Two facts sit uncomfortably next to each other. Nvidia’s ambition is to supply the compute substrate for the world’s robots — a single vendor providing the nervous system for embodied machines everywhere. And according to the Wall Street Journal, China is among the most enthusiastic customers for exactly that.
Those two things are not in conflict commercially. They are in tension architecturally, geopolitically, and — the part I find most interesting as someone who studies agent systems — structurally. Because the question of who supplies robot compute is not a question about chips. It is a question about where intelligence lives, and who controls the layer beneath it.
Why robots are a different compute problem
Training a large model and running a robot are related but distinct workloads. Training is batch work: enormous, parallel, tolerant of latency, happy to live in a datacenter. Robot inference is the opposite. It is a closed loop against physics, running at whatever frequency the control problem demands, with a latency budget measured against a falling object or a moving human.
That difference drives a hardware split you can see in how the industry organizes itself. Datacenter accelerators optimize throughput per rack. Embedded robot compute optimizes deterministic latency, thermal envelope, and power draw against a battery. A humanoid or a mobile manipulator cannot phone home for every decision, which means a meaningful fraction of the policy stack has to run onboard.
So when a vendor talks about running the world’s robots, they are talking about owning three things at once:
- The onboard inference silicon that executes learned policies in real time
- The simulation and training infrastructure where those policies are produced
- The software layer that connects the two — the runtime, the model formats, the tooling
The third item is the one people underweight. Silicon can be substituted. Toolchains build habits, and habits build lock-in.
The agent architecture angle
Readers here care about agent design, so let me place this where it belongs in that conversation. Embodied agents are the hardest version of the agent problem. A text agent that plans badly wastes tokens. An embodied agent that plans badly breaks something or hurts someone. The consequence gradient is steeper, the feedback is noisier, and the state space includes the messy physical world rather than a tidy API surface.
That has practical implications for how these systems get built. A hierarchical split is the standard answer: a slower deliberative layer that reasons about goals and sequences, and a fast reactive layer that handles control. The slow layer might be a large multimodal model. The fast layer is closer to classical control or a compact learned policy. Getting the handoff right between them is, in my view, the central unsolved engineering problem in embodied agents.
Now notice what a compute vendor controlling the whole stack does to that problem. It standardizes the split. Whatever boundary the vendor’s tooling makes easy becomes the boundary most teams adopt, not because it is optimal for their task, but because it is the path of least friction. Architecture follows tooling. That is a real force in this field, and it deserves more scrutiny than it gets.
Eager customers, contested supply
China’s appetite for this hardware is the part that turns a technical story into a strategic one. Robot compute sits at an awkward intersection: it is dual-use by nature, it is manufactured by a small number of firms, and the demand side includes manufacturing economies with enormous scale needs. An eager customer is a good problem for a vendor’s revenue and a complicated one for everyone modeling supply chain risk.
The structural consequence is worth thinking through. When a critical layer has one dominant supplier and a contested customer base, you get pressure in two directions. Buyers hedge by developing domestic alternatives. Suppliers hedge by segmenting products across markets. Neither dynamic produces a clean, unified global platform. Both produce divergence — different silicon, different toolchains, different assumptions baked into the runtime.
For anyone building embodied agents, that divergence is the thing to plan around. If your policy stack assumes one vendor’s runtime, one memory model, one set of quantization behaviors, you have accepted a dependency that may not travel across markets or across years.
What I would watch
I would watch the abstraction layers. The most valuable defensive engineering in embodied AI right now is not squeezing more frames per second out of a specific accelerator. It is keeping the policy representation portable — clean separation between the learned components, the control loop, and the hardware backend. That work is unglamorous and it does not demo well.
It is also what determines whether the next decade of robotics produces a diverse ecosystem of embodied agents or a monoculture running on someone else’s nervous system. Ambition to supply the world’s robots is a reasonable business goal. Whether the field should accept a single answer to a question this foundational is a separate matter, and one that engineers, not just procurement departments, get to influence.
🕒 Published: