Robots Have Fixed Hardware. That's Why Their AI Stays in the Billions While LLMs Hit Trillions.

In LLMs, the model comes first and the hardware serves it. In robots, the on-robot hardware is fixed and the model has to fit. That inversion is the whole reason today's robot foundation models sit between roughly 5 and 14 billion parameters, while frontier language models push into the trillions.

The contrast shows up in concrete models. Physical Intelligence's π0.7 runs at about 5 billion parameters; NVIDIA's DreamZero lands around 14 billion. Frontier LLM labs have stopped reporting parameter counts because the numbers have become unprintable. Both are doing serious work, but they are doing it at sizes separated by two orders of magnitude, and the gap is not a temporary lag. It is structural.

Per SemiAnalysis's "Where Does a Robot Think" analysis, robotics inverts the design order that has powered the LLM buildout. In the LLM playbook, a hyperscaler (one of the handful of cloud giants like Microsoft, Google, Amazon, or Meta that train at frontier scale) picks a target model size, then designs the datacenter, the network, and the power budget to serve it. In embodied AI, the class of models that perceive and act in the physical world, and whose largest slice is robots, the robot ships first. Its on-board compute, its battery, its thermal envelope, and its bill of materials are locked in by the manufacturer, often before any model team gets a seat at the table. The model has to fit what is left.

Two binding constraints define the shape of that leftover room.

The first is real-time control. A humanoid robot does not get to pause and think. Its sense-think-act cycle, the loop where it reads sensors, plans the next motion, and sends it to the motors, has to close inside a deadline measured in tens of milliseconds, because the world has moved on if it takes longer. A control loop is non-negotiable: missing a deadline does not just slow a robot down, it makes it fall down. Offloading inference to a datacenter can work for some decisions, but the control loop has to land on the robot. The model has to be small enough to run inside that budget every single cycle.

The second is per-unit capex. The manufacturer pays for the compute on every robot it ships, not once in a cluster. A datacenter-scale language model can amortize its training cost across billions of user queries; a robot model has to be cheap enough to fit inside a $30,000 home helper or a $200,000 warehouse unit without blowing the bill of materials. Multiply that out and the unit-economics math becomes the whole game. Per SemiAnalysis's framing, in embodied AI "intelligence becomes something you ration against latency and unit economics," not something you scale freely.

5 to 14 billion parameters is roughly what fits on current edge AI accelerators like NVIDIA's Jetson Thor while still meeting a control-loop deadline and a per-unit cost target. LLM scaling laws, the empirical finding that bigger models trained on more data keep getting more capable, do not stop applying to robots. They just stop applying within the budget the hardware leaves over.

The trillion-robot line that follows from this is a what-if, not a forecast. If the world ever does ship a billion humanoid robots, and each one carries even a modest slice of the embodied-AI buildout's total upfront spend, the bill lands in the trillions. Most of that money goes into the robots themselves and the factories that make them, not the compute inside them. That is SemiAnalysis's scenario, and it is the clearest way to see what the design inversion actually buys: when hardware is fixed, intelligence is what gets rationed.

Figure AI's Helix and the π0 family at Physical Intelligence are interesting for the same reason: they target the on-robot budget directly, not because they are racing frontier LLMs. A robot model that gets the control loop right at 5 billion parameters is doing more useful work than a 70-billion-parameter model that has to live in a datacenter. The build problem is different, and the design order shows it.

The next test is whether the edge-AI silicon catches up. Jetson Thor is one step. Whatever ships after it is what determines whether the parameter range moves from 14 billion toward 30, or stays pinned where the unit economics pin it. That is the watch item, and it is the one any future robot announcement will turn on.