What physical AI actually costs a small team

You have a physical problem. A camera pointed at a shelf that nobody has time to count. A gate that should open for your van and not for anyone else’s. A machine that fails in a way an operator can see and a sensor cannot. Somewhere in the last year you watched a humanoid robot demo, read the phrase “physical AI”, and filed the idea under expensive. Then a chip announcement went past with numbers in the thousands of teraflops and you had no way to tell whether it moved your problem or not.

This guide is the arithmetic. It covers where the hardware ladder actually starts, what running the same model in the cloud costs instead, and the four reasons that make a device worth owning. It is written for a solo operator or a small team deciding whether to build one of these things once. It is not for robotics engineers specifying a humanoid platform, and it will bore anyone who has already shipped an embedded product, because most of what follows is the stuff you learn on your first one.

Physical AI is a constraint, not a category

The term covers anything where a model perceives the physical world and something moves or someone waits on the answer. That framing matters because it tells you what the hard part is. A model that has 30 seconds to label a photo has no interesting engineering in it. A model that has 100 milliseconds before a conveyor belt has moved the part out of frame does.

So the first question is not which chip. It is how long your loop is allowed to be, and what happens when the network is not there. If your answer can wait a second or two and the site has reliable internet, you are building a normal software product with a camera attached, and the rest of this guide is about a decision you do not have to make. If the answer has to come back faster than a round trip, or the site is a field, a basement, a vehicle, or a customer’s building where you are not allowed to send images out, you are in physical AI territory whether or not there is a robot in the picture.

The hardware ladder runs from $70 to $5,499

NVIDIA’s Jetson line is the default answer, and the top of it is genuinely expensive. The Jetson T5000 module runs 2,070 FP4 sparse teraflops with 128GB of 256-bit LPDDR5X and draws 40W to 130W [2]; it costs $4,999 in 1,000-unit quantities [1]. The T4000 below it does 1,200 FP4 sparse teraflops with 64GB in a 40W to 70W envelope [2] at $2,999 [1]. The Jetson AGX Thor Developer Kit, which is how you get a T5000 on your desk with a thermal transfer plate and a 1TB NVMe M.2 slot [2], is $5,499 [1].

Those are the numbers that make people file the idea under expensive, and they are the wrong rung to be looking at. The Jetson Orin Nano Super Developer Kit does 67 INT8 TOPS on 8GB of memory in a 7W to 25W envelope [3] and costs $399 [1]. The bare Orin Nano 8GB module is also $399 at volume, Orin NX 8GB is $649, and Orin NX 16GB is $999 [1]. Below the Jetson line entirely, a Raspberry Pi AI HAT+ starts at $70 and comes in 13 TOPS and 26 TOPS versions built on Hailo accelerators [7].

The distance from $70 to $5,499 is the whole point. A single camera running one detection model at a few frames per second is a bottom-rung problem. Reading a gauge, counting stock, spotting an empty pallet, checking that a person is wearing the right gear: none of that needs a robot brain. The top of the ladder exists because a humanoid runs several models at once, in a control loop, on a battery. If you are not doing that, buying for it is the most common way small teams turn a $400 project into a $6,000 one.

Cloud inference is cheap enough that the edge needs a different reason

Before you buy anything, price the boring alternative: send the frame to an API. Gemini 3.5 Flash-Lite is $0.30 per million input tokens on the standard tier [5]. An image with both dimensions at or under 384 pixels costs 258 tokens; larger images are tiled into 768x768 tiles at 258 tokens each [6]. That puts a small frame at roughly eight thousandths of a cent of input.

Run that at 2,000 frames a day and the image input costs about $4.70 a month. Add your prompt and a short structured answer per frame and it is still single-digit dollars. Against $399 of hardware [1] that you also have to mount, power, update, and drive to when it stops working, the token bill is not the argument. At the volumes a small operator actually generates, cloud inference is close to free and the device is the expensive option.

Those four are the real reasons, and each is a yes or no you can check before spending anything. Latency counts when the answer has to arrive faster than a round trip to a data center. Connectivity counts when the thing has to keep working after the link drops, and you know for how long. Privacy counts when a contract, a policy, or a regulator stops the images leaving the building. Bandwidth counts when you are streaming video that costs more to upload than to process locally. If you cannot tick at least one of those, the honest answer is that you are building a cloud app with a webcam, and you should build that instead and stop reading.

Announced hardware is not buyable hardware

On 15 July 2026 NVIDIA introduced two mid-range Thor modules. The Jetson T3000 delivers 865 FP4 teraflops with an eight-core Neoverse Arm CPU, 32GB of LPDDR5X at 273GB/s and 25 GbE, in roughly half the size and power of the T5000; an IGX T3000 variant delivers the same performance with integrated functional safety and runs NVIDIA’s Halos for Robotics stack. Below it, the T2000 does 400 FP4 teraflops on 16GB. Companies building on Jetson include 1X, Agile Robots, Amazon Robotics, Boston Dynamics, FANUC, Hitachi and Techman Robot [4].

Both modules are scheduled to become available in the first quarter of 2027. NVIDIA said T3000 emulation mode would arrive within that same month in JetPack 7.2.1, with T2000 emulation to follow in a later release, so teams can develop against the modules before silicon ships [4]. That gap is the durable lesson, not the teraflops. Between an announcement and a purchase order there is usually a year, and a plan built on a part you cannot order is a plan with a year of slack in it that nobody wrote down. Build the bill of materials from what a distributor will ship you this month, and treat announced parts as a reason to design for a swap rather than a reason to wait.

Prices move up as readily as down

The other half of that lesson is that the price you quoted six months ago is not a fact you can rely on. NVIDIA’s own FAQ now lists the Jetson AGX Thor Developer Kit at $5,499 and the Orin Nano Super Developer Kit at $399 [1]. Those were $3,499 and $249 before a July 2026 revision that raised 19 Jetson products by between 33% and 101%, with the Jetson Nano module going from $99 to $199 [8]. There was no specific announcement about the change; the updated numbers simply appeared in the FAQ [8].

If you have a spreadsheet with per-unit margins in it, the Orin Nano Super devkit line moved 60% [8] on a figure you had treated as settled. The practical response is not to distrust the vendor, it is to re-price the bill of materials before every volume commitment, and to keep a second rung of the ladder in the design so a price move does not strand the product. It also cuts the other way: the same instinct that says “shelved, too expensive” about a decision made a year ago is worth re-running, because a tier that did not exist then may exist now.

The module price is the part of the bill you can look up

A module is not a product. Between the $399 line item and something bolted to a wall there is a carrier board, an enclosure, a camera and lens, a power supply that survives your actual environment, thermal design, and a mounting bracket somebody has to make. There is also the software: a Jetson is an embedded Linux computer, and somebody on your side has to own kernel versions, image flashing, over-the-air updates, and the day a unit in the field stops responding.

That is why the developer kit exists and why buying one first is nearly always right. $399 buys you the answer to whether your model runs fast enough on that class of hardware at all, which is the only question that matters before the rest of the bill comes due. It is also why the list of companies building on the new Thor modules reads the way it does [4]. Those are firms with embedded engineering teams. The platform is aimed at them, and the parts of it that are genuinely for a two-person shop are the cheap end and the developer kits.

What still goes wrong

The most common failure is buying up the ladder. A small team specs a $2,999 module because the demo video used one, then discovers the workload was a single detection model that a $399 board handled at frame rate. The fix is dull and reliable: get the cheapest developer kit that could plausibly work, run your actual model on your actual footage, and only move up when you have measured a number that says you must.

The second failure is the part of the project that has nothing to do with AI. Mounting, lighting, lens choice, cable runs, and the question of what happens when a cleaner unplugs it are where these projects actually die, and none of them are improved by a faster chip. Budget more time for the physical install than for the model, because the model is the part that already works.

The third is treating any of the numbers here as durable. Prices on this line moved by up to 101% in a single revision with no specific announcement [8], and the mid-range Thor modules are not scheduled to ship until the first quarter of 2027 [4]. Anything with a dollar sign in this guide is a snapshot with a date on it, and the sources are linked so you can check them rather than trust the reprint. Re-check before you commit, not after.