Software Companies Are Building Robots Now — And Hardware Engineers Should Be Worried

Something curious happened this week in the convergence of AI and robotics, and it wasn't another humanoid demo or warehouse automation breakthrough. It was the steady drumbeat of software-first companies making moves that position them to own the intelligence layer of physical systems.

OpenAI rolled out health diagnostics assistance, helping physicians diagnose rare genetic diseases. Anthropic saw its most capable models pulled offline over dual-use concerns. Adobe integrated its Firefly AI directly into creative tools. On the surface, these look like typical AI product updates. But look closer at the implications for robotics, and a pattern emerges: the companies building foundation models are no longer content to be the brains behind someone else's robot.

Consider the Intel RealSense D585 Pro depth camera announcement. It's marketed as "AI-native" hardware for robotics, with on-device AI acceleration built in. But whose AI models will these systems run? The camera is hardware. The real value — and control — lives in the perception models, the decision-making layers, the continuous learning systems. Intel is providing the sensor. Someone else will provide the intelligence.

This is where OpenAI's health diagnostics work becomes relevant to robotics. The company isn't just building chatbots; it's building reasoning systems that can navigate complex, high-stakes decision trees in the physical world. Medical diagnosis requires weighing hundreds of variables, understanding context, and making life-or-death calls with incomplete information. Sound familiar? That's exactly what autonomous systems need to do.

Meanwhile, traditional robotics companies are still optimizing motion controllers and servo drives — crucial work, but increasingly commoditized. Elmo's new Maestro controller supports 256 axes. Impressive. But when the real differentiation happens in software that decides what those 256 axes should do, who captures the value?

The Anthropic situation is particularly telling. Their models were restricted not because of what they could say, but because of what they could do. "Dual-use capabilities" is government-speak for systems that can take physical action in the world. We're past the point where AI is just generating text and images. These models are being built to control things.

Traditional robotics companies have spent decades perfecting mechanical systems, safety protocols, and industrial-grade reliability. Those skills remain essential. But the layer that determines whether a robot can adapt to new tasks, learn from experience, or operate autonomously in unstructured environments? That's increasingly owned by companies whose core competency is software, not hardware.

The irony is that while foundation model companies race to control the intelligence stack, hardware manufacturers are still competing on technical specs: payload capacity, precision, speed. Important metrics, certainly. But ask yourself: would you rather own the motor or the brain that decides how the motor moves?

This isn't to say hardware expertise doesn't matter. The CMU research on training robots with internet videos shows that even sophisticated AI needs quality physical platforms to execute on. And projects like the Weedinator demonstrate that combining vision AI with mechanical systems requires deep integration.

But the balance of power is shifting. Software companies are moving down the stack into robotics. Hardware companies, meanwhile, are still largely licensing AI capabilities rather than building them in-house. When Kinova launches its KIMA medical robotic arm, the hardware is impressive. But the intelligence layer? That's increasingly a dependency on external AI providers.

Robotics companies that don't invest seriously in AI capabilities risk becoming contract manufacturers for the software giants. The companies that win won't just build better actuators or more precise arms. They'll control the full stack — or at least the parts of the stack that matter most for autonomous operation.

The next five years will determine whether robotics remains a hardware-led industry or becomes another sector where software companies own the margins. Traditional robotics manufacturers still have the advantage of domain expertise and existing customer relationships. But that window is closing.