The Definitive Guide to Wearables: How Watches & Rings Disrupt AI Product Strategy

Wearables will unlock the power of LLMs from the terminal window. As frontier models push the always-listening use case, the most exciting startups are envisioning a world where LLMs silently power new capabilities that will change how we live and work. Your product's context window stops being a chat history and becomes a life.

Imagine your bra and your AirPods becoming essential to measuring your health and making you productive. Imagine wearables that measure your glucose, dopamine, and other hormones that could unlock individualized health baselines no one in history has had before except the ultra wealthy.

Wearables are defined by much more than watches now. The category Apple and Fitbit defined is being redrawn by Oura, Garmin, Plaud, Meta and OpenAI, because something you wear can combine form factor and utility in ways a phone in your pocket and a speaker on your counter never could. Alexa only ever knew what you said to it. A device on your body knows where you are, what you are looking at, and what your heart is doing while you ask. Sales are climbing and will accelerate as more businesses work out what wearables plus AI actually unlock. The health obsession driving it arrives with influencer hype that gets overbearing, and the science underneath it is real and getting stronger every quarter. For founders, designers and product managers the prize is two things a chat window has never had: individualized data from a customer's real life, and a way to turn an abstract LLM prompt into an action in the physical world.

Wearables are the unlock that AI product need to become valuable

Half of American adults have used an AI chatbot, and Pew found 42% use it to search for information, the most common thing anyone does with it. We answer that with benchmarks. Nobody outside this industry cares that a model gained four points on a reasoning eval.

Wearables are where AI starts saving lives. One of my best friends died of throat cancer at 44, months after he said he felt off and was told he was too young for it to be serious. His data would have held a voice changing too slowly to hear, months of degrading sleep, and a resting heart rate drifting upward. He wore a device that recorded all three and interpreted none of them.

Every large company wants the same thing, which is always-on data. Amazon put a microphone on your counter, television makers watch your screen, Meta put cameras on your face, and startups put pendants around necks.

All of it is a liability. The FTC fined Vizio $2.2 million for capturing second-by-second viewing from 11 million televisions, and Texas later sued Samsung, Sony, LG, TCL and Hisense over the same thing. A television that listens and watches is one somebody else can break into. Glasses are worse, because the exposure lands on people who never bought anything: the ACLU and 75 organizations asked Meta to drop facial recognition from its frames, and New York banned them from its 1,240 courthouses.

A wearable is the same data somewhere safer and more personal. It belongs to the person wearing it, who chose to put it on. They can benchmark their health against their own history, an employer can protect a workforce in conditions that kill people, and a family can watch over a relative without a camera in the living room.

The irony is that form factor, not AI, may be the real design unlock. A bra worn for thirty minutes beats an appointment nobody books, and a smart toilet beats a lab requisition on a fridge door. Expect wild attempts to put sensors wherever daily life already happens.

Types of wearables

The category splits by form factor, and the growth rates diverge hard enough that "wearable strategy" is a meaningless phrase unless you name the shape.

• Hearables — 407.6 million units in 2026, up 4.0%, the largest installed base in the category by a distance, and the cheapest way to reach a customer continuously. Unlocks: conversation that never needs a screen, live translation, and an assistant that hears the meeting you are sitting in rather than the summary you type afterwards.
• Smartwatches — 159.7 million units, down 2.8%, the incumbent and the only shape losing ground. They carry the densest sensor stack on the market, which is why clinical work lands there first. Unlocks: continuous physiology with a screen attached and a notification surface people already tolerate. The unit decline reflects where the growth went rather than any drop in what the platform can do.
• Smart glasses — 13.6 million units, up 41.4%, the fastest-growing shape in the category. They are the only wearable that sees what your customer sees. Unlocks: visual context a model can reason about, and a hands-free output channel for what it decides. The dividing line is camera versus camera-free, and Meta ships cameras while RayNeo iO and Memo Mind deliberately omit them to sell into rooms where cameras are banned. Samsung's Intelligent Eyewear splits the computing with your phone so the assistant can stay on all day, with the stated goal of you reaching for the phone less.
• Smart rings — 4.9 million units, roughly 3% of smartwatch volume, led by Oura and Ultrahuman. People take watches off at night and leave rings on. Unlocks: the highest-compliance physiological baseline available, which is the difference between a dataset with holes in it and one that can support a personal norm.
• Neural input bands — the newest shape and the one nobody has priced. Mudra Band and Mudra Link read the nerve signals running to your fingers, so a barely perceptible thumb movement becomes a command. Unlocks: the missing half of every screenless device on this page. Glasses and earbuds solved output, and neural input is the first answer to how a person replies in a meeting without talking or waving.
• Patches and textiles — no screen, no interface, and almost no competition, running from chest-worn Doppler to circuits woven directly into fabric. Unlocks: hospital-grade measurement outside a hospital, which is also the only shape with a credible path to insurance reimbursement.
• Exoskeletons and actuated wearables — the only category that acts on the body instead of reading it. Hypershell's Halo predicts movement 200 milliseconds ahead; Moonwalkers adjusts motorised gait support to terrain and fatigue. Unlocks: an output channel for AI in the physical world, where a model's decision becomes force applied to a limb rather than a notification.

Category revenue runs $99.36 billion in 2025 to $118.89 billion in 2026 across roughly 626 million units. Revenue is growing faster than units, which tells you the mix is shifting toward sensor-dense devices at higher prices.

The uncomfortable part: if your roadmap says "wearable" and means "watch app," you have anchored to the only form factor losing units, and handed the two growing surfaces, the ones that capture what your customer sees and hears, to whoever shows up first.

How LLMs and AI make this data more personalized and useful than ever

Hardware differentiation is closing, because every serious vendor now has adequate PPG, accelerometry and temperature sensing. The entire contest moved to what happens after capture. Four shifts matter:

• Better hardware stopped being the answer. RingConn cut its heart-rate error by roughly four times without changing a single component, purely by reading two signals against each other more intelligently and training on the ring itself so nobody's data leaves their hand. Accuracy used to be an eighteen-month hardware cycle you waited on. It is now something your software team can ship against this quarter, on devices already strapped to your customers.
• Imagine real benchmarks based on your life, so your wearable spots issues before your doctor can. Your doctor compares you against a population average, measured once a year, in a room that makes your blood pressure wrong. A device that has watched you sleep for four hundred nights knows what your normal actually is, and a deviation from that is the only signal that has ever earned the right to interrupt you. Apple's Series 12 now samples heart rate every five seconds instead of every five minutes, a sixtyfold increase that exists to feed that baseline rather than to draw a prettier chart.
• On-device inference buys trust and latency together. Samsung's xMAE and HiMAE run biosignal analysis in under a millisecond locally. Apple's foundation models, Honor's local cardiac detection, and Memo Mind's fully offline glasses make the same bet: the intimate data never leaves the body. Round-tripping a continuous stream to a cloud is unaffordable in battery and in confidence.
• The LLM layer turns signal into something a person will act on. Abbott wired Lingo glucose data into Google's Health Coach AI. OpenAI Health pulls Apple Health and medical records into ChatGPT for plain-language summaries. Apple's Health app reads uploaded labs alongside years of wearable history.

The world-model argument stops being abstract here. Language models learned from text because text was the corpus lying around. World models need grounded, embodied, continuous observation of physical reality — and wearables are the only consumer device collecting it at population scale, from the first-person perspective, with physiological state attached.

TwinDEX from the intro is the proof. Nothing in a text corpus tells a model what it feels like to pick up a mug, and no amount of scraping the internet will produce it. A person wearing a capture rig produces it continuously, for free, while going about their day.

Strategic implication: your defensibility is the longitudinal context you accumulate and the judgment layer built on top. Neither transfers when a customer switches devices, which makes it the rare moat that compounds daily.

Exciting startups and use cases

The capital has concentrated hard. WHOOP raised $575 million at a $10.1 billion valuation in March 2026 on 2.5 million members and a $1.1 billion run rate. Oura closed more than $900 million at an $11 billion valuation. Plaud expects $500 million in sales this year. Nine deals above $50 million, including VITURE, XREAL, Sesame and RayNeo, account for most of the money in the sector.

Then there is the data point nobody puts in the pitch deck. Limitless, the most direct AI-pendant competitor, was acquired by Meta in late 2025 and stopped selling devices that December. Standalone ambient capture is a feature platforms buy, not a category that stays independent.

The use cases worth studying are clinical, unglamorous, and each one replaces a procedure rather than counting steps:

• Flosonics FloPatch — an FDA-cleared wearable Doppler with FloPredict AI, built with BARDA, recognising the patterns that precede deterioration from internal bleeding in trauma settings.
• Johns Hopkins MOSAIC — reconstructing continuous blood-pressure waveforms from non-invasive signals, aimed squarely at replacing invasive arterial catheters in the ICU.
• Wearable ultrasound (Shanghai) — 91% accuracy on elevated central venous pressure, another catheter replaced.
• PHI-BRA — a smart bra pairing thermography with electrical impedance spectroscopy and an ML layer, worn 30 minutes, detecting breast cancer between mammograms.
• SEERS mobiCARE — the first Health Canada-approved AI-powered ECG wearable, validated across 1,100 institutions and 770,000 tests.
• ZenoWell Luna Plus — closes the loop from tracking to intervention with vagus nerve stimulation instead of stopping at a recommendation.
• Omi's Parent Whisperer — an AI clone of a parent's voice, trained on their communication history, comforting their child while they travel. Technically remarkable, ethically unresolved, and shipping.

The competitive read: every company on that list is replacing a procedure, a catheter, or a clinician's judgment call, and none of them is fighting for space on your customer's wrist. If your competitive set is other apps, you are watching the wrong market. The incumbent being disrupted here is the health system, and it does not yet know it is in a fight.

How wearables are transforming long standing businesses

Wearables are restructuring industries with no obvious connection to fitness.

• Hospitals are pushing monitoring past their own walls. Respiree's clinical-grade cardio-respiratory surveillance, integrated with Philips, follows patients from ward to home with predictive analytics for early deterioration. Cedars-Sinai published an AI clinical decision support framework for hospital integration. The economics are readmission penalties, and continuous monitoring is the cheapest way to avoid them.
• Pharma finally has an adherence measurement layer. DataMeds' patents cover biometric confirmation that a drug was taken, correlated against real-time physiological response. Adherence has been self-reported and unreliable for decades.
• Defence and industrial safety moved fastest. Smack Technologies' Alpha pushes fused battlefield intelligence to soldiers; confined-space systems now predict asphyxiation before the exposure.
• Robotics turned wearables into a training-data pipeline, which is the TwinDEX story above and a genuine category inversion.
• Industries that never considered themselves sensor businesses are about to be. Dyson put real-time camera AI in a toothbrush, which turns a consumer product into a twice-daily diagnostic for gum disease and a data relationship with a dentist. The same move is available to anyone whose product already touches a customer's body on a schedule: a mattress that reports sleep apnoea, a car seat that notices arrhythmia on the commute, a razor that flags a mole that changed. Veterinary medicine is running the same playbook a decade early, because animals cannot describe symptoms and continuous monitoring is the only diagnostic they have ever had.

Those are the disruptions with obvious buyers. The ones nobody has priced yet all follow from a single fact: a wearable produces a continuous, timestamped record of a person's body and their attention, and almost every industry has been guessing at one or both for its entire existence.

• The smart speaker and the phone. Alexa lost because it sat on a counter and knew only what was said to it. A device on your body knows where you are, what you are looking at, and how your body is reacting while you ask. The assistant category gets rebuilt around whichever surface has that context, and the phone slowly becomes the thing that charges it.
• Advertising and attention measurement. Glasses know what a person actually looked at and for how long. Advertising has wanted that measurement for a century and has never once had it, and Meta has already patented facial recognition with expression analysis for its frames.
• Clinical trials. Continuous endpoints collected at home replace periodic site visits. Trials get shorter and cheaper, which changes which drugs are economical to test at all, and small-population conditions that never justified a trial budget start to pencil out.
• Underwriting and employment. A continuous physiological record is the most predictive individual risk data ever collected. Life insurers, lenders and employers will eventually ask for it, and whether your device makes that export possible is a decision your team makes at the schema level long before anyone in legal hears about it.
• Evidence and liability. Wearable data already turns up in criminal and civil cases. A device logging heart rate, location and movement is a witness your customer cannot cross-examine, which turns your data retention policy into a legal exposure with a dollar value.
• Training any skilled physical trade. TwinDEX proved a human wearing a capture rig can teach a robot. The same rig can teach a person. Surgery, welding, physiotherapy and machine operation all currently depend on an expert standing next to a novice, and that constraint is what caps how fast expertise spreads.

The largest restructuring is in insurance and payers, and it comes with the trap most product teams have not priced.

On January 6, 2026, the FDA substantially narrowed oversight of AI-enabled software and consumer wearables. The general wellness exemption widened to cover non-invasive devices reporting blood pressure, oxygen saturation and glucose-related signals, provided they are marketed for wellness. Less regulatory burden reads like a gift.

It is a revenue ceiling. The money in health wearables is reimbursement: CMS pays $50 or more per 30-day remote patient monitoring period, and providers running RPM programmes bill $150 to $300 per patient per month across setup, data collection and care management. Medicare's ten-year ACCESS pilot extends it further. Those codes require clinical validation and regulated status, and a device marketed strictly for wellness is ineligible by construction.

So the exemption is a fork. Stay in wellness and sell hardware and subscriptions to consumers who churn. Take the regulated path, carry the trials and quality systems, and reach a payer who does not.

Strategic implication: pick your side of that line before anyone writes positioning copy, because your marketing language determines your regulatory class, and your regulatory class determines whether a payer can ever pay you.

Research papers that point to the future of wearables

The literature is moving faster than any hardware cycle, and it has split into three clear years of argument.

What 2025 established: sensor data is a corpus, not a readout.

• Foundation models built on behavioural data from wearables (July 2025) beat models trained on the raw sensor streams. How someone moves through a week carries more health signal than any individual reading does.
• AnyPPG (November 2025) trains an ECG-guided PPG foundation model on more than 100,000 hours of recordings, so the cheap optical sensor in every ring and watch inherits what the expensive clinical one knows.
• Benchmarking Egocentric Multimodal Goal Inference for Assistive Wearable Agents (October 2025) was the first serious attempt to measure whether a wearable can work out what you are trying to do before you tell it. That benchmark is the scoreboard your proactive feature will eventually be judged against.
• npj Digital Medicine (2025) reviewed AI plus wearables across diabetes and prediabetes and found the combination performing across prediction, classification and clinical decision support rather than in one narrow task.

What 2026 proved: scale works, and the body is a training set.

• SensorFM (arXiv:2605.22759) is the number that should reset your planning. Google pretrained on over one trillion minutes of sensor data from more than five million Fitbit and Pixel Watch users across 100+ countries, and the resulting model transfers to 35 health prediction tasks and grounds a personal health agent. The per-condition classifier your team is scoping has already been made obsolete by a general physiological model someone else paid to train.
• EgoWAM (CoRL 2026) settles the world-model argument. Training a robot on in-the-wild egocentric human data, while forcing the model to predict how the scene changes rather than just copy the action, improved out-of-distribution generalization by up to 4x. Footage from a person wearing a camera is better robot training data than the robot's own attempts.
• 11 million days of longitudinal wearable data (January 2026) surfaced future health insights that no cross-sectional study design could have reached, because the finding lives in the duration rather than the measurement.
• Will wearable technologies transform clinical trials? (Nature, 2026) documents the pharmaceutical industry working out that continuous home endpoints can replace site visits.
• Wearable AI in the Era of Large Sensor Models is the best single survey of where all of this converges, and it names the category: large sensor models, treated the way text was treated in 2020.

What is sitting in preprint for 2027: none of it has cleared peer review, and all of it will shape next year's roadmaps.

• Wearable Foundation Models Should Go Beyond Static Encoders argues the current generation is built wrong, because a model that encodes a fixed window cannot reason about a person changing over months. The fix is the same architectural shift that took language models from embeddings to agents.
• OpenMHC and Inertia-1 are both attempts to open-source wearable foundation models and the data to evaluate them. If either lands, the proprietary sensor model stops being a moat sometime in 2027.
• Sonata builds a hybrid world model for inertial kinematics specifically to work around clinical data scarcity, which is the constraint that has blocked every serious medical wearable from shipping on schedule.
• Researchers at Pusan National University put neuromorphic compute inside stretchable material, so a patch senses a physiological change and responds locally with no chip and no cloud. The end state is a bandage that adjusts its own compression.

Put those three years next to each other and the shape of the next product generation is visible. The model you fine-tune, the behaviour you capture and the compute you push into the material all converge on a device that never asks the user anything. A few things that become buildable on the other side of that:

• A condition detected from how someone moved through their week, with no test, no clinic and no symptom reported.
• A garment that treats what it measures, closing the loop between sensing and intervention with nothing in between.
• A physiological model that transfers across manufacturers, turning today's ecosystem lock-in into a commodity and putting the value back into judgment.
• Population health monitored continuously rather than surveyed annually, which is either the best public health instrument ever built or the most complete surveillance infrastructure ever deployed, depending on decisions your industry is making this year.

What Product Leaders Need to Win at Wearables

The sensors are solved and the models are commoditising. What stays scarce is the judgment to decide what a device should notice, when it should speak, and what it should refuse to do. Here is the playbook, in the order the decisions actually have to be made.

1. Choose your regulatory class before anyone writes a line of positioning copy. Wellness or regulated, and you only get one. Your marketing language determines your class, your class determines whether a payer can ever pay you, and reversing the choice later means running the trials you skipped. Write the decision down and circulate it to marketing, because they will break it by accident inside a quarter.
2. Pick the form factor from the job, not from what your team already ships. Ears for anything conversational, glasses for anything that needs to see what the customer sees, ring for a baseline you need people to actually maintain, patch for anything clinical, wrist only when you genuinely need the screen. Shipping a watch app because you have iOS engineers is a staffing decision wearing a strategy costume.
3. Instrument for a baseline in your first release, not your third. Value lives in deviation from a personal norm, and a norm needs months of continuous data before it is worth anything to anyone. The sampling rate and retention schema you set at launch decide what you are allowed to build in year two, and they are the hardest things to change retroactively. Build that pipeline before you build the feature that needs it.
4. Make wear-time a tracked product metric with a named owner. A ring worn 95% of nights beats a watch with better sensors worn 60% of nights, every time, and almost nobody's dashboard shows this number. Compliance is your real retention metric, and it is also your accuracy metric, because a model cannot learn a norm from a device sitting in a drawer.
5. Run inference on the device wherever the data is intimate. It is the only privacy claim a person believes about something strapped to their body, it removes the cloud cost of a continuous stream, and it is a differentiator you can put in the marketing. Budget the engineering early, because retrofitting on-device inference is a rewrite rather than a sprint.
6. Assume the general physiological model wins, and decide now what you own when it does. SensorFM-scale pretraining means the per-metric classifier your team is building is a temporary advantage with a visible expiry date. What survives is the longitudinal record of one specific customer and the judgment layer sitting on top of it, and neither of those transfers when they switch devices.
7. Write down what you will not do with the data before someone asks you to do it. Meta patented facial recognition with expression analysis for its glasses. Omi built an AI clone of a parent's voice to comfort their child. Both are impressive, and the decision about whether either should exist will be made by teams like yours rather than by a regulator that has just stepped back. That answer is worth far more written down now than improvised later under commercial pressure.

Wearables give a model something it has never had: continuous, embodied evidence of how one specific person moves through the world. A product built on that stream knows a customer instead of a segment, and the same stream is what teaches machines what physical reality is. Most teams will treat it as a sensor spec and ship another step counter into the one shrinking form factor. The relationship goes to whoever treats it as the context layer under everything else they build, because that context does not transfer when a customer buys a different device.

Working that out with product teams is what we do at PH1 Research. If you are deciding where wearables fit in your AI product strategy, the hard part was never the hardware.