On September 17, 2026, two events happened roughly eleven time zones apart that will do more to shape the humanoid industry than any demo video this year. In Hong Kong, D-Robotics announced the close of a $400 million Series C, the largest financing ever recorded for a dedicated robotics semiconductor company, built to power Chinese embodied AI. That same week, a Tesla robotics team landed in Ningbo to audit component suppliers for Optimus mass production, and new drone footage showed the steel skeleton of Tesla’s dedicated robot factory at Giga Texas five column grids from closure.
Three data points, one direction of travel. The humanoid industry is decoupling into two largely self-sufficient spheres: a Chinese ecosystem scaling around domestically designed silicon, and an American ecosystem responding to an import ban by pulling brains, bodies, and fabrication onto its own soil. The split is no longer a forecast. It is visible in funding closes, factory pours, and federal register entries.
What the FCC Rule Actually Does
The proximate cause is regulatory. On July 28, 2026, the Federal Communications Commission added foreign-made humanoid robots, quadruped robots, and connected power inverters to its Covered List under the Secure and Trusted Communications Networks Act. The practical mechanics matter more than the headline. Robots already certified for sale in the US can continue to be imported by retailers, as CNBC reported at the time. What is blocked is anything new: a new Unitree humanoid, a materially revised model, or any machine requiring fresh FCC equipment approval now faces a denial unless the Department of War grants a Conditional Approval. The rule is formally country-agnostic but was explicitly aimed at China over national security concerns, and it technically reaches qualifying European-built robots too.
The effects surfaced quickly. The largest US distributor of Unitree machines pivoted its business within weeks. A parallel Section 232 probe into robotics imports, noted in a recent Eurasia Review analysis, keeps the door open to tariffs on a much wider set of components. And the ban exposed an uncomfortable dependency: as KrAsia reported, Western robot builders themselves lean heavily on Chinese actuators, harmonic drives, magnets, and battery cells. Washington banned finished machines it could not build cheaply, while the parts pipeline remains deeply intertwined.
This is the backdrop against which the D-Robotics round should be read. The company raised $400 million to expand the exact category of product that the Covered List makes radioactive in the United States. The investors, led by South Korea’s Mirae Asset with strategic participation from a leading global internet company and existing shareholders, are effectively underwriting the non-American robot stack.
D-Robotics and the Case for Robot-Native Silicon
D-Robotics was spun out of Horizon Robotics, the Hong Kong-listed automotive chip designer, in early 2024, with a mandate to adapt edge-inference silicon for machines that think with their bodies. The company’s Sunrise family spans 5 to 560 TOPS of INT8 compute, from toy-grade companion robots to full humanoids. According to the funding announcement, cumulative Sunrise shipments have passed 8 million units, revenue grew several times year over year in the first half of 2026, and the embodied AI business has entered mass production. Gasgoo calculates the company has now raised nearly 4.5 billion yuan within 2026 alone.
The flagship S600, launched November 2025, embodies a design thesis worth understanding because it diverges from the Western default. D-Robotics splits the robot brain in two. The “Brain” is an 18-core ARM Cortex-A78AE cluster that runs language, vision-language, and vision-language-action models: the slow, deliberative layer. The “Cerebellum” is a proprietary Nash-architecture neural accelerator tuned for the fast layer: 500 Hz joint-torque loops, locomotion stabilization, sensor fusion. Separating deliberation from motor control on dedicated silicon avoids the pipeline contention that occurs when one processor must juggle a transformer inference call and a balance-critical command in the same queue. Within six months of launch, more than 20 embodied AI customers adopted the S600, including UBTECH, Fourier, Booster Robotics, PaXini, Astribot, and X Square Robot.
The Western benchmark in this segment is NVIDIA’s Jetson AGX Thor: 2,070 FP4 TFLOPS on Blackwell, 128 GB of LPDDR5x in a 130-watt envelope, already adopted by Agility Robotics and Boston Dynamics. The two philosophies are not directly comparable on spec sheets. Thor’s floating-point throughput favors large, lightly quantized models; the S600’s INT8-first design prioritizes cheap, efficient deployment of quantized networks at much lower cost. But the strategic asymmetry is stark. NVIDIA sells a compute platform to everyone; D-Robotics, backed by Meituan-adjacent capital and 500 university programs producing 100,000 developers on its toolchain, is building a vertically captive ecosystem that no US rule can revoke.
Tesla’s Answer: Vertical Integration at Continental Scale
The American response to this bifurcation is being built in Austin, Texas, in reinforced concrete. Tesla’s dedicated Optimus factory, under construction on Giga Texas’s North Campus since late March, is approaching structural completion six months after groundbreaking, with steel five column grids from the north perimeter and concrete already pouring on three upper floors in parallel. The building is part of a 5.2 million square foot expansion that will eventually run more than 4,000 feet, nearly the length of the adjacent car factory, toward a long-term target of 10 million Optimus units per year.
Two details distinguish Tesla’s approach from every competitor. First, the Fremont pilot line began mass-producing the current Gen 3 robot in January and is expected to build tens of thousands of units this year, with a stated capacity around one million units: not for sales, but for data generation, feeding real-world teleoperation and autonomy telemetry back into the model stack before Texas ramps. Second, the same campus hosts Terafab, the joint Tesla and SpaceX chip fabrication plant intended to supply Optimus silicon directly. Co-locating robot assembly and chip production is a deliberate hedge against exactly the kind of supply chain weaponization the FCC rule represents.
The commercial signals are catching up to the construction. Tesla has reportedly placed its inaugural large-scale production order of approximately 5,000 Optimus units with component suppliers, and the Ningbo audits this month focused on verifying supplier exclusivity and consistency for mass production. There is a deep irony here that defines the whole moment: Tesla is qualifying Chinese suppliers for a robot that US policy is simultaneously walling off from Chinese competition, because in actuators, reducers, and battery cells, the dependency still runs one way. Tesla’s answer is not to escape that dependency overnight but to compress it into a transitional window while Terafab and domestic suppliers ramp.
China’s Parallel Universe Is Already at Volume
While Washington legislates, China ships. Roughly 40,000 humanoids were shipped in China in the first half of 2026, though, as the Eurasia Review analysis cautions, most went to demos, exhibitions, and research deployments rather than productive work. Unitree’s hardware costs are approaching $6,000 per unit even while the company remains loss-making, a price point no Western manufacturer can touch. XPeng’s Iron line runs an 80 percent automated assembly process with mass production targeted for end of 2026. The 15th Five-Year Plan embeds embodied AI data standards and 70-plus data collection sites as national infrastructure.
Add D-Robotics’ 8 million shipped chips and the picture completes itself: China is assembling a full-stack, domestic, low-cost humanoid supply chain, from silicon to actuators to data pipelines, at volumes the US cannot currently match at any price. Goldman Sachs projects roughly 6.5 million humanoids shipped by 2035; on current trajectory, the majority of that hardware will be built in one sphere and sold mostly in the other’s absence.
The Bill for Decoupling
Decoupling is not free for either side, and the costs land differently. For American builders, the bill is near-term and denominated in unit economics: domestically sourced actuators, rare-earth magnets outside Chinese processing, and NVIDIA Thor-class compute push bill-of-materials far above the $6,000 frontier Unitree is testing, before any margin. The FCC’s own logic concedes this: the ban raises prices and shrinks the menu of machines available to US buyers, as the Eurasia Review op-ed bluntly argues. For Chinese builders, the bill is deferred and denominated in markets: losing US and likely European access caps the addressable customer base at precisely the moment production capacity is scaling, forcing the domestic ecosystem to absorb millions of units through state procurement, factory pilots, and eventually consumer channels at loss-leading prices.
The swing variable is the rest of the world. Southeast Asia, the Gulf, Latin America, and Africa will choose robots on price and availability, not on Washington’s Covered List. If D-Robotics-backed machines at Chinese price points win those markets while Western humanoids remain five to ten times costlier, the two-world structure hardens into a durable global settlement rather than a transitional friction.
What to Watch Through 2027
Three signposts will tell us how deep the split goes. First, whether Tesla’s Fremont line hits its tens-of-thousands target this year: that number is the US sphere’s first real production datum, and it doubles as the data engine for Optimus autonomy. Second, whether the first Conditional Approvals emerge from the Department of War for any new Chinese model, which would reveal whether the ban is a wall or a negotiating instrument. Third, whether D-Robotics’ S600 ecosystem produces a humanoid whose landed cost in a neutral third market undercuts the cheapest Western equivalent by an order of magnitude, the point at which the price gap becomes strategic rather than commercial.
The humanoid industry spent 2024 and 2025 pretending geopolitics was someone else’s problem. September 2026 ends that pretense. The robots are still learning to fold laundry, but the map on which they learn it has just been redrawn.