Nobody ships a shared robot brain
A deep dive on twenty US and twenty-eight Chinese robotics players, the state of multi-robot memory, on-robot compute and the parts supply chain, and what it says about where a Singapore full-stack team can do better.
We ran four research passes on 5 September 2026: the US players, the Chinese players, the state of the art in multi-robot memory and on-robot compute, and the humanoid parts supply chain. About 230 searches and 190 full-page reads. This is what they say, with the single-source claims and our own inferences marked as such in the underlying reports.
The gap
No company in the United States or China ships an on-site, shared short-term and long-term memory for a mixed robot team. What exists sorts into four families. Shared weights, offline: Figure uploads at the dock and retrains in the cloud in hours to days, and AgiBot's May 2026 paper broadcasts a new policy to sixteen robots every fifty training steps. Shared intent and traffic: Open-RMF, RoMi-H and Amazon's DeepFleet, the only one with fleet-scale economic proof. Shared operations data: Boston Dynamics Orbit, Agility Arc, InOrbit, Viam. Shared language: Google's Gemini Robotics ER 2, released 30 July 2026, where each robot runs its own copy of the model and hands off sub-tasks in natural language over the cloud, in a demo, with healthcare excluded by the model card.
UBTech's BrainNet 2.0 claims a group brain for its Walker S2 fleets. The public evidence is one press example, a blocked aisle that other robots route around, with no architecture, no latency figure and no third-party deployment. BAAI's RoboOS-NeXT is the real technical reference, a shared spatial, temporal and embodiment memory, and it is a lab framework with two or three robots and a cloud brain whose authors list memory update and retrieval as open problems.
The consistency problems of a memory many robots write to were only formalised in June 2026: leakage across scopes, stale propagation, contradiction persistence and provenance collapse. No robotics product names them.
The chip ceiling
Jetson Thor runs a 3-billion-parameter vision-language-action model at 10 to 25 Hz and a 7 to 8-billion-parameter planner at 40 to 80 tokens per second. Its memory bus, 273 GB per second, is the same as a DGX Spark and a quarter of a desktop RTX 4090, and it is the binding limit: a 16.7-billion-parameter policy drops to 2 Hz, and holding 1,000 timesteps of context in the policy drops it to 1.3 Hz. Memory therefore has to live outside the policy and be retrieved. A 30 to 70-billion-parameter planner, an LLM verifier, long-term memory consolidation and fleet arbitration belong on a site server, where one RTX Pro 6000 class box serves five to twenty robots.
Radio decides the rest. Private 5G and Wi-Fi 6E measure 10 to 30 ms typical round trip with tails of 100 to 150 ms and multi-second dropouts in lifts. Running the manipulation policy off the robot is faster in the lab and unsafe on those tails, which is why no humanoid vendor does it.
Thor supply is the other constraint. NVIDIA raised Jetson prices by 50 to 100 percent in July 2026 as LPDDR5X lead times reached 26 to 58 weeks. The classification is 5A992.c, licence-free to Singapore; the pending US Chip Security Act would touch a site server GPU but not Jetson. Huawei Ascend carries the Bureau of Industry and Security's "anywhere in the world" language and is out.
The United States
Figure has one audited unit at BMW with 1,250 hours, and BMW chose Hexagon for its German plant. Agility has nine facilities and 65,000 hours, and GXO's chief executive said in August that none of its five humanoid vendors will be in production in 2026. Boston Dynamics' 2026 Atlas units all go to Hyundai and Google. Physical Intelligence's memory work is two seconds of video and fifteen minutes of text within one episode on one robot. The only US hospital incumbent, Diligent's Moxi, sold for 29 million dollars in January after MultiCare removed fourteen robots that needed an escort between floors. No US humanoid or VLA vendor targets ports. Gartner expects fewer than twenty companies in production by 2028. No vendor publishes uptime or mean time between failures.
China
China built about 12,800 humanoids in 2025. IDC finds more than 85 percent of deployments are performances, education, data collection and guided tours, and Unitree's prospectus shows more than 70 percent of humanoid revenue from research and education. UBTech's founder concedes its robots are half as efficient as people; at its Liuzhou training centre, trainers get one usable teleoperated motion per fifty to three hundred tries. The verified productive cases are narrow and mostly wheeled: Galbot's licensed pharmacies at 95 to 99.5 percent pick success, Siasun's two hundred vehicles at PSA under PSA's own fleet brain, Xiaomi's nut fastening at 98 percent.
The trust gap is documented rather than argued. Unitree's Bluetooth root exploit with one hardcoded key across the fleet went unpatched from September 2025 to at least August 2026, its robots report to vendor servers every five minutes, and the Federal Communications Commission put all foreign-produced mobile robots on its Covered List on 28 July 2026. China's May 2026 lifecycle registry requires joint wear, battery and training-history telemetry to reach a national platform for domestic market access, so firmware roadmaps point toward more phoning home. Japan's compliance stance blocks bipeds, so ZEALS built a wheeled robot on a Unitree chassis for a hospital trial. Saudi Arabia's flagship humanoid deal went to a British company with local assembly.
At the same time China is giving the model layer away. RynnBrain, Qwen-Robot, LingBot-VLA 2.0 across seventeen brands, WALL-OSS and UnifoLM are open, and more than half of the first-half 2026 embodied-AI funding went to brain companies rather than body makers.
The body
The rotary joint is a commodity at 110 to 480 US dollars a module, with harmonic reducers halved in price since 2024. Planetary roller screws are the one scarce actuator part. Hands are where the money and the failures are: 6-degree-of-freedom hands cost a few thousand yuan, the top end costs fifty thousand dollars, and the documented failure modes are tendon wear, coreless motor burnout and tactile skins that drift. UBTech rates the Walker S2 hand at 80,000 cycles, about two weeks at four hundred picks an hour. No vendor publishes ingress ratings, operating temperatures or a preventive-maintenance schedule for a humanoid. China's extraterritorial rare-earth rules are suspended only until 10 November 2026.
What it means for Tidewell
Tidewell Brain has two modes. Solo runs everything a robot needs on the robot and is a first-class state, not a fallback. Crew connects every robot on the site network to one governed memory with three tiers, core, short-term and long-term, and promotes what robots learn into the site's core memory only through a named human. The contract between the two modes, link-loss detection under two seconds, automatic downgrade, append-only re-sync, no policy and no safety function off the robot, is published as a specification and will be measured.
On the body we buy joints, screws, sensors, cells and chassis, and build the rated-life sealed hand, the calibrated fingertip, the secure body bus, the certified pack and the safety file. The wheeled W1 is the volume product. The biped B1 goes only where PSA has said legs are needed. A Brain Kit brings the Unitree and AgiBot bodies already in Singapore's testbeds under the same brain, with its safety limit stated on the datasheet.
What we will not claim: a manipulation policy served over wireless, weights that update across the fleet in real time, a 70-billion-parameter model on the robot, a whole shift in the policy's context, a cross-site global brain. The reports that back every line here are in the company repository and will be published in full as the Insights in this category.