How Physical AI Will Make The Network Part Of The Machine
Venkatesh Gokulraju, technical director of forward deployed-engineering at Lancesoft, specializes in AI-driven network automation.
gettyPicture what a regional airfield could look like around 2030.
An autonomous baggage tractor loses satellite navigation to interference, yet keeps working. The cellular network carries its commands, helps track its position, detects a drone near the apron and provides a trusted clock. When the tractor moves beyond terrestrial coverage, a satellite keeps its remote supervisor connected, and the machine shifts into a more cautious mode.
In my work leading telecom product development and forward-deployed engineering across 5G, non-terrestrial networks (NTNs), network automation and service assurance, I work at the boundary between lab design and field reality.
The building blocks for this scenario—edge inference, network sensing, closed-loop automation, satellite connectivity and resilient timing—already exist or are moving through standards and product road maps.
Together, these capabilities point to one of the biggest telecom redesigns since 5G: networks that stop only carrying data for machines and increasingly participate in how those machines sense, decide and act.
But challenges remain for bringing this vision to life, many of which are organizational, as operations and security teams must increasingly engineer the robotics, network and cloud capabilities as one system.
Factories are on track to install more than 700,000 industrial robots a year by 2028, according to the International Federation of Robotics, and mobile machines are spreading from factory floors to yards, ports and airfields.
Unlike generative AI, which answers on a screen, physical AI runs a continuous loop of sensing, computing and acting through a real body.
While networks were traditionally scored on coverage, capacity and throughput, physical AI scores them on whether the right decision arrives while it still applies. No machine can carry every computation it needs, and no distant cloud is fast enough for every reflex, so intelligence is splitting across the machine, edge and cloud or satellite layers.
• Radio sites will become AI sites. AI-native radio networks will share processors between wireless functions and AI inference, turning cell sites into compute locations close to machines.
• The edge will become a decision layer. Camera and lidar feeds will increasingly be fused at the edge, while machines exchange task-relevant meaning rather than raw video, reducing the load on every link.
• The core will become an orchestrator. Network analytics can already predict whether service quality will hold, allowing workloads to move before a link degrades.
• Determinism will become a product. Network slicing, ultra-reliable service, time-sensitive networking and precision timing will increasingly be delivered as guaranteed performance requested through software.
• Satellites will become intelligent infrastructure. Regenerative payloads will run network functions in orbit, while laser links route traffic between satellites with AI helping to allocate capacity in space.
The most underestimated change is that the network itself will become a sensor.
The first 6G standards, scheduled to be finalized in early 2029, build on work on integrated sensing and communication, which uses radio signals to detect and track objects such as drones, much as radar does.
For fleet operators, machines could gain perception beyond their own line of sight. A forklift could be warned about a person its cameras cannot yet see.
The catch is assurance: Network-sensed data will need the same validation as onboard sensors before a machine is allowed to act on it.
Satellites are usually bought as backup coverage. In a physical AI architecture, they become a distinct operating tier with their own physics.
Standards studies put the maximum round trip from propagation alone at 25.77 milliseconds for a low-orbit satellite at 600 kilometers and more than half a second, 541.46 milliseconds, for a geostationary one.
The realistic future is not satellites running split-second control. It is machines that know which tier they are on and adapt: continuing approved local work, reducing speed or autonomy, deferring nonessential data or pausing tasks requiring continuous supervision.
As NTNs become more deeply integrated into 5G-Advanced and future 6G architectures, those transitions should become more systematic.
That said, two lesser-known dependencies deserve board-level attention.
The first is timing and position. GPS signal-loss events in aviation rose 220% between 2021 and 2024. Machines that rely only on satellite navigation inherit that exposure, which is why standards bodies have studied cellular networks as a resilient time source.
The second is cryptography. A draft federal transition plan proposes deprecating today’s public-key algorithms after 2030 and disallowing them after 2035. Machines purchased this year may still be working then. Post-quantum keys are many times larger than today’s. Also, over a satellite hop, bigger handshakes can add costly round trips. Test them on real links now.
Based on my experience, one of the biggest remaining hurdles is what happens during integration and validation. Difficult failures often appear at system boundaries: a handover leaves data stale, an edge node overloads, a route becomes asymmetric, timing drifts, a satellite path adds delay or an automated recovery action conflicts with policy.
Organizations that manage robotics, networking, cloud, security and operations independently can miss these cross-domain failures until deployment. Here are a few ways to manage that challenge:
• Write a decision budget for each autonomous task: deadline, data freshness, degraded behavior and recovery.
• Rehearse transitions in a digital twin, then on-site: cut links, overload the edge and switch to satellite.
• Contract for measured outcomes, such as on-time decisions and recovery time, alongside coverage and speed.
• Require crypto agility and timing resilience in every machine and network purchase.
• Name one owner across operations, robotics, networking and security.
By 2030, telecom will be judged less by how many people it connects and more by how safely machines complete useful work under changing conditions. To scale physical AI with confidence from the factory floor to the airfield to orbit, leaders must now start designing their networks as part of the machine.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
