Humanoid Robots Are Following AI's Split Into Blocs and Open Weights

Humanoid robots are no longer a separate story from AI. The robot market is splitting into blocs along the same lines as compute, taking in open-weight models the way software did, and running into regulators who so far have written rules for models and machines separately, not for machines that act on their own.

Key points

• US rules now block foreign-made humanoids the same way they block chips, so the US-China split in AI compute is being copied into robot bodies.
• Rare-earth magnets do for robot actuators what advanced chips do for AI: they are the physical chokepoint, and China holds it.
• Open-weight robot models, together with bodies costing under $10k, spread capability faster than any import ban can contain it.
• EU AI rules regulate models and treat AI agents as unsettled, while the rules for AI in machinery are delayed to 2028, so a humanoid working in a shared space falls between the two.
• The likely result is two robot ecosystems with different standards, and nobody overseeing both.

Where the lines cross

Humanoids and compute geopolitics

In AI compute, the split is well advanced. The US has closed export-control loopholes and extended them to cloud access¹, while DeepSeek is moving its training onto Huawei Ascend and Huawei plans to roughly double output². Robots now show the same pattern. China shipped about 97% of the world's humanoids in the first half of 2026 and has published its own standard system³. Unitree and AgiBot are expected to take about 80% of shipments⁴. In July the FCC banned newly authorized foreign-made humanoids, citing surveillance and remote-takeover risks⁵.

The two splits reinforce each other. A US humanoid will likely run on NVIDIA's GR00T stack⁶. A Chinese one will increasingly run on a domestic model trained on domestic chips. Materials add a second chokepoint that does not depend on chips. Chinese firms make about 90% of NdFeB magnets, and a humanoid needs dozens of high-torque magnet actuators⁷. We think it plausible, at about 65%, that Western humanoid unit costs stay at least 3x above Chinese costs through 2028. The current gap is 5 to 10x⁸, and the FCC ban shields US makers from that price competition without closing the gap.

Humanoids and open-weight models

The open-weight debate in language AI has shifted from security worry to a question of national competition: more than 270 organisations signed Nvidia's letter against early restrictions⁹. Robotics is taking in the same argument. GR00T N1.6 weights are public on Hugging Face⁶. Meanwhile small open models now run useful agent tasks on laptops and phones¹⁰, and the number of frontier-scale models is set to climb into the hundreds¹¹.

Combined with bodies priced below $7k⁸, this means anyone can put a capable, locally run control model on cheap hardware. The FCC ban covers the bodies, not the model weights. A robot assembled from permitted parts and running open weights can do much of what the ban was meant to prevent, and once released, those weights cannot be recalled.

Humanoids and EU AI regulation

Europe now enforces obligations on general-purpose models, with fines of up to 3% of global turnover¹². The Commission says its thinking on AI agents is "only preliminary"¹³. The machinery side is running later still: high-risk obligations for AI in machinery moved to August 2028¹⁴, and ISO 25785-1, the safety standard for legged robots, is still in draft¹⁵. A humanoid is an agent with a body. Its reasoning layer might count as a general-purpose model, its motor policy is part of the machine, and the gap between the two sets of rules is where incidents are most likely. Agility's cage-free Digit 5 is already in that gap¹⁶.

What the overlap opens up

The opportunity is real. Open robot models licensed across many bodies⁶ could let European makers such as NEURA compete on integration and trust rather than on cheap hardware¹⁴, turning a slow rulebook into a certification advantage by 2028. The risk is larger and only visible when the fields are read together. Robot control is concentrating on bloc-aligned chip and model platforms, even as open weights spread capability past any border. Home robots piloted remotely by staff¹⁷ carry camera and microphone access into a world of rival blocs. The first serious injury or data breach involving a humanoid will probably be judged under whichever bloc's rules apply, and the other bloc will not recognize the verdict. The open questions from part two about liability and who controls the robots' data become questions of foreign policy.