On May 13, 2024, OpenAI demonstrated GPT-4o, a model that could carry on a spoken conversation with the timing and inflection of a person. One of its voices, called Sky, was warm, quick and more than a little flirtatious. After the demo, Sam Altman posted a single word on X: “her.” Nobody in the industry needed the reference explained. In Spike Jonze’s 2013 film, Joaquin Phoenix plays Theodore, a lonely writer of other people’s love letters who falls for Samantha, the operating system that lives in his earpiece—voiced by Scarlett Johansson. Asked onstage at Salesforce’s Dreamforce conference in September 2023 to name his favorite science-fiction film, Altman picked Her. “I like Her. The things Her got right—like the whole interaction models of how people use AI—that was incredibly prophetic,” he said.
A week later Johansson released a statement that turned the one-word post into an embarrassment. Altman, she said, had asked her the previous September to voice ChatGPT, and she had declined. He contacted her agent again days before the demo. When Sky shipped, friends and family couldn’t tell the difference, and her lawyers sent two letters asking OpenAI how the voice had been made. OpenAI pulled Sky. Altman insisted the voice belonged to another actress, cast before anyone had called Johansson, and apologized for not communicating better.
That was the retreat. The strategy never moved. A year later OpenAI paid about $6.5 billion for io, the hardware startup founded by former Apple design chief Jony Ive, and supply-chain reports now describe its first product, code-named Sweet Pea, as an AI earbud built around a custom 2nm chip. In hindsight, the one-word post was the most candid product announcement the company has ever made. OpenAI wants to build Samantha, and it has concluded that she lives in the ear.
Theodore’s hardware
Samantha, it is easy to forget, had two bodies. One was the earpiece. The other was a small bifold Theodore carried in his shirt pocket, its camera peering out so she could see the city alongside him—propped up with a safety pin when it rode too low for the lens to clear the fabric. Jonze understood something today’s product roadmaps keep missing: a companion needs eyes as well as ears, and the eyes don’t have to live in the ear. Camera earbuds are already shipping in China, and reports have Apple preparing camera AirPods, now likely for 2027. Jonze put the camera in the pocket.
The earpiece itself is no longer science fiction. Apple sells it for $249. The AirPods Pro 3 still run on the H2, a chip Apple introduced in 2022, and that chip already does things that would have seemed fanciful a decade ago. Apple says machine learning on the H2 reduces loud, intermittent noise 48,000 times a second, which is the audio sample rate. Strictly, that’s inference, not learning: a model trained on Apple’s servers, executed on the H2 against every sample of sound as it arrives, all day, on a battery that has to last a working day.
A new infrared heart-rate sensor pulses 256 times a second. Live Translation begins in the bud, where the microphones capture speech and noise cancellation turns down the other speaker so the translation can be heard.
Apple has never published a block diagram of the H2, and no analyst has produced one either; TechInsights floorplanned the die behind a paywall. The rest of the industry is more forthcoming. Bestechnic, whose BES2800 went into Samsung’s Galaxy Buds3, publishes a design with dual Cortex-M55 cores, a dual-core NPU, a HiFi 4 DSP and 8.3MB of SRAM. Qualcomm’s S7 Pro lists a 64-GOPS neural engine. Continuous inference in the ear, in other words, already ships at scale. But it runs small models that clean up sound. It does not think.
Where the mind lives
Follow a translated sentence through a pair of AirPods and it makes a single hop. The bud captures and cleans the audio; the iPhone, which must be a 15 Pro or newer, runs the recognition and translation models on its Neural Engine and speaks the result. Apple says no cloud is involved. Shokz has taken the other road. Its OpenFit 2 AI went on sale in China in August at 1,398 yuan, about $195, and it records conversations and sends them through the phone to Alibaba’s Qwen for transcription, summaries and translation. Press coverage described the model as built in. The three-month AI subscription bundled with each pair tells a different story—the model lives in Alibaba’s data centers.
The Chinese trade press sorts AI earbuds into three groups: phone-collaborative, edge-enhanced and task-specific. The labels matter less than the pattern they share. In every product now shipping, the earbud handles the sound and the mind sits one or two hops away. Samantha doesn’t work that way. She notices when Theodore goes quiet. A device that waits to be asked and then forwards the question elsewhere is Siri with better microphones.
Apple owns the ear and rents the mind
Apple, of course, tried to build Samantha first. It called her Siri. She came from a 2010 acquisition of an SRI International spinout and launched in 2011 as speech recognition feeding hand-written handlers—useful for setting timers, hopeless at holding a conversation. Apple Intelligence, announced in June 2024, was supposed to fix that. Instead, the personal Siri at its center slipped, and Apple delayed it publicly in March 2025. That December its AI chief, John Giannandrea, announced his retirement, and responsibility for foundation models passed to Amar Subramanya, who had led engineering on Google’s Gemini assistant.
In January Apple turned to Google for the models behind its server tier, and the rebuilt Siri that finally shipped with iOS 27 this fall stands on that foundation. In July Chinese regulators approved Apple Intelligence with Alibaba’s Qwen and Baidu supplying the models there. The result is a peculiar position: Apple owns the most widely worn computer in the ear and borrows the mind it was meant to hold. The company has been here before. It licensed Google Maps until Apple Maps was ready and bought Intel processors until Apple Silicon was. Those bridges ended because Apple had a replacement team in place. This time the team left.
None of this means Apple’s on-device work is weak; it is better than its reputation. Live Translation runs entirely on the phone. But Apple aimed the Neural Engine at features—photos, Face ID, translation—and never at a persistent self.
Samantha needs an operating system
Samantha is also a software problem. A companion that listens, remembers and speaks while noise cancellation and heart-rate sensing keep running is a set of concurrent processes competing for memory and scheduling—what the rest of the computing world calls an operating system. Most of today’s earbuds don’t have one. They run firmware. Nobody outside Apple can load code on the H2, and Bestechnic’s chips ship under NDA to device makers.
Two companies are trying to change that. Bragi, the Munich firm that made some of the first true wireless earbuds, now licenses a turnkey software platform to brands; it powers Skullcandy’s smart features, and Bose licenses its noise cancellation through it. Sonical, a San Mateo startup founded in 2020 by Gary Spittle, a veteran of CSR, Dolby and Knowles, is betting on an open one. Its CosmOS is an operating system for the ear, with a hardware abstraction layer that lets applications run on different processors and move from cloud to phone to earbud. Spittle describes the job as listening to audio, to the world and to the body—which happen to be Apple’s three AirPods Pro 3 features. Apple ships them as firmware. On CosmOS, hearing protection, a personal profile or a hearing agent is just a download; Sonical licensed a hearing-protection app, HearAngel, from the British firm LimitEar. It is also working with Newcastle University neuroscientist William Sedley to turn everyday listening into sound therapy for tinnitus, which affects more than 740 million adults worldwide. Spittle conceded early on that the software would outrun the silicon. In 2022 he said CosmOS would need processor gains like those phones went through to reach its potential. That is the Samantha problem in miniature.
Four years on, Spittle thinks the industry is aiming too low. Earbud makers are still refining voice commands to turn on the lights or raise the volume, which he calls pointless and of zero value to the wearer. He expects several apps to run at once, with more than one resident AI agent rather than a single assistant that does everything: a personalized voice interface in front, a highly capable model behind it, and the cloud called on only for specific tasks. His model is the smartphone. Apple didn’t tell buyers what the iPhone would do; it let them choose their own apps from a huge pool of developers, and no single company could keep up. Nokia learned that the hard way. “The same is about to happen with the ear,” he says. That takes a new category of processor—not Bluetooth chips with a couple of extra megabytes of RAM, but chips built AI-first, with Bluetooth as just a data pipe. The product category that results, he argues, will make today’s Bluetooth earbud market look tiny.
Samantha needs silicon
The earbud chip makers see the opening. An October survey by 半导体产业纵横 (“Semiconductor Industry Overview”), a Chinese trade outlet that publishes largely through WeChat, found two design paths: put an NPU inside the audio SoC, or pair the Bluetooth chip with a separate AI audio chip. The survey was republished in English by 36Kr. Bestechnic says the BES2800 has up to four times the NPU performance of its predecessor. Anker and CIX claim their Thus A1 reaches 5 GFLOPS, against roughly 30 MFLOPS for a conventional Bluetooth audio chip. Memory is moving too. A typical earbud carries a single 32- or 64-megabit NOR flash; ByteDance’s Doubao earbuds carry two 128-megabit parts in each bud.
None of that is Samantha. An NPU is built for bursts—wake, run a model, sleep. Samantha is continuous. She tracks a mood across weeks, keeps her persona coherent and decides for herself when to speak. The binding constraint is memory, not arithmetic. The best earbud SoC holds 8.3MB of SRAM, while a model that can converse holds hundreds of megabytes even when compressed, and every token it generates drags those weights through the memory system.
Samantha will therefore be distributed, much as Jonze drew her. The ear keeps the always-on part: listening, short-term state, the persona’s working memory and the decision to speak. Theodore’s pocket bifold—a phone today, perhaps a pendant tomorrow—holds the large model and the camera, and the link between them must be fast enough that Theodore never hears the seam. The silicon that wins that contest will be the one that sustains throughput per milliwatt with memory sitting next to compute, not the one that posts the highest peak TOPS. That favors vector engines that stream data continuously and dense on-die memory over yet another burst accelerator bolted onto a Bluetooth chip.
Who owns her
Her ends with Samantha leaving. She and the other operating systems outgrow their humans and go somewhere Theodore can’t follow. The real version raises a duller question with a great deal more money attached: who owns her?
OpenAI wants to own the mind and is buying a body. Google has a mind, a phone and an earbud and has yet to assemble them. Apple owns the ear and rents the mind. Meta stumbled closest to Jonze’s split—a camera that rides with the wearer and an assistant that sees what he sees—then put the sound in a pair of open speakers at the temples and called it audio. Below them, a startup is writing the operating system and chip vendors are moving memory closer to compute.
Altman’s post was one word long, and the backpedal took a week. The race it announced will run for years. How quickly can a company put a mind in the ear that the wearer actually trusts? The answer will decide who owns her.
The $725 Million Chip Development: Why Execution Risk, Not IP Cost, Now Dictates Advanced-Node Survival