This article was originally featured in The Robot Report’s State of Physical AI and Robotics report. Read the full report here
In physical AI, the model does not capture value. The orchestration layer that runs the operation does.
The orchestration layer generates proprietary data every time it runs in a live environment. Unlike language models, this data does not exist on the internet. It can only be created by running real operations at scale.
As live warehouse deployments accumulate across sites, the company that owns the orchestration layer builds a compounding data asset with time-based defensibility. Every live site generates proprietary operational data like failure scenarios, recovery patterns, and unit economics that a new entrant can’t replicate. That is a software-like margin profile sitting inside what the market currently prices as a hardware or services business.
The verdict on the other players is already visible. Hardware will commoditize. Model builders will either commoditize or consolidate. Systems integrators will earn services multiples.
The orchestration layer is where the multiple re-rates. Models produce intelligence, but operations need to learn from live data.
Over the past five years, AI has made its biggest progress in text-based tasks, where you only need a model, data to train it on, and compute to run it. Deploying these models in real-world settings is a different problem.
Physical environments are not constant. Lighting changes between shifts. Conveyor belts jam. A person walks into the path of a robot. None of this appears in training data, and when it happens, it forces the operation to a halt.
Customers pay for uptime, throughput, and not having to put a human on the floor every time something unexpected happens. The model cannot guarantee any of those things on its own. The orchestration system that sequences tasks, catches failures, and gets the operation back on track has to. Those functions determine whether a site hits its SLA.
Model builders, including NVIDIA, DeepMind, and OpenAI, produce the intelligence but not the orchestration loop. Some stop at the API; others go all the way down to a robot foundation model. Either way, none of them runs the loop that keeps a warehouse running at 3:00 a.m. with no engineer on site. Model companies will not capture the value of deployment data because it’s generated in the execution loop, and they sit upstream of it.
Vertically integrated OEMs like Amazon Robotics and Locus Robotics and humanoid OEMs like Apptronik and Figure own the full stack. But their recovery logic is optimized for their own hardware in their own environments. The data they generate stays inside a closed system. It does not transfer to the heterogeneous, multi-vendor sites that most operators actually run.
Traditional systems integrators (SIs) look like platforms from the outside, but they’re not, because each project is custom. The fix that solved a sensor failure on site two lives in an engineer’s head, not in something site three inherits. The data stays in the project once it’s finished.
The capitalized SI is an integrator that owns the customer relationship and delivery accountability, with the orchestration layer embedded rather than bolted on. Each deployment feeds failure data, recovery patterns, and operational benchmarks back into a shared layer that every subsequent site inherits. It starts from everything the previous sites already learned, never from zero.
Most current robotics platform narratives describe software sitting on top of hardware without owning the data that the operation generates. That’s not a platform. It’s a dashboard. The companies that will define physical AI embed the orchestration layer inside a proven delivery model.
The model decides what should happen. The orchestration loop decides whether anything actually does. If a robot stops unexpectedly, the exception handler logs the downtime, the trigger, and the fix. The next time that failure appears, the recovery is faster. The site sees better throughput and lower cost per unit moved.
That data doesn’t exist until the system has run in real conditions. Synthetic data can inform the model, but it can’t replicate what the orchestration layer accumulates across live failures.
With the right orchestration layer, every site that goes live, be it greenfield or brownfield, runs on data a competitor has to generate from scratch. At 20 live deployments, this starts to show in uptime and throughput margins.
At 50 deployments, a new entrant isn’t just behind on features. They’re behind on every failure mode the platform has already seen and fixed. A better model can’t close that gap. Only time and deployments can, which is exactly what a new entrant doesn’t have.
Here are two live deployment examples I’ve seen in the field.
In the first, a robotic workflow produced optimal plans. A single jammed component forced a full system reset, repeatedly because no one had built what the exception handler should do when that happened.
Each reset cost roughly 15 to 20 minutes of downtime. Across a shift, that translated to 10% to 15% throughput loss. The model kept working. The operation didn’t.
In the second, dust and packaging debris interfered with sensors, triggering perception failures that cascaded across the warehouse floor. Restoring normal operation required a human to intervene each time.
The model was fine, but nobody had built the layer that catches a bad sensor reading before it takes down a line.
These are not edge cases. Physical environments are dusty, variable, and unpredictable. A system that needs human intervention every time one of these things happens is not a deployable product. Operators learn this fast.
Deployability is determined less by how often a system fails than by what happens when it does. A system that recovers autonomously in seconds beats a more capable one that waits minutes for a human, every time. Operators will pay a premium for a system that does not require them to staff for failure.
The orchestration layer is the investment. Not the model, not the hardware, not the integration services. The company that owns the orchestration layer owns the asset that compounds operational data across every live environment. I see two opportunities for investors. Pureplay horizontal orchestration, software that runs heterogeneous fleets across vendor sites, wins on fragmented, mid-market deployments where no single integrator controls the physical AI stack.
The capitalized SI wins on large enterprise accounts where customers want one counterparty and don’t want the deployment risk on their balance sheet. The vehicles differ; the asset doesn’t. In both cases, the orchestration layer is the moat.
The model builders have already caught on. They’re investing in components across the physical AI stack. For instance, Alphabet owns a foundational model in DeepMind, a skills library in Intrinsic, silicon chips in TPU, orchestration infrastructure via Google Cloud Robotics, and partners with humanoid OEM Apptronik.
Orchestration can not be as easily acquired as the other layers, and this is where model builders will compete. These organizations are flush with worldclass engineers who have never stepped foot inside a warehouse.
Pure-play horizontal orchestration providers bring the institutional knowledge of how to operate physical AI in live, variable environments. Model builders will look to acquire these players who can complete their missing domain expertise.
Value will concentrate in the orchestration layer. The delivery relationship is what puts you inside it. Everything else in this market will trade at what it is: a necessary input, not the asset itself.
Thomas Stapp is an investor, operator, and member of the executive leadership team at GreyOrange, where he leads strategy and corporate development. He is also an investor and advisor to Feather Robotics and serves on the board of Cartesian Kinetics. Prior to GreyOrange, Stapp held senior roles across management consulting at Bain & Co., principal investing, and C-suite leadership. His work focuses on the commercialization and deployment of robotics, physical AI, and industrial automation technologies at scale. Follow Thomas on LinkedIn.