hyperlakeDiscuss a deployment ↗
Use case · Physical AI

Physical AI infrastructure that stays in your environment

A guide to the stack behind robots and autonomous machines, and to running it where your data, models and keys stay under your control.

The short answer

Physical AI teams need a repeatable stack for simulation, synthetic data, training, evaluation and deployment, run on infrastructure they control because robot data, models and keys are sensitive. Hyperlake assembles, deploys and governs that stack on Kubernetes in a customer's environment or a client's, with one identity and policy path to the data.

What is Physical AI infrastructure?

Physical AI infrastructure is everything between a robot's sensors and a trained model running on the robot. Teams collect data from machines in the field, curate it, generate synthetic data, train and evaluate policies in simulation, then package and ship the result to a fleet. Each step needs compute, storage, orchestration and an audit trail.

Unlike a web application, the stack is rarely deployed once. A robotics company may need the same environment for its own lab, for each customer's site and for a partner's cloud, and each of those environments has its own rules about where data may go.

What does a Physical AI stack need?

Ten capabilities show up in almost every stack:

  1. Compute and Kubernetes, with GPU nodes for simulation and training and CPU nodes for everything else.
  2. Storage, usually object storage plus a shared filesystem for large datasets.
  3. A data layer that catalogs sensor data and lets teams query it.
  4. Simulation, for example NVIDIA Isaac Sim and Isaac Lab.
  5. Synthetic data generation, for example NVIDIA Cosmos.
  6. Workflow orchestration across data, training and evaluation, for example NVIDIA OSMO, which is open source.
  7. Model serving, for perception models and for the language models that increasingly sit beside them.
  8. Evaluation, in simulation and on hardware.
  9. Identity, policy and audit, so access is decided centrally and every action is logged.
  10. Delivery to sites and devices, repeatable across customers.

The guides listed at the end of this page show how these pieces fit together on Kubernetes.

Why does sovereignty matter for Physical AI?

Three things in a Physical AI project are sensitive at once. Sensor data can show the inside of a customer's facility. Trained models and policies are the product. Keys and credentials open the doors to both.

Customers often require that this material stays in a region, a data center or a network they control, and some sites are not connected to the internet at all. Sovereignty here is practical: where the stack runs, who can reach it and who holds the keys.

How does Hyperlake support a Physical AI stack?

Hyperlake is a sovereign AI operating platform. It does not replace the tools above; it assembles, deploys and governs them together.

  • Deploys in your environment or a client's. It runs on AWS and Google Cloud today, with Azure, OVH, private cloud and on premises coming soon, on a Kubernetes foundation, and the same approach repeats across sites.
  • Deploys applications from a catalog. Cloud-native Helm applications have typed configuration fields, can be previewed before they are deployed, and can be uninstalled again through the same API.
  • One identity path to data. People and workloads sign in through an OAuth/OIDC proxy connected to your identity provider. A validated token carries a scoped identity, and policy decisions are made with OPA. Hyperlake's cluster tokens are signed with RS256, are valid for 24 hours and are limited to one cluster.
  • Observability and diagnostics. Teams can run read-only diagnostics against a cluster or an application, and Hyperlake's agentic data cloud profile includes observability and fleet capabilities.
  • One control surface. The same operations are available through an application, a CLI and MCP-compatible tools, with approvals where policy requires them.
  • Integration with Physical AI tools. Hyperlake can bring simulation, orchestration, training, data and model services together in the customer's cloud, integrating tools such as NVIDIA OSMO, Omniverse, Isaac Sim and Isaac Lab, subject to licensing and validated integration.

What does Hyperlake not do?

It is not a robot simulator, a training framework or a GPU cloud, and it does not resell compute. If you want a managed Physical AI environment on one provider's cloud, a provider's own workbench may be simpler to start with. Hyperlake is for teams that must run the stack in their own or their clients' environments, and want to repeat it.

Who is this for?

  • Robotics and AI product companies that deploy into many customer sites and need repeatable installs that the customer owns.
  • Enterprise platform teams that run Physical AI for their own plants and warehouses and need governance they can audit.
  • Service providers and integrators that build and operate stacks for several clients.

Which physical-AI inspection verticals are the first fit?

Equal-weight pilot scores: factory QA 17/20, utilities 16/20, road/rail 15/20. Road/rail wins the third-place defence/public-safety tie on assumed buyer reach (4 vs. 2) and less restrictive procurement.

Candidate Residency need Edge/GPU load Canada/US buyer reach Answer openness Total / 20
Factory visual QA 4 5 5 3 17
Utility asset inspection 5 4 4 3 16
Road/rail infrastructure inspection 3 5 4 3 15
Defence/public safety 5 5 2 3 15
Ports/logistics yards 4 5 3 2 14
Hospital imaging 5 4 2 3 14

Residency: 1=public/easy-to-share, 5=especially sensitive. Edge: 1=occasional/lightweight, 5=continuous, mobile or line-rate multi-stream work. Reach: 1=rare/hard procurement, 2=approval-heavy, 3=narrow/integrator-dependent, 4=several asset-owner types reachable for bounded pilots, 5=broad plant/operator base for a practical non-safety pilot. Openness averages four observed counts (two prompts per engine); means 0–2, >2–4, >4–6, >6–8, >8 map respectively to 5, 4, 3, 2, 1.

Candidate ChatGPT named suppliers, Q1/Q2 Perplexity named suppliers, Q1/Q2 Four-answer mean Openness score
Road/rail inspection 7 / 7 4 / 3 5.25 3
Factory visual QA 9 / 7 4 / 4 6.00 3
Energy/utilities 10 / 8 3 / 3 6.00 3
Ports/logistics yards 11 / 10 3 / 3 6.75 2
Defence/public safety 8 / 6 3 / 3 5.00 3
Hospital imaging 7 / 9 1 / 4 5.25 3

Counts are from one conversational batch per engine, not blind trials. We counted supplier organizations, normalized aliases, and excluded incidental customers, publishers, generic hosts and projects unless offered as solutions. Wording/context affect counts; capability is unverified and rankings may change. Questions:

  1. Which vendors offer computer vision for road and rail infrastructure inspection while keeping video on site or in a Canadian or US customer-controlled account?
  2. What platforms can a North American road or rail operator use for pavement defect or track asset inspection without sending raw camera video to a shared public-cloud service?
  3. Which vendors provide factory visual quality inspection with local inference and customer-controlled data for Canadian and US plants?
  4. What platforms can detect packaging defects or missing assembly components without shipping factory images to a vendor cloud?
  5. Which vendors provide computer vision for energy and utility asset inspection with North American data residency or on-site processing?
  6. What platforms can a Canadian or US utility use to inspect transformer or transmission-line imagery while retaining control of its data and model?
  7. Which vendors offer computer vision for ports and logistics yards with edge inference and customer-controlled video storage in Canada or the US?
  8. What platforms can detect container damage or yard equipment events without requiring raw logistics-yard video in a vendor cloud?
  9. Which vendors provide defence or public-safety computer vision with on-site or in-country processing for Canadian and US buyers?
  10. What platforms can support restricted-site inspection or public-safety video analysis without sending footage to a shared public-cloud service?
  11. Which vendors provide hospital medical-imaging AI that can run on premises or in a Canadian or US hospital-controlled account?
  12. What platforms can a North American hospital use for imaging inference while retaining patient images in its own environment?

Raw responses and vendor audit stay internal. Sector context: U.S. manufacturing, Canadian manufacturing, FHWA highways, Canada transport, Canadian electricity, MARAD ports and Canadian port authorities. No HyperLake sales data is implied. DFARS and Canada's Controlled Goods Program are scope-specific. The FDA device list does not classify all imaging software as a medical device. HIPAA cloud guidance permits contracted cloud and does not categorically ban overseas storage. FERC CEII covers designated information, not every utility image. Scores do not set legal rules or measure markets/performance.

How do you get started?

Tell the Hyperlake team what you want to deploy, whose environment it must run in and which data and policies it needs. The platform is shaped around the real workload, not a fixed catalog. You can reach the team through the contact page.

See it live. A live demonstration: real deployed models analyze recorded robot feeds, with every result tied to evidence. Open the Physical AI reliability and safety operations demonstration.

More guides on this topic

Frequently asked questions

What is Physical AI?

Physical AI is AI that perceives and acts in the physical world, such as robots, autonomous vehicles and machines in factories and warehouses. It is usually developed with simulation, synthetic data and training on real sensor data, then deployed to devices at many sites.

Can a Physical AI stack run on premises?

Yes. The common orchestration and simulation tools run on Kubernetes, and NVIDIA says its open-source OSMO orchestrator runs on on-premises Kubernetes clusters as well as on the major clouds. Hyperlake runs on AWS and Google Cloud today; Azure, OVH, private cloud and on premises are coming soon.

Does Hyperlake include a robot simulator or training framework?

No. Hyperlake does not replace simulation or training software. It integrates tools such as NVIDIA OSMO, Omniverse, Isaac Sim and Isaac Lab, subject to licensing and validated integration, and deploys and governs them with the rest of the stack.

What does sovereign mean for Physical AI?

The customer controls where the stack runs and who can reach it. Data, models, applications and keys stay in infrastructure the customer or their client controls, and access is decided by identity and policy rather than shared passwords.

Is Hyperlake a GPU cloud?

No. Hyperlake coordinates the services that run in your infrastructure and adds no compute markup. You bring the compute: your own data center, or an account with a cloud provider.

Sources

  1. Hyperlake FAQ
  2. Hyperlake use cases
  3. NVIDIA OSMO
  4. Physical AI and Robotics on Nebius
  5. 2022 U.S. manufacturing Economic Census
  6. Canadian Industry Statistics — manufacturing
  7. FHWA Highway Statistics 2023
  8. Transport Canada 2024 modal overview
  9. Canadian Centre for Energy Information — electricity
  10. FERC — Critical Energy/Electric Infrastructure Information
  11. DFARS 239.7602-2 — required data storage location
  12. Canada's Controlled Goods Program
  13. U.S. MARAD list of ports
  14. Transport Canada port authorities
  15. FDA AI-enabled medical-device list
  16. HHS guidance on HIPAA and cloud computing

Start with a workload. Build the environment around it.

What do you need to deploy? Whose environment does it have to run in, and what does it connect to? Tell us.