The ear as a sensing site

The one-line read

The ear is genuinely the best non-invasive site on the body for three things — EEG, core temperature, and cephalic (head-directed) blood flow — and merely an alternative site for everything else, where the wrist and finger are already good enough and already own the user. Every independent consumer in-ear sensing company to date has died, sold its hardware, been acqui-hired, or been shut down by its own parent. Zero survivors. The survivors are component licensors and regulated-medical plays.

What the ear is uniquely good for

Use this as the first screen. If a pitch’s modality is not on this list, the ear is a form-factor choice, not an advantage.

• Ear-EEG. The only discreet, wear-all-day EEG site. The wrist cannot measure EEG at all. In-ear recordings showed >98% validity against 46–78% for frontal setups. Still inferior to scalp EEG, and single-channel in-ear sleep staging is 90.5% binary wake/asleep but only 65.1% four-class — the accuracy gap is the product problem.
• Cephalic blood flow. Nothing at the wrist tells you about blood flow to the head. Stat Health optically senses a shallow ear artery as a cerebral-flow proxy; published in JACC (Mar 2023), tested at Johns Hopkins, predicts fainting minutes ahead. Cleanest “only the ear can do this” claim in the map.
• Core body temperature. The tympanic membrane reflects carotid temperature. But you cannot wear a sensor on the TM — you sit at the canal entrance and a thermal gradient from ambient air corrupts it. Right in principle, contested in practice.
• Cochlear function (otoacoustic emissions). Only observable from the ear. See the graveyard: this is what nura did, and it exited as an acqui-hire.
• Ear-canal acoustic biometrics. Only site with the geometry. Also a 30-year-old idea — see below.

Everything else — HR, HRV, SpO2, blood pressure — is an alternative site, not a better one. The real ear edge for PPG is motion immunity (the canal is mechanically shielded), not accuracy.

The power inversion (matters for any silicon thesis)

Naive intuition says the amplifier is the constraint. It is not. EEG front-ends are nearly free (~200 µW for 8-channel acquisition ASICs; sub-µW/channel front-ends in 28nm). PPG is the power hog — you are pushing photons through tissue, and Apple pulses LEDs >100×/sec. The scarce resource in an earbud is the optical budget, not the compute or the amplifier. Any pitch claiming an inference-silicon power win should be asked what fraction of the earbud’s budget its block actually addresses.

(Confidence: directional. Sourced from academic ASIC figures; no shipping-product teardowns found.)

The graveyard — the load-bearing evidence

The message: zero independent consumer in-ear sensing survivors. Sample is small (~5) and pre-dates the current sensing generation, so hold at ~85% not 100%.

The structural read — Apple absorbs

Apple’s pattern is to absorb proven modalities as free features, not to compete on price:

• Hearing aid: Sept 2024 FDA De Novo, first OTC hearing-aid software. Not a hearing-aid business — a software toggle on a product it already sells 100M+ of.
• Hearing test: 5-minute pure-tone audiometry on hardware you own.
• Heart rate: Powerbeats Pro 2 (Feb 2025) → AirPods Pro 3 (Sept 2025) → iOS 26 Fitness integration.

The kill mechanism is distribution, not capability. IDC forecasts 407.6M hearable units in 2026 (+4.0%), explicitly citing biometric health monitoring as a driver. Apple gets biometric distribution free on an install base an independent cannot buy. Bose Health is the clean control experiment.

Evidence against the absorption thesis (thin, but real): Apple frames ear HR as complementary to the Watch, not a replacement. It has shipped no ear-EEG and no core temperature, with no signal it intends to. Note the shape of that counter-evidence — it says the niche is currently un-absorbed, not un-absorbable. Apple didn’t ship a hearing aid until it did.

Where the defensible positions are

1. Regulated medical — the only one with a proof point. Naox Technologies (French) received FDA 510(k) clearance in Jan 2026 for Naox Link, the first-ever in-ear EEG clearance (CEO Hugo Dinh; already in select neurology and sleep centres). This works precisely because Apple won’t go there — Apple’s regulatory posture is the lowest-risk path touching the most people. Epilepsy monitoring and Alzheimer’s-risk epileptiform detection need trials, reimbursement codes, and a neurologist channel. But this is a medical-device business, not a semiconductor one — value accrues to the clinical dataset and the clearance, not the silicon, so the winner here is whoever can fund trials and a neurologist channel. Confidence it’s durable: ~65%, one proof point and zero exits.
2. Defence / industrial. Stat Health holds a US Air Force grant. Core temp + cephalic blood flow under heat stress and G-load is a real military requirement with a buyer indifferent to consumer distribution economics. Small, non-commoditisable. One datapoint, not a market.
3. Component / IP licensing — where the bodies are. Valencell is the successful case and it’s a modest licensing business, not a venture outcome. Bragi pivoted to licensing and disappeared. Nura’s IP ended up inside Denon.
4. Consumer device: no defensible position found. Zero counter-evidence.

Ear biometrics is a 30-year-old idea in a patent minefield

• US5787187A — “Systems and methods for biometric identification using the acoustic properties of the ear canal,” Sandia, filed 1 Apr 1996, granted 1998, now expired. The core idea is public domain.
• NEC has pushed ear acoustic authentication since 2016 with a live patent family; Apple has an ear-canal biometrics filing. A pre-seed company selling an ear-print module to headphone OEMs walks into NEC plus Apple — and the OEMs it wants to sell to are the parties filing the patents.
• The “99% accuracy / like Face ID” claim does not survive contact with the literature. 99% accuracy is a meaningless metric; the operating point is what matters. Independent work reports 22% false-rejection at 0.1% false-accept — one in five legitimate unlocks fails. Face ID’s false-match rate is ~1 in 1,000,000, roughly 1000× tighter. The Face ID comparison is not defensible.
• NEC did ship — via crowdfunding, in 2020, ten years after announcing, with no OEM design win. Likely causes: reinsertion/fit variance (a physical property of the measurement, not a software bug), the FRR problem above, and no job-to-be-done (your phone already authenticated you and pairs by proximity).

Attention decoding — read this before any “we read your focus” pitch

• The established route is EEG, and it’s crowded. KU Leuven (Bertrand) reports 91.8%/81.1% at 0.1s windows; Eriksholm (Oticon/Demant) ran the EU COCOHA project and built a real-time prototype. ~15 years, EU consortium money, in-house research arms of the major OEMs — and still no shipped neuro-steered hearing aid. That is the base rate.
• The otoacoustic-emission route is contested to the point of artifact. Francis, Zhao & Guinan Jr, Frontiers in Systems Neuroscience 2018: “Auditory Attention Reduced Ear-Canal Noise in Humans by Reducing Subject Motion, Not by Medial Olivocochlear Efferent Inhibition.” Across 8 subjects the apparent attention effect was subjects holding still and breathing differently; only 2 of 8 showed an efferent signature. A 2020 replication (n=20) split 9-of-20 — a coin flip. Guinan is the senior figure in efferent physiology.
• Diligence screen: any demo where a subject concentrates and the ear-canal signal changes is exactly what Guinan 2018 predicts from postural stillness, with zero cochlear physiology. Ask whether subject motion was controlled. If the answer is vague, that is the answer.
• Framing correction: “below the noise floor” is backwards. The clinical criterion for an OAE is 3–6 dB above the noise floor, recovered by averaging over many repetitions. That averaging costs time, set by noise statistics — you cannot engineer it away with a better microphone. Fatal for a real-time interface.
• Direction has no mechanism. The efferent reflex is a cochlear gain control — a volume knob. Spatial direction is computed centrally from interaural timing/level differences. No literature suggests a per-ear gain signal carries azimuth.

Market size — the right denominator

Wrong denominators: “hearables market $55–62B” is the audio market; sensing is a feature inside it. Any slide multiplying 400M earbuds by a health ARPU.

The layers are three, and they are not the same business.

• Device layer (407.6M hearable units): the biggest number and the most captive. Apple/Samsung/Chinese ODMs own it, and proven modalities arrive as free features. Independent consumer devices here: see the graveyard.
• Medical-device layer (23.16M hearing aids + the un-absorbed clinical niches): ~$9.7B in device value, growing ~2%/yr, 92.4% held by five vertically-integrated firms. The growth is not here; the defensibility is, in the regulated corners (Naox) those five and Apple don’t enter.
• Silicon / IP layer: 23.16M units/year across ~5 buyers. At a plausible $5–20 sensing/SoC content, the entire addressable chip layer is ~$115M–$460M/year, contested, against buyers who design in-house (Sonova, Demant, Starkey) and who bought from an incumbent (Starkey→Valencell) the one time they didn’t. The device denominator is 17× larger and inaccessible — nobody is selling a sensing ASIC into AirPods.

The size conclusion, vehicle-agnostic: the ear is a hundreds-of-millions opportunity at the component layer and a single-digit-billions, low-growth, highly concentrated one at the device layer. Neither shape supports a venture-scale independent outcome on current evidence. The asymmetric money here, if any, is in regulated-medical clearances and in the listed incumbents’ exposure to OTC commoditisation — not in the chip.

⚠️ The $5–20 content figure is an estimate, not a sourced number — no independent hearing-aid BOM teardown found. Treat $115M–$460M as an order-of-magnitude check that the answer is “hundreds of millions, not billions.” That conclusion survives a 3× error either way. The unit count and concentration are solid; the BOM is not.

So what

This page confirms Hearing aids as an edge-AI silicon beachhead arithmetically, at ~80%. That page argued the beachhead is structurally thin; EHIMA’s 23.16M units at +2.1% and 92.4% top-5 concentration are the numbers behind it. Attach them there rather than re-deriving.

Three screens for the next ear deal:

1. Is the modality on the “uniquely good for” list? If not, the ear is a form factor, not an edge.
2. Which graveyard entry is this? Bose had brand, distribution and engineering. What does this team have that Bose didn’t?
3. If it’s attention decoding — was subject motion controlled?

All-routes read

Suggested prediction to log

“No independent consumer in-ear sensing company reaches >$50M standalone revenue by end-2028.” — prior ~75% it holds. Resolvable, and it’s the crux of this page.

Connected

Sources

Research sprint 15 Jul 2026, three parallel agents. Key primary sources verified directly by Lawrence’s session (not agent-relayed): EHIMA hearing aid sales (23.16M/2025, +2.1%); Francis, Zhao & Guinan 2018; US5787187A Sandia 1996 (expired); Naox FDA clearance Jan 2026.