From models to the physical world – what it takes to scale Physical AI
Physical AI is moving AI into warehouses, factories, robots, vehicles, and other environments where software has to interact with the physical world. That means models are learning from real-world data, updates need to reach devices in different locations, and some decisions have to happen within milliseconds.
We spoke with Dr. Xiong, Physical AI technology lead at Amazon Web Services (AWS), about where Physical AI stands today, what’s needed for it to land and scale, and what companies need to plan for as they move into production.
Here are seven takeaways from our chat.
1. Physical AI goes beyond embodied intelligence
Embodied intelligence focuses on giving a robot the intelligence to perceive and act within an environment. Humanoid robots are one of the most visible examples.
Physical AI is broader. According to Xiong:
“A Physical AI system needs to interact with the physical world. It has to perceive its environment, make a decision based on what it perceives, and then take an action that changes that environment.”
The system also has to learn rather than follow fixed preprogrammed instructions.
AWS describes this as a Physical AI flywheel: a device collects data from the physical world, that data goes into model training and simulation, the updated model is validated and deployed back to the device, and the cycle starts again.
“Every time that cycle completes, the system evolves.”
Xiong says agentic AI can automate more of that loop, including data labeling, model training, simulation, validation, and controlled deployment.
2. Different Physical AI use cases are at very different stages
Xiong groups Physical AI use cases into three levels of maturity.
The first includes applications already running at scale, such as autonomous mobile robots in warehouses and logistics environments and inspection robots in industrial facilities. Amazon operates more than one million robots across its facilities and logistics network.
The second includes systems that work well in one environment but struggle when conditions change. For instance, a quadruped or wheeled robot might be able to perform inspection or security tasks inside a factory but its perception and planning systems may need to be tuned before deployment at a construction site.
“The system may scale within one scenario, but once the environment changes, the perception and planning systems often need to change with it.”
General-purpose humanoid robots sit at the third level. Progress is moving quickly, and Xiong points to companies now capable of shipping more than 5,000 units for certain use cases.
But reliably completing complex tasks across unstructured environments at scale is no easy feat. Variations in lighting, surfaces, objects, movement, and human behavior can all make it harder to collect useful data and generalize across environments.
3. Global deployment adds new data, operational, and infrastructure requirements
Operating Physical AI across regions adds data compliance, fleet management, local infrastructure, and network requirements.
The first is data governance and compliance. Robots and physical devices can collect visual, spatial, force, telemetry, and other data. Europe places strict requirements around personal data like imagery containing identifiable faces and how that information can be transferred across borders. Some Middle Eastern markets have data-localization requirements, while requirements across Southeast Asia can vary significantly between countries.
Xiong stresses that “compliance needs to be part of the architecture from day one.”
The second is fleet operations. Once thousands, or eventually tens of thousands, of devices are deployed across regions, companies have to manage software versions, model updates, device health, telemetry, failures, and ongoing operations across the fleet.
The third is market adoption. Local systems integrators might hesitate to build around a newer platform while end customers wait until it has been proven in similar environments.
The fourth is networking and infrastructure. Physical AI devices exchange information with cloud and edge infrastructure for data backhaul, model distribution, OTA updates, monitoring, and coordination. For some real-time use cases, the difference between 100 milliseconds and 500 milliseconds of latency can materially change system behavior.
4. The network has to connect three layers of intelligence
Physical AI is more sensitive to latency than many digital AI applications. An extra second or two from a large language model isn’t a big deal, but a warehouse robot that’s 200 milliseconds late avoiding an obstacle could hit a shelf…or worse.
Xiong describes Physical AI using a three-layer cognitive architecture.
- The cloud-based superbrain handles complex reasoning and longer-horizon planning and can tolerate response times measured in seconds.
- The brain turns instructions or perceived conditions into more immediate planning decisions that may need to happen within hundreds of milliseconds, usually closer to the device.
- The cerebellum handles motion control and safety, making decisions within milliseconds that need to remain on-device or at the immediate edge.
“The network is the nervous system connecting those three layers.”
The network supports three parts of the Physical AI flywheel: data backhaul from devices into training and analysis systems, model distribution and OTA updates back to the fleet, and real-time coordination between cloud and edge intelligence.
Some decisions can run in the cloud. Others have to stay at the edge or on the device. The infrastructure has to match those response requirements.
5. How Zenlayer supports Physical AI deployments
As Physical AI expands, companies need infrastructure for model access, inference, data movement, and connectivity across distributed environments. We’re helping Physical AI companies address those needs in three areas:
Model access
Physical AI developers may use their own vision-language-action (VLA) or physical world models while also calling foundation models from multiple providers.
For one embodied AI company, researchers needed access to both proprietary models and a broad range of external models. Zenlayer AI Gateway and MCP Gateway gave them a centralized way to access those models, orchestrate calls across them, and monitor usage.
The company reduced model management and operational costs by nearly 80%.
Distributed GPU compute
Self-built infrastructure may cover development and initial deployment in one location, but it’s much harder to extend that capacity everywhere demand grows.
Our globally distributed infrastructure can provide GPU resources closer to the environments and users that need them, giving Physical AI companies more flexibility in where they deploy inference and how they allocate compute across regions.
Multi-cloud data access
Physical AI and industrial AI datasets may sit across clouds while researchers in different regions need access.
For one customer, we combined bare metal compute, private connectivity, and cloud connectivity to create a more direct path between those environments. Researchers in different regions could access pretrained datasets while continuing to work in the cloud environments they already used. The architecture reduced regional egress costs, and the new AI infrastructure platform was delivered within days.
Across these use cases, the common requirement is connecting cloud, edge, and device environments while accounting for where data can move, where compute needs to run, and how quickly those resources need to be available.
That is where our distributed compute and network infrastructure complements AWS capabilities across AI compute, simulation, compliance, and ecosystem support.
Customers can also access our services like Zenlayer AI Gateway and global acceleration offerings through AWS Marketplace, and AWS and Zenlayer have worked together across multiple joint customer deployments.
6. What’s next for Physical AI
Drawing on his experience at AWS, Xiong sees Physical AI moving through three stages: building the system, expanding into new markets, and operating it at global scale. As companies move through those stages, the business model may evolve as well.
He expects more companies to move from one-time hardware sales toward Robotics as a Service, or RaaS, where customers pay for ongoing outcomes instead of the robot itself. That also makes consumption-based infrastructure a more natural fit as compute, storage, and network requirements rise and fall with usage.
“The leading companies will increasingly sell intelligent outcomes in the physical world, rather than simply selling hardware.”
Amazon’s own experience operating more than one million robots also gives Xiong a view into what that scale looks like in production. Managing a successful system in one environment is one thing, but operating tens of thousands of devices across countries, infrastructure environments, and regulatory requirements is another entirely.
7. How infrastructure providers can support Physical AI
AWS supports Physical AI companies across the three stages Xiong outlined:
- Building: Support the development cycle from data collection and model training through simulation, validation, and deployment, with agentic AI increasingly automating processes such as data labeling, training, simulation, and controlled rollout.
- Expansion: Help companies bring Physical AI systems into new markets through global infrastructure, AWS Marketplace distribution, solution architecture support, and connections with customers, partners, and investors.
- Operations: Support fleets running across regions with compliant global infrastructure, cloud-edge coordination, device management, and consumption-based infrastructure that can support business models like Robotics as a Service.
AWS is one example of how infrastructure providers can support Physical AI as deployments grow.
Infrastructure providers have a critical role in delivering the compute, connectivity, edge-cloud coordination, and global infrastructure needed to move data and models between environments while accounting for different latency, operational, and regulatory requirements to ensure Physical AI systems can operate reliably at scale.
The next advantage is how fast you can turn the flywheel
As Physical AI deployments scale, the data generated in real environments will be one of the hardest advantages to replicate.
High-quality Physical AI data is difficult to collect, and companies with larger deployed fleets can use it to improve their models, push those improvements back into production, and start the next cycle with even more data.
Xiong sees the next three years as an important window for building that advantage.
As he puts it:
“The competition in Physical AI comes down to who can get the flywheel turning first, keep it turning, and make it run faster, better, and across more environments.”
To keep that flywheel moving across more environments, the infrastructure underneath it has to keep up. We help Physical AI companies connect cloud, edge, and device environments so they can move models, data, and compute where they’re needed as deployments grow.
Connect the infrastructure behind Physical AI
Use Zenlayer AI Gateway to simplify access to leading models, and Zenlayer Cloud Networking to move your data securely and efficiently from where it’s generated to the cloud, data center, or compute environment where it needs to be processed and used for training.
Start building and scaling AI applications faster with zenConsole today.



