NVIDIA had a big week. During its robotics push it made the new Jetson T4000 generally available, and the spec sheet is the kind of thing that makes you sit up. It's built on NVIDIA's latest Blackwell architecture. It does on the order of 1,200 trillion operations a second. It carries 64 gigabytes of memory. It runs about four times more efficiently than the generation before it. And it does all of that inside a 70-watt power budget, small enough to bolt onto a moving machine.
That is an astonishing amount of intelligence to hang on the edge of the world, away from any data center. If you build physical AI, the natural reaction is to feel the floor move: the edge is getting serious, I'd better keep up, I'd better spec a bigger box.
I build physical AI for small buildings. A sump pump in a basement in Watertown. Forty devices in a building in Northampton. And when I read the Jetson T4000 announcement, I had the opposite reaction. Not "I need one of these." Closer to: my pump needs less compute than the video doorbell by the front door.
Why a robot needs 1,200 teraflops
That number isn't marketing. A humanoid or an autonomous mobile robot genuinely needs it, because of what it has to do every single moment it's powered on. It's reading several camera streams at once. It's running a vision-language-action model that turns "pick up that box" into joint angles. It's planning a path through a space full of people and forklifts. It's running safety logic on top of all of that, in real time, while the whole machine is in motion and the scene never stops changing.
That is a general-purpose problem. The robot doesn't know in advance what it will see, what it will be asked to do, or where it will be standing. It has to be ready for almost anything, which means it has to carry a brain big enough for almost anything. Hence Blackwell at the edge, 64 gigabytes of memory, and a 70-watt power draw that needs real cooling. The compute is large because the job is open-ended and the machine moves.
Why a pump needs almost none of it
Now stand in front of the sump pump in that Watertown basement and count what it actually has to think about. There's one vibration signal. There's one current reading from a clamp on its power line. The job is not open-ended at all: learn what this one pump sounds and draws when it's healthy, and notice when that normal starts to slip. The pump isn't moving through a changing world. It's bolted to the floor, doing the same thing it did yesterday, narrating its own health on two channels.
A model that only has to know one machine's normal is tiny. It runs comfortably a few times a second on a board the size of a deck of cards, drawing single-digit watts, with no cooling fan and no special silicon. It will happily run on a Raspberry-Pi-class computer, or even a microcontroller, with room to spare. Put concretely: the little box watching that pump draws less continuous power, and needs less compute, than the smart doorbell most people already have at their front door.
| A general robot (why Jetson T4000 exists) | A single-asset watchman |
|---|---|
| Many camera streams, vision-language-action, planning, safety | One vibration channel, one current channel |
| Must handle an unknown, changing world | Must learn one machine's normal |
| Moving through space, in real time | Bolted to the floor, doing one job |
| ~1,200 teraflops, 64 GB, 70 watts | Pi-class board, single-digit watts |
| Brain big enough for almost anything | Model just big enough for one pump |
The trap in "AI needs giant chips"
Here's the conclusion a small building owner reaches from a week of headlines about a 1,200-teraflop edge module. They think: AI is getting more powerful and more power-hungry. The chips are getting bigger and more expensive. This is enterprise-scale stuff, and it'll be years before it's small enough and cheap enough for my boiler.
That's backwards, and the reason is the same reason the robot needs all that compute in the first place. The chip is getting bigger because the job is getting more general. The teraflops are buying generality and motion. A single fixed asset needs neither. The watchman on your pump isn't a small version of a robot brain that's waiting to grow up. It's a fundamentally smaller problem that was always going to fit on a tiny, cheap, low-power board, no matter how large the frontier of edge compute gets. The two numbers are moving in opposite directions on purpose: the robot's brain grows because its world is open, and your pump's brain stays tiny because its world is one machine.
Right-sizing is the whole advantage
This is the quiet edge a small building has over a fleet, and it shows up everywhere once you see it. A robot needs a $2,000 edge module because it has to be ready for anything. Your boiler needs an $80 board because it only has to be a boiler. A robot's brain has to be replaced and upgraded as its tasks expand. Your pump's watchman learns one machine and is essentially done. The robot pays for flexibility it can't avoid. You get to refuse to pay for flexibility you'll never use.
And the savings compound past the chip. A few-watt board doesn't need a cooling system, doesn't need a special enclosure, doesn't need its own circuit. It can run off the same outlet as the equipment it's watching and you'll never notice it on the power bill. The whole watchman, sensor and board together, costs less than the Jetson module alone, because it was scoped to one machine instead of scoped to the unknown.
What this means if you own a building, not a robot
When the biggest name in AI hardware spends a week showing off a 1,200-teraflop brain that fits in 70 watts, it's easy to hear that as "AI hardware is huge and getting huger, and the small, affordable version is somewhere off in the future." But that giant brain exists to do a giant, open-ended job: perceive an unknown world and move through it. Your building doesn't have that job. It has a handful of fixed machines, each one quietly narrating its own health on a couple of channels, waiting for something small and cheap to listen.
So don't wait for the frontier to come down to your basement. It doesn't need to. The asset on your wall was always a small problem, and small problems fit on small, low-power, inexpensive boards, the kind already sitting inside your doorbell. Put one of them next to your pump, your boiler, your compressor. Let it learn that one machine and warn you before it fails. You don't need a robot's brain. You need a deck-of-cards board that knows one pump cold.
Right-size it. Watch the one asset that matters.
Each critical asset gets its own small detector, built from off-the-shelf sensors, trained on its own measured history, running on a low-power board on the wall, local and watching 24/7. No giant edge module, no cloud round-trip, no compute you'll never use. It learns your pump from your pump and warns you early when something drifts. $99 to $199 per month, hardware under $3,000.
See how it worksSources: NVIDIA National Robotics Week coverage and newsroom announcements, including general availability of the Blackwell-architecture Jetson T4000 module (reported at ~1,200 FP4 TFLOPS, 64 GB memory, ~4x prior-generation energy efficiency, ~70-watt configurable envelope) and new open physical-AI / Isaac foundation models (NVIDIA Blog and Newsroom, June 2026); Weekly Robotics #365 and June 2026 coverage of open robot foundation models including Alibaba's Qwen-RobotNav / Qwen-RobotManip / Qwen-RobotWorld; edge-AI funding context (SiMa.ai $85M; ~$310M average round size for last-round-2025/26 edge-AI companies vs ~$140M for 2023-or-earlier; Wayve $1.2B Series D; Figure ~$1.7B total) per StartUs Insights and New Market Pitch edge-AI fundraising trackers; 2026 predictive-maintenance trade aggregates (65% of maintenance teams planning AI adoption by end of 2026; shift from predictive to prescriptive maintenance). Field deployments at The Intersecto Watertown sump-pump site and Northampton 40-device building. Companion brief: /Users/tdeshane/lobster/research/physical-ai-brief-2026-06-26.md.