If you squint at the future OpenAI and Jony Ive are sketching, it looks like a tiny, polite, pocketable computer that watches and listens just enough to be helpful — but not so much that it becomes creepy. The reality, according to people briefed on the project, is messier: the team still hasn’t agreed on what the device should be like as a companion, how often it should interrupt you, or how to afford the enormous computing power it would need to feel genuinely useful. Those are not small details when you’re trying to make a new class of personal hardware.

Reporters who’ve spoken to sources inside the project say OpenAI and Ive are building a screenless, palm-sized device — roughly the size of a phone — that you can carry or set on a table. It relies on microphones, speakers and at least one camera to sense the world and respond, and is being designed to act like “a friend who’s a computer who isn’t your weird AI girlfriend,” in the memorable phrasing used by one source. The hardware is described as the first in a family of devices aimed at being a new, ambient touchpoint for AI.

OpenAI’s push into hardware accelerated after it acquired Jony Ive’s device outfit — commonly reported as the io/LoveFrom team — in a high-profile deal earlier this year, bringing Ive and a cadre of former Apple engineers into OpenAI’s orbit. That acquisition set the expectation inside and outside the company that OpenAI would make something physical to sit alongside phones and PCs. The timeline people are talking about suggests a consumer product aimed for late 2026 or early 2027, though sources warn that it could slip.

The hard, unglamorous problems

Designing a new product is always partly about trade-offs; designing a new class of product — one meant to live in your life and occasionally speak up — forces an argument about personality, privacy and raw infrastructure.

Personality and timing. Engineers and designers are locked in something of a philosophical debate: if the device is too talkative, it will annoy people; if it’s too quiet, it will feel useless. Teaching a model when to interrupt, how long to hold a conversation and how to bow out politely is actually a social-science problem as much as an engineering one — and it’s proving thorny. The team is testing different interaction models (wake-word vs. “always listening” contextual engagement) but hasn’t settled on one.

“Always on” vs. privacy. Several sources say OpenAI is exploring an “always on” approach that continuously gathers ambient clues so the device can act contextually — not unlike some of the ambitions behind wearables and smart speakers. That raises obvious privacy and regulatory questions: what data stays local, what gets sent to servers, how long is a user’s “memory” kept, and who audits that behavior? Those details matter because consumer trust collapses fast when a device seems to be listening or watching without a clear benefit.

Compute and cost. There’s a blunt, financial bottleneck here: OpenAI’s large language and multimodal models are compute-hungry. Big tech incumbents like Amazon and Google already run the huge server farms that make always-on assistants cheap and reliable; OpenAI has the models, but — sources told the Financial Times — not the same scale of dedicated infrastructure for a consumer product yet. That imbalance could force compromises in latency, capability or price.

Why hardware is harder than software

You can ship an app update next week if something goes wrong. You can’t push a physical recall if users hate the way a device talks to them. The history of recent AI gadgets offers a cautionary tale: Humane’s AI Pin launched with ambition but stumbled on latency, UX and sales; other early entrants like Rabbit’s R1 launched rough and required months of software fixes to feel useful. Those rollouts underline a basic truth: great industrial design and a glinting prototype aren’t enough if the underlying AI, connectivity and economics don’t line up. OpenAI’s device — intentionally screenless and socially aware — needs all three to work at once.

The product choices that will determine success

Below are the kinds of decisions that will shape whether OpenAI’s gadget becomes helpful or just another gadget people return:

• Local vs cloud processing. More local inference reduces latency and privacy risk — but increases device cost, battery draw and complexity.
• Interaction model. Wake word (like “Hey Siri”), button activation, or contextual, always-on prompts? Each has trade-offs in discoverability and intrusiveness.
• Personality design. How human should a voice be? How much small talk is useful versus distracting? Designers worry about the line between friendly and anthropomorphic.
• Manufacturing and supply chain. Sources named partners and assembly options being explored; final choices affect price and geopolitical risk.

So — will it talk to you like a friend?

That’s the aspiration: an assistant that understands context, remembers useful things, and interjects in a helpful way. But the more helpful a device tries to be, the more it must observe. That observation creates privacy pressure, which in turn forces engineering, policy and product compromises. The FT’s reporting suggests OpenAI is painfully aware of that circle — and is treating the personality question as central to the product’s trustworthiness.

The calendar: optimistic, cautious, or delayed?

The public timeline floating around — late 2026 to early 2027 — looks optimistic to industry watchers, given the open questions about compute, privacy safeguards and the subtleties of conversation design. If OpenAI needs to build more bespoke infrastructure or radically change the model to run faster and cheaper, the launch could slip. Historically, the devices that win are not the ones with the best demo alone; they are the ones that show up reliable, useful and priced so that ordinary people will tolerate small oddities in exchange for net benefit.

Why this matters beyond a single gadget

If OpenAI gets this right, a successful, trusted ambient AI device could redraw where AI lives in people’s routines: not in phones or on screens but in the room with you. That has big implications — for privacy laws, for which companies control what we see and hear, and for the competitive landscape between cloud giants and newer AI players. If OpenAI earns that access to daily life, it could remake how we think about personal computing; if it fails, the result will probably be another cautionary chapter in AI hardware trials.