Tokens Built The Digital AI Business. Physical AI Needs A Different Unit
Brandon Barbello is co-founder and COO of Archetype AI, with over 10 years of product leadership across Fortune 500 and startups.
gettyAlmost every industrial AI deployment I’ve seen lately eventually reaches the same meeting. The pilot worked, the operations team wants to expand to more machines and finance asks what the numbers look like at 300 assets instead of three.
Increasingly, that question comes with another: Will you charge us by the token?
Enterprises have spent the past few years learning a new vocabulary around AI economics: tokens, inference, context windows, consumption-based pricing and discovering how hard those costs are to predict. Accenture recently reported that one in three organizations exhausts its AI token budget before year’s end.
For digital AI, token economics are justifiable. Jensen Huang has said he’d be “deeply alarmed” if a $500,000 engineer didn’t consume at least $250,000 of tokens a year, because the spend buys output. For knowledge work, he’s right.
But a compressor is not a $500,000 engineer. It produces signal every second it runs, regardless of whether anyone asks it a question. It makes more sense to analyze that signal locally, on hardware the customer already owns, than to ship every second of it to the cloud.
Before Archetype AI, which builds physical AI for industrial customers, I led privacy-preserving, on-device AI projects at Google, where decisions about where intelligence runs were central to building practical products. That experience taught me that physical AI’s economics must account for the continuous flow of sensor data and the hardware already available to process it.
We have an opportunity to build business models that reflect those differences instead of importing LLM economics into the physical world.
Most LLM interactions are episodic: A person asks a question or gives an agent a task, and the model produces an output.
In the physical world, cameras, vibration monitors, pressure gauges and other sensors generate data continuously, and applications are inherently always-on. Meter by the token, and the bill grows with every hour the equipment runs.
The underlying data is different, too. Language models are trained on text and images. They don’t understand what a bearing sounds like at 4,000 RPM. That takes a model built to interpret physical signals directly.
Industrial data is also often too proprietary to leave the facility. Streaming that data to the cloud puts the company’s core IP at risk. This affects how physical AI gets deployed, scaled and priced.
The practical answer is to run intelligence where the data originates. Smaller, general-purpose models can increasingly run on customers’ own hardware, keeping sensitive information on-site and letting decisions happen where the data is produced.
One municipal customer of Archetype AI, for example, uses cameras across hundreds of intersections to improve pedestrian safety. Based on internal analysis, each intersection produces roughly 43 gigabytes of camera data daily. Moving that data to the cloud would cost an estimated $850,000 to $1 million annually before any AI processing. Distilling a larger model into a smaller one let physical AI run on hardware the customer already owns, using its existing cameras.
Running at the edge, the system can determine if a pedestrian is still in the crosswalk without sending data to a remote data center and waiting for a response. That round trip could take longer than the window to act.
Training, fine-tuning and fleet-level analysis can continue to happen in the cloud, while time-sensitive decisions happen closer to the edge. Use the cloud where scale is valuable and the edge where speed, cost and responsiveness matter most.
When AI runs primarily as a metered cloud service, consumption-based pricing makes sense: Someone has to pay for the compute.
But once physical AI runs on the customer’s infrastructure, compute is already paid for, so pricing doesn’t have to correspond to every unit of inference. It can follow from the unit where the customer actually receives value. Industrial businesses already think this way, and it’s how they buy software.
This points to a different model: licensing per asset, machine, facility or application, rather than charging when the system processes data. That brings predictability. Finance teams can count machines more easily than forecast tokens; a company expanding from 10 machines to 1,000 knows the cost before deployment.
The operational value becomes clearer at scale. In one Archetype AI customer deployment, analysis of data from more than 40 sensors across 63 wind turbines identified nine previously unknown failure modes, revealing critical problems before they cause downtime.
The trade-off is that per-asset licensing can be less economical for equipment that needs only occasional analysis. Companies should consider how often the system will be used and the operational value it delivers.
Inference is getting cheaper. Gartner predicts inference on a one-trillion-parameter model will cost providers over 90% less by 2030 than in 2025. Elsewhere, however, Gartner expects inference spend per agentic workflow to climb more than fivefold through 2028, since agentic systems use five to 30 times more tokens per task than chatbots.
In other words, falling unit costs don’t guarantee savings. Every new sensor adds volume, regardless of operational value.
NVIDIA researchers argue that smaller LLMs are already powerful and economical enough for many agentic tasks. Physical AI takes that further, with models built from the start to understand physical signals and run where signals originate.
Physical AI is still in early days. We shouldn’t assume what worked for digital AI is right for physical. Pricing per asset means vendors grow as customers expand, and customers expand as the system proves its value.
Before expanding a pilot, businesses can bring operations, IT and finance together to compare the total cost at three assets, 30 and 300, including licensing, hardware, integration and ongoing support. Weigh those costs against measurable outcomes such as reduced downtime or maintenance hours, and agree on what would justify expansion.
When finance understands what 300 assets will cost and operations can demonstrate the value, the conversation can move from how usage will be billed to whether—and how quickly—to scale.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
