate of growth investors at progressively higher valuations.

Adding a hardware product line to that operating picture is non-trivial. Consumer hardware requires upfront tooling, inventory commitments, marketing spend, and a working-capital cycle that does not exist in pure software. A 100-million-unit ramp implies meaningful CapEx that does not earn revenue until the device ships and sells.

There are two structural mitigants. The first is that the device is, financially, a customer-acquisition cost for ChatGPT subscriptions, not a standalone revenue line. A user who buys a $499 device and subscribes to a $20/month ChatGPT tier produces roughly $240/year in recurring revenue for as long as they stay subscribed. At decent retention, the device pays for itself in 18-24 months even at zero hardware margin. That is the model. OpenAI does not need to make money on the device. It needs the device to lock in subscriptions.

The second mitigant is that OpenAI's funding base is unique. The company has access to capital at terms no normal hardware startup could match. The total cost of the Ive program -- acquisition, R&D, tooling, launch -- is real, but it is a single-digit-percent line item against OpenAI's overall capital plan, which includes the much larger Stargate infrastructure commitment.

The risk is not that the hardware bet bankrupts OpenAI. The risk is that it distracts the company. Hardware is operationally hostile to software cadence. Releases are annual, not weekly. Mistakes cost months of inventory rather than a hotfix. The cultural compatibility between a hardware org and an AI research org is famously difficult; Apple has spent two decades managing it.

If the device ships and underperforms, the financial cost is contained. The strategic cost -- the message it sends that OpenAI's distribution problem cannot be solved -- is the larger risk.

Five Predictions And What To Watch For

Putting the bets in order, from most to least confident:

1. The first device will ship later than the current target. Hardware programs of this scope slip. A late-2026 internal target almost certainly becomes a Q1 2027 launch at the earliest, with first units in customers' hands probably Q2 2027. The first sign of slippage will appear in supply-chain reporting about engineering validation builds, not in OpenAI's public statements.

2. The first-year unit number will be a fraction of the internal target. A 100-million-unit two-year plan is a forcing function, not a forecast. First-year shipments in the 3-8 million range would be a strong launch by any historical comparison and would still leave OpenAI far short of its internal goal. Anything above 10 million in year one would be unprecedented for an AI-native device category.

3. The device will not be sold standalone. OpenAI's structural advantage is recurring subscription revenue, not hardware margin. Expect the device to launch bundled with a new ChatGPT subscription tier -- possibly free hardware with a 24-month commitment, possibly a higher-priced tier ($30-$40/month) that includes the device on a refresh cycle. Carrier-style economics, not Apple-style economics.

4. Apple will respond by deepening ChatGPT integration, not by killing it. Apple's interest is in being the platform on which all AI runs, not in picking a winner. The 2026-2027 Apple Intelligence updates will likely include even tighter ChatGPT integration alongside Gemini and possibly Claude. The Ive device's strategic threat is real, but Apple's response is more likely to be cooption than confrontation. If Apple ever removes ChatGPT from Siri's default options, that is the moment the OpenAI device thesis gets validated.

5. Anthropic, Google, and Meta will all have answers within 18 months of launch. If the Ive device achieves any breakout traction, expect Anthropic's hardware partner device to materialize publicly, Google to push a Pixel-adjacent companion device, and Meta to accelerate a non-glasses ambient form factor. The category, if it works, will not stay OpenAI's alone. If it does not work, OpenAI will be the only one holding inventory.

What to watch for, concretely: supply-chain whispers about engineering validation builds at Foxconn and Luxshare; FCC filings (which become public 90-120 days before launch); developer documentation drops; OpenAI carrier partnership announcements; and the cadence and tone of Apple Intelligence updates through the second half of 2026.

The Real Question

The Ive device is not really about the device. Almost nothing important about it depends on whether the camera is in the right place or the projected interface works or the on-device wake word fires reliably. Those are engineering problems. They have engineering answers, and they are being worked on by the people who solved the same problems at Apple between 2007 and 2020.

The real question is whether the AI-native consumer device is a category. If it is -- if there is a real product space between the smartphone and the smart speaker, occupied by an always-on ambient companion that consumers want -- then OpenAI's bet is correctly sized and well-staffed and likely directionally right. If it is not -- if ambient AI is best delivered as a feature inside the smartphone people already own -- then OpenAI has spent $6.5 billion in equity and a few years of executive attention to ship a beautifully designed object that ends up filed next to Humane and Rabbit in the museum of AI hardware that the market refused.

Altman has said, repeatedly and on the record, that he thinks the device will be the most successful consumer hardware launch in history. That is not a forecast. It is a forcing function. Inside OpenAI, the message is that distribution is the company's single largest strategic risk and the device is the answer. Inside Apple, the message is that ChatGPT is a feature of iOS. Both companies cannot be right.

What Jony Ive built at Apple was a product family that made everyone forget there was ever an alternative. What he is being asked to build now is a single object that makes consumers forget they ever needed to pull out their phone. The first project required 13 years and a category-defining touchscreen. The second project has about 18 months and an AI model whose marginal value over a phone-based ChatGPT app is, today, narrower than the marketing suggests.

The bet is the right one for OpenAI to make. That does not mean it works. The history of hardware is full of products that were strategically correct and commercially fatal. The Ive device might end the iPhone's monopoly on consumer computing. It might also be the most expensive proof yet that the phone won, and we just didn't notice when it did.