Edge AI chips are everywhere. The software to use them isn't.
AI chips that run on the robot itself are finally fast and cheap enough for real work, but the software layer that decides who can build with them has not caught up.
AI chips that run on the robot itself are finally fast and cheap enough for real work, but the software layer that decides who can build with them has not caught up.
The robotics industry just hit the same wall the PC industry hit in the early 1980s. Hardware got fast and cheap before the software caught up, and for a while the only people who could actually use the machines were the ones who built them. Edge AI has flipped that problem, or at least pushed it up the stack. AI accelerators from NVIDIA, AMD, Qualcomm, and Hailo can now run perception, planning, and control models on a robot's own board, with no cloud round-trip. The chips are ready. The software between those chips and a person who wants to build with them is not.
That gap is the story, and the reason the PC analogy keeps coming up. In a recent column, The Robot Report argued that edge AI software layers are doing for robots what Windows did for PCs: moving robotics from a niche engineering craft to something a much wider set of teams can actually use. It is a useful frame, and a leaky one. The PC analogy works because it captures a real shift, the move from a world where using the machine meant writing the machine, to a world where most of the plumbing is hidden behind a layer people can compose with. The analogy leaks because Windows was not inevitable, and no single "Windows for robots" is either.
A concrete example of the layer question sits at Numurus, a Seattle-area company that builds NEPI, the Numurus Edge Platform Interface, on top of ROS and ROS 2. The company describes NEPI as handling roughly ninety percent of robotics software, including hardware driver recognition, AI model management, an automation engine, data collection, and a browser-based UI, so that teams can focus on the ten percent that is actually their logic, models, or application. The product is source-available and free to try without an account, which on its own says something about who the company thinks the next wave of users is. As one trade publication argued, the pitch is that edge AI software layers are making robotics broadly usable the way a graphical operating system once made personal computing broadly usable (The Robot Report).
The friction is real, and it is structural, not cosmetic. Current edge AI stacks mostly assume a desktop PC's ergonomics: a keyboard, a mouse, a display, a file system, a word processor on the other end. A robot has cameras, motors, lidars, GPS receivers, and live sensor streams feeding models that drive actuators, often with no screen at all. Numurus pitches plug-and-play sensor, camera, and actuator integration, edge AI model deployment, low-code event-driven automation, and browser-based monitoring, all aimed at making that loop feel less like writing device drivers from scratch (Numurus). The first customer reference on the company's own site, Roger Fellows, president of Inov8v Marine Group, says they delivered smart sonar sensors to customers within five months using NEPI and Numurus professional services. That is a vendor-published testimonial, not independent validation, but it is also the kind of timeline that, if it holds across more than one shop, would matter.
The interesting question is not whether NEPI in particular wins. It is whether the layer above the silicon settles into a single dominant platform, a handful of open-source stacks, or a thicket of vendor-specific toolchains that each lock in their own customers. The PC answer was a near-monopoly for about fifteen years, then a contested duopoly, then a mobile platform shift that broke the whole frame. The robot answer could move in any of those directions, and it will be decided by software choices that are being made right now inside engineering teams who would rather not have to write a new sensor driver every time they swap a camera.
Independent analyst data on how wide the Linux-on-chip gap actually is, and how many non-Numurus teams are closing it, would make the load-bearing claims in this story sturdier. Without that, the safe read is narrower: edge AI silicon has clearly crossed an inflection, the software layer above it is the new bottleneck, and whoever defines that layer is the one deciding who gets a seat at the robotics table.