Intel advanced SAP’s physical AI footprint with an architecture offering open choice across robotic systems and deployment environments.

In collaboration with the German enterprise resource planning (ERP) giant, Intel demoed a proof of concept centered around SAP’s high-volume warehouse operations management system, SAP Extended Warehouse Management (SAP EWM). As demonstrated at Intel’s Chengdu site in China, business context and workflow orchestration within the system were powered by robotics hardware and software with AI inference run on the edge. Enterprise requests were initiated in the SAP ERP and dispatched to autonomous mobile robots (AMRs), closing the loop with IT and operational technology (OT) working in conjunction.

In a recent interview with this title, Nagesh Puppala, GM of Intel's Physical AI and Robotics Segment, explained the chip giant sells robotics as a combined package of chips and developer software. The chips start with Core Ultra SoCs, with Intel offering 40 to 1,500 tera operations per second (TOPS), spanning from SoCs to discrete GPUs, with AMRs running on lower-end chips.

On the software side, Intel concentrates on deployment, the last stage of the robotics lifecycle following pre-training, simulation, and development.

Intel's package includes PyTorch, which allows models trained elsewhere to run on Intel CPUs, GPUs, and NPUs. Its Physical AI Studio allows engineers to fine-tune those models in a real deployment environment, while OpenVINO Physical AI connects the model to the robot hardware.

According to Puppala, Intel's current robotics strategy is "train anywhere, simulate anywhere, deploy on Intel", and its longer-term aim is enabling a turnkey platform for "any sensor, any model, any robot". The SAP collaboration adds ERP integration on top, as designed for cloud or on-premises deployment.

Physical AI on the edge

On the edge factor, Puppala told SDxCentral that the edge server is the likely middle ground between the robot and the cloud, but that the industry is yet to settle the question.

“There's a lot of debate of where the reasoning and those things should run,” Puppala explained. “I think most people will agree that it will not happen in the cloud because of data sovereignty and privacy issues.”

Intel's own factory robots use private 5G to keep traffic on site, with Puppala stating Intel “will never let our data out to somewhere on the cloud.

“Cybersecurity is a big issue. Someone takes control of that robot, it could be a disaster.”

While SAP's bread-and-butter nature as an ERP vendor gives it a considerable footprint in the warehouse space, the firm has gone beyond AMRs to dabble with more advanced humanoid machines. A January 2026 case study saw it link EWM to a humanoid robot executing picking tasks, with the automaton trained virtually using Nvidia Isaac Sim before deployment.

Intel rival Nvidia has various tentacles in the space, with its AI Enterprise suite powering robots for SoftBank on the RAN (radio access network), and its Nvidia Cosmos Reason vision-language model, designed for robotics and any device supporting computer vision, embedded in HPE Private Cloud AI as of last August.

Recent research from Dell'Oro Group contended that physical AI, alongside agentic AI applications, will drive cumulative spending on switches for AI back-end networks to $1 trillion by 2030. The firm described an "unprecedented pace of investment in accelerated computing to support the next wave of AI journey" as fueling a rapid expansion of the AI-centric networking market, including scale-up, scale-out, and scale-across networking.