Two Ways to Win the Robot Race: Smarter vs. Cheaper
TrendForce · 2026
"The US and China are running two structurally different playbooks for humanoid robots — the US is betting that intelligence (AI models trained on real-world data flywheels) will be the deciding moat, while China is betting that supply-chain scale and manufacturing speed will out-iterate any intelligence gap, echoing how China caught up in electric vehicles."
US firms like Tesla, Figure AI, Boston Dynamics, and Apptronik are racing to make their robots' AI smarter through real-world data collection. Chinese firms like Unitree and AgiBot are racing to make robots cheaper and more numerous, scaling one firm's production from 1,000 to 10,000 units in about sixteen months.
TrendForce frames this as a rerun of the electric-vehicle playbook: build a complete domestic supply chain, localize core components (servo motors, reducers, batteries) quickly, drive costs down, and win on volume and iteration speed rather than a single technological edge. The US counter-bet is that hardware specs matter less than they used to — value is shifting to the AI layer (NVIDIA's Cosmos and Isaac Lab, Google DeepMind's Gemini Robotics), and whichever company accumulates the most real-world interaction data fastest compounds an intelligence advantage that manufacturing scale alone can't close. TrendForce's own supply chain index (CSCII) shows this split concretely: the US dominates the 'Mental' plane (AI chips and compute, via NVIDIA and Qualcomm's near-monopoly), while China has the most comprehensive presence across all four component categories and leads the 'Power' plane outright.
What is the central strategic difference TrendForce identifies between the US and Chinese approaches to humanoid robots?
Read more about the topic
The explanation above is written with AI assistance. These are the originals — go to them to check it.
- Humanoid Robots Part 1: The US-China Divide and Who Controls the Supply ChainTrendForce Insights (Substack)
The Scaling Hypothesis
"Intelligence may emerge from scaling up compute and data. GPT-3 shows the hypothesis holds."