Compute vs data: where the robotics bottleneck really is

The 100-Watt Paradox

An H100 GPU pulls up to 700 watts of electricity at peak load. A human brain operating a pair of hands to thread a copper wire through a surgical needle pulls roughly 20 watts. If the physical AI revolution were purely a question of compute vs data, the path forward would be simple: stack more silicon, scale the cluster size, and wait for emergent reasoning to solve the physical world. But training a humanoid robot to fold a shirt is not like training a Large Language Model to write Python code.

We have reached an inflection point where the compute vs data debate in robotics has been settled by physical reality. While frontier AI labs run massive clusters to train next-generation multimodal models, the developers of physical AI are hitting a different kind of bottleneck. We do not lack the tensor cores to run inference on a 10-billion parameter Vision-Language-Action (VLA) model. We lack the high-quality, high-frequency, multimodal physical demonstrations required to ground that model in the chaotic, non-linear dynamics of our world.

The bottleneck in humanoid robotics is not the compute budget. It is the physical data layer.

The Math of Physical Scaling

To understand why this bottleneck exists, we must look at the raw data requirements of modern robotic foundation models. Large language models train on trillions of tokens scraped from the open internet. Text is cheap, static, and abundant. Physical AI models, such as those discussed by Physical Intelligence in their work on generalist robot policies, require continuous trajectories that pair high-resolution visual inputs with precise force, torque, and proprioceptive motor actions.

A typical human demonstration captured at a facility requires recording multi-angle RGB video at 30 to 60 frames per second, synchronized with robot joint positions, velocities, and tactile feedback at 100 Hz to 240 Hz. When we look at standard open-source datasets like the Open X-Embodiment dataset or the DROID dataset, we see a collection of diverse but highly fragmented trajectories. While these datasets are invaluable milestones, they represent a drop in the ocean compared to the scaling laws that governed the transition from GPT-2 to GPT-4.

The hard truth of physical AI is that simulated data cannot bridge the gap alone: high-fidelity human demonstrations remain the ground truth for real-world generalization.

Why Simulation is Not a Silver Bullet

A common counter-argument is that we can bypass the physical data bottleneck by training robots inside simulated environments. Platforms like NVIDIA Isaac GR00T have demonstrated immense success in teaching robots basic locomotion and control policies through reinforcement learning in simulation. However, simulation suffers from the "sim-to-real" gap.

While simulators can model rigid-body physics with high accuracy, they struggle with the infinite complexity of contact mechanics, deformable objects, fluid dynamics, and optical noise. A robot trained purely in simulation to grasp a plastic bottle may fail in the real world because it cannot accurately predict how the thin plastic deforms under its fingertips, or how the liquid inside shifts the center of gravity. To cross this gap, models must be fine-tuned or pre-trained on real-world egocentric data, such as the multimodal trajectories found in the Ego4D and Ego-Exo4D datasets.

The Dimensionality of Action Space

In language, the action space is discrete: the model chooses the next token from a fixed vocabulary of perhaps 100,000 words. In physical manipulation, the action space is continuous and multi-dimensional. A humanoid robot with two arms, two hands, and a torso has dozens of degrees of freedom. At every millisecond, the controller must output precise torque values for every single actuator. A tiny deviation in one joint propagates through the kinematic chain, changing the entire physical state of the system.

The Shift to End-to-End VLA Models

This complexity explains why the industry is moving away from classical, hand-coded control loops and toward end-to-end learning. The trend is evident in the work of companies like 1X Technologies and Skild AI, who are building general-purpose physical agents that map sensory inputs directly to motor outputs.

These models perform best when they are fed massive quantities of human-demonstration data. Humans possess intuitive physics: we know how much force to apply to open a stuck drawer, how to catch a falling object, and how to balance our weight on a slippery surface. Capturing this tacit knowledge through egocentric, multimodal teleoperation or wearable capture systems is the only known way to bootstrap generalist robotic systems.

The Path to Physical Scaling

If we want robots that can operate safely in hospitals, warehouses, and homes, we must treat physical data collection with the same engineering rigor we apply to GPU cluster design. The bottleneck is no longer the speed of our neural networks; it is the scale, diversity, and fidelity of our training data. Only by building dedicated infrastructure to capture the richness of human touch and movement will physical AI move from lab demonstrations to reliable, deployed systems.