Sensors of the future won't be more connected. They'll be more decisive.
An early-decision fabric that lives in the sensor and asks one question, continuously: is there enough evidence here to justify waking anything else? When the answer is no — which it almost always is — nothing downstream spends a thing.
Every sensor we add is there for a good reason.
Continuous sensing is how we catch the fall, find the leak, stop the machine before it breaks, and turn the heat off in an empty room. A great deal of it exists precisely to save resources — which is why the deployment curve isn't going to bend, and why what each sensor costs to run is the thing worth fixing.
Safety
Falls, intrusion, gas leaks, glass break. Someone is told while it still matters, instead of after.
Health
ECG and activity monitored on-body. The arrhythmia is caught on the beat it happens, not at the next appointment.
Uptime
Vibration and acoustics flag a failing bearing days early — a scheduled fix instead of an unplanned line stop.
Conservation
Occupancy sensing cuts heating, cooling and lighting in empty space; leak detection saves water directly. Sensors already pay resources back.
Autonomy
Vehicles, robots and tools that understand what's around them, without asking a person or a server first.
Reach
Measurement where no person can stand — inside a tire, a turbine, a pipeline, a body.
Volumes are industry-estimate ranges — orders of magnitude, not point forecasts. The takeaway isn't a single number: continuous sensing is scaling faster than the resources available to process it.
We pay for it in resources — mostly to conclude that nothing happened.
An always-on sensor that streams doesn't only spend its own power. It pulls a chain behind it: radio, network, datacenter compute, cooling, and all the physical hardware that has to exist to serve it. The overwhelming majority of that spend produces one finding — no change.
Electricity
a year — the world’s data centres already draw around this much electricity — more than the entire United Kingdom uses — and it is set to roughly double to ~945 terawatt-hours by 2030, on the order of Japan’s whole national grid. Every forwarded frame lands on that meter, plus the device power to capture and send it.
IEA, Energy and AI (2025)Water
a year — what the world’s data centres already draw for cooling and the power behind it, rising fast as AI scales toward a projected ~9.3 trillion litres a year by 2030. Chip fabs draw heavily too.
IEA · UN estimatesBandwidth
Spectrum, backhaul and uplink are provisioned and powered for traffic whose typical payload is an unremarkable frame. Radio time is usually the largest single draw on a connected sensor's battery.
device power budgetsPhysical hardware
Racks, radios, cooling plant and the wafers behind them all have to be manufactured before the first frame arrives — an embodied cost paid up front, then again at refresh.
embodied-carbon literatureBatteries
Internet of Things batteries are discarded worldwide every day, under 40% recycled. Devices are built to last 10+ years; their batteries often last two. The cell is frequently the biggest, heaviest, costliest item in the bill of materials.
EnABLES position paper (EU)The common case is nothing happened — and today it costs very nearly as much as something did.
Third-party resource figures above are published global estimates from the cited sources — not Spot Vision measurements.
Decide at the sensor. Let everything else stay asleep.
When the answer is "nothing worth waking for," no host wakes, no radio keys up, no packet is sent, no server spins, no water is evaporated to cool it. The energy isn't in the decision — it's in everything after it. So the further left you decide, the less you spend.
An always-on primitive cheap enough to simply leave running — microjoule-class in sensor-integrated silicon (the ASIC target, proven correct on FPGA today). A rejected event stops here and never pays the tiers below.
Multiplier arrays, external memory, clock trees and data movement all wake — for every event, including the ones that don't matter.
Capture, compress, key the radio, traverse the network, run inference, cool the rack. Billed per device, per month, for the life of the fleet.
| Spot Vision fabric | Larger on-device NPU | Cloud | |
|---|---|---|---|
| What must wake | Nothing beyond the fabric itself | Host processor, memory, clocks | Radio, network, servers, cooling plant |
| Resources drawn | A sliver of device battery | Battery, plus silicon area and heat | Electricity, water, bandwidth, server hardware |
| Recurring cost | None | None | Per device, per month, forever |
| Latency | Immediate, deterministic | Bounded by host wake-up | Round-trip; fails without connectivity |
| Raw data leaves device | No | No | Yes |
Every event you reject skips the expensive path entirely — and in always-on sensing, rejection is the common case, not the exception. That's the entire product.
Battery & bill of materials
- Kept always-on, an order of magnitude less energy than a conventional accelerator on a host — so the same runtime fits a far smaller cell
- Or keep the cell and multiply service life instead — fewer recharges, fewer swaps
- Shrink the biggest, heaviest line in the bill of materials and the whole product shrinks
Operating cost
- No per-device cloud inference bill, and no storage bill for footage nobody watches
- Uplink only for events worth sending — smaller data plan, less airtime
- Fewer truck rolls and maintenance windows across a deployed fleet
Product
- Smaller, lighter, longer-lived hardware in the customer's hand
- Works with no connectivity, and answers in real time
- Private by construction — the decision already happened locally
Energy figures are relative magnitudes for a rejected event and vary with workload, radio and duty cycle — they compare where a decision is made, not one chip against another. Spot Vision's own measured energy to date is stated below as an upper bound on isolated compute.
A small circuit, almost no memory, and results you can trust.
The energy win comes from not being a neural accelerator. There are no multiplier arrays and no DSP blocks — the most power- and area-hungry parts of a conventional engine are absent, not merely optimised. What's left is small enough to leave switched on forever.
No multipliers, zero DSP
Multiplier-free inference — zero multipliers in the datapath. The arithmetic that dominates NPU power and area simply isn't in the design.
No external memory
Models stream from cheap serial flash. No DRAM, no host processor babysitting the block, no memory bus to power up.
Deterministic and bit-exact
A fixed cycle count per compiled model. Timing closure and verification are ordinary engineering, not a research project.
Tiny and field-updatable
A small logic footprint that fits beside the sensor, with new models loaded from flash long after the chip has shipped.
Accurate where it counts
The gate rejects the common case reliably and passes anything ambiguous up the chain — so accuracy is spent where it changes the outcome.
It ships as an IP core and drops into your floorplan.
Licensed as a block you integrate like any other. It asks almost nothing of the chip around it — no DRAM, no DSP blocks, no host supervision — which is why it fits into an SoC, an ASIC, or a sensor’s own silicon without disturbing the rest of the design.
Evaluate on hardware
Run the fabric on a commercial FPGA with our reference models and measure energy and accuracy on your bench, against your data — not from our slides.
We build your model
You bring the use case and the data; we train the detector and compile it onto the fabric. The model engineering is ours — not a toolchain you have to run.
Integrate the core
Drop the block into your silicon. Small footprint, deterministic timing, standard flow, serial flash for the model library.
Field-updatable
As we develop new or improved detectors for your product, they load to flash after the chip ships — one silicon design becomes a family of products, no re-spin.
A cheap always-on model wakes a heavier one only when something happens — the common case costs almost nothing.
Progressively more capable models handle progressively harder cases — accuracy where needed, battery everywhere else.
Detect, classify, verify — all inside the one fabric, with no host processor needed to orchestrate the chain.
Sit it directly after the sensor and the question is always the same. The signal changes; the primitive doesn't.
Decide early. Spend resources only when it matters.
An early-decision fabric for sensor systems — any signal, always on, delivered as licensable IP for integration into low-power silicon. If you build always-on sensing products, we'd like to show you the hardware.
info@spotvision.ioResources saved
- An order of magnitude less energy per device, on the events that dominate the duty cycle
- No datacenter electricity or cooling water spent to conclude that nothing happened
- Uplink and spectrum reserved for traffic that carries information
- Fewer cells manufactured, shipped and landfilled across a fleet's life
Money saved
- A smaller battery in the bill of materials, or a much longer interval between swaps
- No recurring per-device cloud inference or storage bill
- Less radio airtime and a smaller data plan per unit
- Fewer service visits; deployments that were uneconomic become viable
Product gained
- Smaller, lighter, longer-lived hardware, no longer designed around its cell
- Real-time response that works with no connectivity
- Private by construction — raw data need never leave the device
- One licensed block that ships into many products and improves after launch