Evolution spent 3.7 billion years perfecting perception and only a sliver of that on reasoning. Yet today's AI efforts are focused on reasoning and blind to most of what could be sensed. Machines perceive the world through a tiny slice of modalities (mostly cameras and microphones), ignoring chemistry, smell, electric and magnetic fields, radiant heat, polarisation and force that biology exploits routinely. We believe large-scale opportunities can be unlocked by building "sensory intelligence": co-designing novel sensors with AI so machines can perceive reality in fundamentally new ways and develop capabilities that haven't been possible before. This post explores (1) sensing in nature, (2) the state of machine sensing, (3) the opportunities ahead, and (4) a rough sketch of venture-compatible company shapes.

Contents

• Sensing in nature and machines

• Sensing opportunities

  • Quantum sensing

  • Electronic nose

  • Fine-force touch

  • Electroreception

• Company shapes

• Non-consensus beliefs

Sensing in nature and machines

Nature's senses sort by the physical stimulus being measured. A bee's ultraviolet vision and a pit viper's heat pit are both photoreception, sampling different bands of one spectrum. Seven stimulus families cover almost everything alive, plus a plant-specific cluster. The oldest senses measure the most fundamental things — molecules, fields; image-forming vision and hearing are late, expensive refinements.

Machine perception first evolved in the 1940s through RADAR (radio detection and ranging) and SONAR (sound navigation and ranging); the CCD (charge-coupled device) image sensor in 1969; LIDAR (light detection and ranging) and thermal imaging in the 1980s; smartphone sensor fusion (camera, microphone, GPS, inertial unit, magnetometer, barometer) in the 2000s; the deep-learning vision breakthrough in 2012; and digital olfaction and commercial quantum sensors in the 2020s. Two senses — vision and hearing — plus ranging and navigation instruments dominate; smell, taste and touch barely register.

The chart below sorts each capability by how far it reaches — what humans can feel, what only other life can, and what only machines can — with two five-point reads on the machine side: maturity today and opportunity ahead (untapped headroom). The gap between them is where the value is.

Two things jump out. (1) Every human sense is a fainter copy of some animal's; the AI we train on human-legible data inherits that same narrow slice. (2) Quantum sensing and chemoreception (sensing molecules — the "e-nose") are the lowest-maturity, highest-opportunity frontier. Those are where we focus.

Sensing opportunities

This section breaks down the opportunity sets from the top down. First, an overview of the four most promising sensing technologies by their 5–10-year readiness and their novelty (the degree to which they enable new behaviours). Then we double-click into each of the four most promising capabilities — quantum sensing, the electronic nose, touch and electroreception — covering the leading technical approaches, the bottlenecks, and the application potential, including market-size estimates.

Capabilities as opportunities

The chart below reflects the opportunity sets with the highest asymmetry: low maturity today, large market opportunity in the future. The market sizes are approximations derived from various market reports, linked at the bottom. Importantly, these approximations are derived from integrated capabilities — they do not reflect the expected market of the underlying sensing technology alone.

Opportunity: quantum sensing

What it is. Quantum sensing uses single quantum systems — atoms, ions, electron spins, photons — as measurement instruments. These systems are so fragile that the smallest change in their environment shifts their quantum state in a precisely readable way. The fragility that makes quantum computers hard to build is what makes quantum sensors work: the disturbance is the measurable signal.

The prize is sensitivity orders of magnitude beyond classical sensors, across magnetic and electric fields, gravity, acceleration, rotation, time and temperature. A first generation — atomic clocks and SQUIDs (superconducting quantum interference devices, magnetic-field sensors that require cryogenic cooling) — existed for decades. The next generation works at room temperature and can be shrunk to chip size.

Four method families are the most promising approaches today. No single method dominates; each use case picks its own quantum system, depending on context.

Where the opportunity sits. The sensor hardware itself is a small market — McKinsey estimates roughly $1–6B by 2040. Value accrues in the capabilities each sensor unlocks (navigation ~$7.4B by 2034, brain–computer interfaces ~$14B by 2035). See the section on company shapes below.

Sources — science & approach: Degen, Reinhard & Cappellaro, Quantum sensing, Rev. Mod. Phys. 2017 · High-sensitivity nanoscale NV quantum sensors, Communications Materials 2025 · A portable OPM-MEG platform, Imaging Neuroscience 2024 · Quantum-enhanced navigation with atom interferometry, arXiv 2025 · Optical clock resolving gravity across a millimetre, Nature 2022 Sources — market sizes (one representative forecast each; firms differ): quantum-sensing market — McKinsey 2024 · quantum navigation & sensing outlook — McKinsey 2021

Opportunity: electronic nose

What it is. An electronic nose (e-nose) is machine olfaction — the artificial sense of smell. It copies the architecture of the biological nose through an array of cross-reactive sensors, each responding a little differently to many molecules, with AI reading the pattern across the whole array. This is combinatorial coding — the same trick the nose uses, where ~400 receptor types encode millions of smells. The target is VOCs (volatile organic compounds — carbon-based molecules light enough to evaporate into the air we breathe), the chemical signatures that leak from a food, a mate, a disease or a buried explosive.

Conventional lab instruments like GC-MS (gas chromatography–mass spectrometry, the current standard) name every constituent molecule, slowly and expensively. An e-nose reads the overall scent character in real time and learns to tell foreground from background — closer to how a dog works than a mass spectrometer. Two barriers long held it back: the limit of detection (reacting to single molecules) and the limit of recognition (decoding a smell in a noisy, shifting plume). Both are now falling — though there is still a long way to go (see bottlenecks).

Bottlenecks. The gating problem is data: olfaction has no ImageNet (the structured visual database that trains and benchmarks AI models) equivalent yet. Vision has PNG, audio has WAV, language has tokens — olfaction has no digital standard that captures a smell. Representations are fragmented, which limits benchmarking. And chemical sensors age, each type drifting differently, so a model trained today degrades tomorrow unless it keeps recalibrating. The challenges are enormous and likely to take the better half of a decade to overcome.

Where the opportunity sits. The e-nose sensor market is real but mid-sized — about $30B in 2025, projected to ~$77B by 2032 at roughly 14% CAGR (forecasts vary widely by firm). As with quantum sensing, the leverage is one layer up, in the served markets each nose competes in — not e-nose revenue itself: food and pharma quality control (food-safety testing ~$56B by 2035), CBRN (chemical, biological, radiological, nuclear) threat detection (~$30B), and non-invasive cancer screening from breath and urine (multi-cancer early detection ~$6.8B). The moat is not the sensor, which commoditises, but the labelled scent library — the proprietary dataset of clinical and field samples that trains the classifier and compounds with every deployment.

Further reading — science & approach maturity: Mershin et al., Machine Olfaction and Embedded AI, arXiv 2025 · Advanced electronic noses for future robotic olfaction, npj Robotics 2025 · France & Daescu, AI and Olfaction: A Survey, ChemRxiv 2025 Sources — market sizes (one representative forecast each; firms differ): e-nose market — Mershin et al. · multi-cancer early detection — Nova One Advisor · CBRNE defence — GM Insights · food-safety testing — Expert Market Research

Opportunity: fine-force touch

What it is. Fine-force touch is machine tactile sensing — an electronic skin that measures the mechanical contact between a robot and the world: force (how hard it presses), shear and slip (whether an object is sliding out of the grip), local shape, texture and hardness. Vision is too limited here, which is why touch is the critical bottleneck to dexterous manipulation: at the moment of a grasp the hand occludes the very thing it is holding, and force, slip and compliance are not optical quantities. Vision tells a robot where something is; touch tells it how hard, whether it is slipping, and what it is made of.

Touch lagged vision for the same reason smell did: no ImageNet-scale training data, and no skin that was cheap, durable and large-area at once. Two shifts are breaking that. First, vision-based tactile sensors (optical tactile sensing) turn touch into an image an ordinary neural network can read, giving sub-millimetre contact geometry from a standard camera. Second, coverage is going whole-hand: the F-TAC hand (Nature Machine Intelligence, 2025) embeds high-resolution touch across ~70% of the hand's surface at 0.1 mm resolution and beats non-tactile baselines across 600 real-world grasping trials.

Where the opportunity sits. The tactile-sensor market is mid-sized — roughly $14.5B (2025) rising to ~$47.5B by 2035. As with olfaction, the leverage is one layer up, in the platforms touch unlocks: dexterous manipulation for humanoid robots (~$38B by 2035), surgical and medical haptics (surgical-robotics ~$46B by 2035), and consumer and XR haptics (~$7.1B). The moat is the contact dataset and grasping policy — the proprietary flywheel of real manipulation episodes that trains the model. Whoever collects the most contact data wins, which is exactly the shape of startups now teaching dexterous hands from sensor-glove human demonstrations.

Sources — science & approach: Tactile Robotics: An Outlook, arXiv 2025 · Embedding high-resolution touch across robotic hands (F-TAC Hand), Nature Machine Intelligence 2025 · Biomimetic multimodal tactile sensing, Nature Sensors 2025 · Visuotactile field guide · tactile-sensor comparison — SVRC Sources — market sizes (one representative forecast each; firms differ): humanoid-robot market — Goldman Sachs · tactile-sensor market — SNS Insider · surgical-robotics market — Precedence Research · haptics technology — IDTechEx

Opportunity: electroreception

What it is. Electroreception is sensing electric and bioelectric fields — the faint voltages that every nerve, muscle and heartbeat produces, and the fields around any live wire or charged object. It is a sense humans lack entirely, but sharks, rays and the platypus rely on it: a shark's ampullae of Lorenzini detect the bioelectric field of hidden prey down to a few billionths of a volt. For machines it opens two otherwise-closed domains: the body's own electrical activity — the firing of nerves, the heart and the muscles, normally reached only with gels, needles or implants — and the electrical state of infrastructure — the voltage and faults inside a live grid, read without touching it.

Sensors sort by what field they read and how close they must get. The set below is focused, not exhaustive.

Where the opportunity sits. Electroreception is under-invested relative to olfaction and touch, and asymmetric — the body's electrical signals are continuous, information-dense and today mostly unread outside a clinic. The machine analogues tap large end markets: bioelectronic medicine and electroceuticals (~$48B by 2035), non-contact and wearable biopotential monitoring (EEG/EMG equipment ~$7.9B by 2035), and industrial electric-field sensing (~$4.8B by 2032). The moat is the same shape as in the other modalities — the labelled physiological dataset and the model that reads it, not the electrode, which commoditises.

Sources — science & approach: Bioelectronic Medicine and Neural Interfaces, J. Bio-X Research 2025 · Next-generation bioelectronic medicine: non-invasive closed-loop neuromodulation, Bioelectronic Medicine 2024 · Chi et al., Dry-Contact and Noncontact Biopotential Electrodes: A Review, IEEE Rev. Biomed. Eng. · Bioinspired soft electroreceptors for artificial precontact somatosensation, Sci. Advances 2022 Sources — market sizes (one representative forecast each; firms differ): bioelectronic-medicine market — Metatech Insights · EEG & EMG equipment — Market Research Future · electric-field sensor — Verified Market Research

Company shapes

Leverage principles. We are eager to back companies that use many or all of the principles below to build durable businesses.

1. Prevention beats cure; continuous beats episodic. Most problems are cheap to fix early and catastrophic to fix late, so a sensor that catches a problem while it is still small is more valuable the steeper the cost-over-time curve (stage-1 cancer vs. stage 4; the first cracks in a bridge; an annual check-up vs. a continuous breath monitor).

2. Resolving scarcity bottlenecks. Some perceptions are locked inside a rare, expensive, trained human or animal. Sensing plus AI turns that capability into something cheap, everywhere and always-on (trained dogs and bomb-disposal experts are rare).

3. Making the invisible priceable to create markets. You cannot trade, insure or regulate what you cannot measure. When a sensor makes a hidden quantity continuously measurable, it can be priced (methane leaks became a tradable, finable emission once satellites and ground sensors measured them).

4. Reaching the inaccessible. Some places are physically or economically closed to existing sensors: inside a living body, behind a wall, under rubble, underground, in jammed airspace. A modality that reaches them opens a domain that was simply shut.

5. Fusion compounds. Two sensors that fail in different ways are worth more than the sum of their parts, because each covers the other's blind spot (vision + lidar in self-driving).

The durable sensing company is not a sensor company. In every historical case the raw transducer commoditises; value accrues to whoever owns the layers above it.

On data loops

The most valuable companies of the smartphone era ran on a sensor that shipped in every phone for free (GPS, the microphone). None of them made the sensor. Each owned the proprietary data loop on top of it. Note what they are, though: they mostly sell you an outcome — a ride, a song ID, a match — not the engine itself. The flywheel is the moat that defends a vertical; it is not the same thing as being a platform. The sensing prize is to run this engine and extend to selling the capability.

On business models

A sensing company is a stack. The bottom layer — the sensor — is the one layer that commoditises. The value and the defensibility sit in the three layers above it, and in the flywheel that connects them: every deployment produces data no one else has, which trains a better model, which wins more deployments. Own more than the sensor.

If the above is the "engine", consider the "chassis" — the business models taking shape around it. Sensing companies fall into three archetypes:

1. Component manufacturers stay durable only behind scale and process moats (Sony, Illumina). For most startups this path leads to the Velodyne trap (merged with Ouster, ~$3B market cap): hardware-first sensing gets commoditised to death — not venture-compatible.

2. Vertically integrated businesses "sell the outcome" by bundling sensor + model + data — in self-driving (Tesla, Waymo), diagnosis (Cala Health, ~$272M raised) or brain–computer interfaces (Neuralink, ~$9B valuation). TAM is bounded by the vertical. Venture-compatible.

3. Platform businesses "sell the data engine" by co-designing sensors and models, winning a lighthouse market, then selling the capability broadly across domains. Osmo (ca. $130M raised) intends to build the "ImageNet for smell" and sell it across domains; Tacta Systems (ca. $75M raised) pursues the same for full-body robotic touch. Venture-compatible.

Non-consensus beliefs

AI is bottlenecked by perception, not reasoning — and the field is spending on the wrong constraint. The consensus unlock is bigger models and more compute. We believe the next jump in AI capability comes from widening the senses — the way nature built capability out across every animal and plant. Evolution spent 3.7 billion years on perception and a sliver on cognition; today's AI inverts that ratio and is starting to hit the ceiling of it. The scarce input is not more of the same data, it is new modalities of data — chemistry, fields, force, neural current — that no model has ever seen.

The sensor itself is usually a red herring. The deep-tech reflex is to pay a premium for defensible transducer IP — the novel e-skin, the exotic chemical sensor. We think that, for nearly every modality, the transducer commoditises the moment it works. The durable company points a commodity readout — a camera, a phone, a cheap electrode — at a quantity nobody was capturing, and owns the data loop on top. A GelSight is a camera doing touch; Shazam was a microphone; Uber was the GPS already in every pocket. Backing novel-transducer IP is backing the one layer that evaporates. Quantum is the exception that proves the rule: there the physics of capture is genuinely hard and interpretation is trivial, so the transducer stays the moat.

If you're building a vertically integrated or platform company on next-generation sensing, let's talk.

Related