Sensors for the NextGen AI

Field Notes · Physical AI · September 11, 2026

By Dr. Raymond King, Founder, Applied Sensor Co.

In July 2026, Travis Kalanick raised $1.7 billion for a company called Atoms, led by a16z, to build physical AI. The same year, investors put more than $3 billion into world-model startups betting that the next leap in AI will not come from reading more text.

I run a sensor company in Woodinville, Washington. I have been refining this argument with other hardware founders for years, usually over a workbench. Now the funding data makes it for me.

Here is the argument in one paragraph. To move AI into the real world, we need machines that can perceive, predict, and act autonomously, what Travis Kalanick has started to refer to as Physical AI. That kind of intelligence cannot exist without sensors that measure real spaces and real people in real time (presence, pressure, movement, temperature, time). Imagine you are caring for an aging parent and you have an AI model trained to recognize the health patterns of aging adults. Today, that means sitting down every day to describe how they acted, when they ate, how much they were on their feet, how long they slept, before the model can give you a thoughtful read on whether they are doing fine or starting to decline. Now imagine that same model combined with sensors in the home, feeding it what actually happened all day.

You do not have to describe anything. It notices the pattern shift itself, comes to you, and says it found an anomaly, asking you to step in before it turns into a fall, or worse. This is just one example of how physical sensing turns our current era of text-based AI into the physical AI that will improve a range of industries, from manufacturing efficiency to building usage to patient outcomes. That shift is not just an opportunity, it's becoming a necessity, because the fuel that built the current generation of AI is starting to run out.

The words are running out

Large language models got where they are by mostly reading the internet. That worked because the internet was enormous and nobody had trained on it yet. Both conditions are expiring. Researchers at Epoch AI project that models will exhaust the stock of public, human-written text between 2026 and 2032. The finding was striking enough that AP and PBS covered it as a coming supply crisis.

There are only so many words. Nobody has an answer for that on the text side except synthetic data, which is a model eating its own cooking.

But almost nothing that happens in physical space gets written down. Nobody blogs the temperature of their refrigerator. Nobody logs the minute they got out of bed, how long the kitchen was occupied, or how many hours a chair held someone today. That information exists for a moment and evaporates.

Sensors are how it stops evaporating. Instrumented spaces are, in a real sense, the next training corpus.

Context is the product

Think about what an AI system can do with physical context that it cannot do without it.

Without sensors, you have to tell the model what is happening before it can help you. It is reactive by construction. With sensors, the model already knows. Say you operate two hundred commercial kitchens and a heat wave hits Atlanta. A model with live temperature data from every refrigerator does not wait for a question. It tells you which units are drifting, which food is at risk, and where to send a technician first.

Same model. Different inputs. One is a chatbot. The other is an operator.

The companies I find most interesting right now are building exactly this in eldercare. They deploy sensor networks in senior living facilities, then use AI to turn raw events into patterns: reconstruct the day, compare it to last month, and flag when a habit is drifting in a direction worth a doctor's visit. The value is not in any single reading. The most important thing the data gives you is history. A resident who averages seven trips out of bed a week and suddenly logs fifteen is not an alert. She is a trend. Trends are where health outcomes get caught early.

That work only functions if the underlying measurements are true. Which brings me to my bold statement about AI.

It's easier to do AI

I say this with respect for the people doing serious AI work: it is easier to do AI, the same way it has always been easier to do software. You can edit and iterate in minutes. Your marginal cost of trying again is close to zero. Your supply chain is a keyboard.

Hardware does not iterate in minutes. Hardware has materials, tooling, yield rates, and a bill of materials that has to survive contact with real unit economics. When hardware fails, it fails in someone's building, and you drive there.

So the market has sorted itself the way markets do. There are more AI companies than hardware companies right now, because that is where the money went. Follow the incentives and you get a thousand teams building intelligence and a handful building the instruments that feed it.

I mean it kindly, but plainly: there is a lot of AI fluff out there, and the AI fluff needs information. A model without good physical data is a very articulate guess.

This imbalance is not a complaint. It is the opportunity. Every one of those AI companies needs a sensing layer, and almost none of them should build it themselves. Spending eighteen months of runway learning why a connector fails in the field is not why their investors wrote the check.

What good physical data actually requires

Not all sensor data is good for a model. Three principles separate data an AI team can build on from data that quietly poisons a model.

Measure is better than infer. A lot of "presence detection" is inference stacked on inference. A vibration stopped, so the bed is probably empty. Motion happened nearby, so someone probably walked through. OEM teams tell us the built-in bed-exit features they have tried run behind reality and misfire on the wrong subject entirely. A pressure sensor under the body does not guess. Presence means presence. We do not infer that someone entered or exited. We measure it directly.

I hear the same story from nearly every monitoring company that reaches out to us. They have tried to derive bed presence from motion sensors in the bedroom, or from millimeter wave, or from a consumer health gadget repurposed for a job it was never designed to do. The consumer device is the most painful path, because its maker has no integration support and no interest in providing any. The startup burns months reverse-engineering a product whose roadmap they do not control, feeding their model data of a quality they cannot verify. Everybody who has reached out to us has been disappointed with what exists for bed presence. That disappointment is a market telling you the input layer is broken.

Redundancy makes truth. The more independent sensors that say the same thing, the more likely it is true. That is the core of the Kalman filter, and it has been quietly running your GPS for decades. For an AI company, this means the sensing layer should be designed as a system of agreeing witnesses, not a single loud one.

Collect for the decision, not the pile. Someone described a failure mode to me recently as being data drunk. There is so much you could collect that you collect all of it, and then even an AI system stares at the pile and asks: what do I do with this? The fix happens at the hardware layer, decide what matters first. Sample at the rate the decision needs. Output events, durations, and frequencies rather than an undifferentiated firehose.

They bring the AI. We bring the sensors.

Here is where Applied Sensor Co. has planted its flag, and where we have deliberately chosen not to go.

We are not building AI monitoring products. The companies doing that are our partners and customers, not our competition. I think of us as the pickaxe seller to the people mining with AI. We sell the physical context they are missing.

The pickaxe is real. US Patent 12,612,720 covers a sensor architecture in which conductive polymer and textile elements are embroidered into fabric. Sensing pads, signal traces, and interconnects form in one construction, on mature industrial embroidery equipment, with existing material supply chains. The stitch design is the sensor. Pad geometry, density, sensitivity, and dynamic range are design parameters we tune per application, from a simple on-off presence pad to pressure mapped across an entire surface.

Because the sensor is a textile, it goes where rigid electronics cannot: under a mattress, beneath a rug, into a seat, onto surfaces that bend and fold and get slept on. And because we build in the US, at our facility in Woodinville, WA, an AI team can go from conversation to working prototype in weeks, not quarters.

The output side speaks software. MQTT, REST API, ESPHome, with Wi-Fi, BLE, Zigbee, Z-Wave, PoE, USB, and more on the transport side. Your stack, your protocol, your call. The details live on our OEM integrations page.

What a partnership actually looks like

Because this question comes up in every first call, here is the shape of an engagement.

It starts with your decision, not our catalog. What does your model need to know, at what resolution, how often, and in what places? From there we design the sensing element around your answer: pad layout, sensitivity, dynamic range, shape, and size, tuned to the surface it will live in and the events it must catch. Prototypes come off the same industrial embroidery equipment that runs production, so the thing you test is the thing you will ship. Working prototypes arrive in weeks.

Then we get out of the way. The data flows from our cloud directly into your stack, under your protocol, into your model. Your product stays yours. We are the sensing layer, deliberately and permanently.

The next data is physical

The text era of AI was built on data that happened to already exist. The physical era will be built on data that someone has to go get. That takes instruments, and instruments take a hardware partner who has already made the mistakes.

It's harder to do hardware, but it's necessary. If you are building the intelligence, we would like to build the instruments. Start at asc.com/oem/integrations or write to oem@asc.com.

FAQ

What is physical AI? Physical AI describes AI systems that learn from and act on measurements of real environments: presence, pressure, temperature, motion, and time. Instead of training only on text and images, these systems use sensor data to understand how spaces and people actually behave.

Why do AI companies partner for sensing hardware instead of building it? Hardware carries long iteration cycles, supply chains, certification, and field failure costs that burn startup runway. A hardware partner with proven architecture and US manufacturing compresses years of learning into a purchase order.

What data do textile pressure sensors provide? Pressure, presence, duration, and frequency, delivered over MQTT, REST API, or ESPHome, with transport over Wi-Fi, BLE, Zigbee, Z-Wave, PoE, USB, and more.

Does Applied Sensor Co build AI monitoring products? No. ASC builds the sensing layer. AI monitoring companies bring the models and the application; ASC brings patented textile sensor hardware, built in Woodinville, WA.

---

Author: Dr. Raymond King, Ph.D. Mechanical Engineering, founder of Applied Sensor Co. Before ASC, Raymond led early wearables EMG research at Meta.