US Restrictions on Chinese Humanoid Robots: The Full Map

Key Takeaways

1，Washington is targeting finished robots, not parts — and drawing the line at government procurement. Section 163 of the FY2027 NDAA, passed by the House 216–212 on July 22, 2026, bars only the Pentagon from procuring, leasing or operating adversary-linked humanoid robotic systems; commercial sales remain legal. The GUARD Act — the FCC Covered List route to a full commercial ban — was excluded from the NDAA and remains in committee.

2，The only measure in force that touches components is the Pentagon 1260H list. Unitree was added on June 8, 2026: direct DoD contracting is barred from June 30, 2026, and from June 30, 2027 the bar extends to any product containing listed companies’ components. The bigger component-level weapons — the Section 232 robotics investigation and a potential ICTS rule — remain undecided.

3，America cannot afford to hit components, and that is the root of its self-restraint. Teardown-based estimates put a fully de-Sinicized Optimus Gen 2 bill of materials at roughly $131,000 versus $46,000 (about 3x). China controls ~90% of NdFeB magnet processing, and each humanoid needs 3.5–4 kg of high-grade magnets.

4，For Chinese makers the restrictions are near-symbolic today; the real variables are November 2026 and allied follow-on. Chinese firms shipped roughly 80–90% of the world’s humanoids in 2025, and their revenue and capital are overwhelmingly domestic — Unitree’s RMB 4.2 billion STAR Market IPO was approved on July 3, 2026, three weeks after its 1260H listing. But the US-China export-control truce expires in November 2026, colliding with Optimus V3’s production ramp and the US midterms.

5，Humanoid decoupling runs deeper than the drone precedent: it is dual-layer. Drones were a hardware problem; humanoids stack an AI service layer — models, data flywheels, compute — on top of the machine. FSD’s experience in China (data cannot leave, training compute cannot enter) previews Optimus’s bind, and Tesla’s Shanghai model is likely unreplicable for humanoids.

6，Decoupling will not be uniform — it follows a four-tier scenario gradient. Government and defense are already decoupled; the home and consumer-service market will split fastest, driven by the extreme sensitivity of in-home audio-visual and spatial data; commercial and public venues form a gray buffer; industrial deployment — offline-deliverable, ROI-driven, price-sensitive — decouples slowest and will be the last shared market.

7，The most likely endgame is Brain–Body Bifurcation by roughly 2035. China dominates the low-cost, high-volume “body” (hardware and manufacturing); the US holds the “brain” (AI compute and models); the intelligence layer decouples as model forking rather than service cutoff. Europe, Japan, Korea and the Middle East become the contested middle ground where each side replicates its own “Shanghai model.”

4. Why Components Are Untouchable: The Supply-Chain Math

• Bill of materials: teardown-based estimates cited by McKinsey put a fully de-Sinicized Optimus Gen 2 at roughly $131,000 versus $46,000 — about 3x; SemiAnalysis reckons an identical robot arm costs about 2.2x more to build in the US than in China.

• Actuators exceed 50% of BOM (Bank of America: linear ~27%, rotary ~24%), with dexterous hands ~19% — precisely where Chinese suppliers (Sanhua, Tuopu, Leaderdrive and others) are strongest.

• Magnets: each humanoid needs 3.5–4 kg of high-grade NdFeB (about twice an EV’s content), and China controls ~90% of magnet processing and ~93% of refining. MP Materials has a $400 million DoD equity deal and a price floor, but its 10,000-ton Texas “10X” plant only commissions from 2028 — realistic US magnet self-sufficiency lands in 2028–2030 at the earliest.

• China has demonstrated the reverse lever twice: April 2025 rare-earth magnet licensing stalled Optimus production (Musk on the earnings call: “China wants assurances these are not used for military purposes”), and Beijing’s October 2024 sanction of Skydio left America’s flagship drone maker rationing batteries within weeks.

• Countdown: the November 2025 truce suspended China’s expanded October 2025 rare-earth and battery controls (including the extraterritorial 0.1% content rule) for one year, expiring November 2026 — colliding with Optimus V3’s million-unit line plans, the overdue Section 232 decision and the US midterms. It is the single biggest uncertainty node of the next twelve months.

The DJI episode is the complete cautionary tale of “ban first, supply chain later”: only 23 of 300+ applicants passed the Pentagon’s Blue UAS vetting; the entire US drone industry builds fewer than 100,000 units a year; the ban did not conjure an American DJI — it delivered pricier, scarcer choices. IEEE Spectrum’s irony still lands: US robot makers win only if they ban Chinese competitors while keeping Chinese components — two things that rarely coexist politically.

5. Deeper Than Drones: Dual-Layer Decoupling — Humanoids Are Smart EVs, Not Drones

Drone decoupling was single-layer: a hardware supply-chain problem. Humanoids, like intelligent EVs, are dual-layer — beneath the finished machine sits an AI service layer of models, data flywheels and compute. That makes their decoupling both deeper and less reversible.

Tesla’s Shanghai model works only at the hardware layer. Model 3/Y — Chinese supply chain, Shanghai build, sold to Chinese and Asian customers — was a win-win because a car’s value lies mostly in manufacturing, and data localization proved solvable (Tesla built its Shanghai data center in 2021). But FSD previews Optimus’s bind: Chinese road data cannot leave the country under export-security rules, US-controlled training compute cannot enter under chip controls, and Tesla ended up running a locally trained, feature-limited FSD that has effectively forked from the US mainline. If cars are that constrained, humanoids — which record audio, video and spatial maps inside homes and factories — will be regulated harder, in both directions.

Hence a triple lock on Optimus:

• Keep the Chinese supply chain → carry US policy risk (the 232 ruling, the 1260H component clause, tariffs) plus China’s export-license leverage, already exercised once in April 2025.

• De-Sinicize → the BOM roughly triples toward $131,000, forfeiting price competitiveness across Asia — including China itself. The hardware simply cannot sell into the Chinese market.

• Even if the hardware sells into China → the AI stack must localize; no US compute, no US data flywheel. And mirror-image: Beijing will almost certainly bar American humanoids from sensitive government, factory and home settings on the very security logic Washington uses against Unitree.

A buffer variable: unlike autonomous driving, humanoids can do on-device, offline inference — model weights ship with the hardware, with no cloud dependence — which softens the “can’t use the other side’s AI infrastructure” constraint. But the moat in this industry is fleet learning: deployment data flowing back into training. Data and training compute are exactly the two nodes each side has locked. The intelligence layer therefore decouples in the end — as model forking, not service cutoff.

The Gradient: Decoupling Is Not Uniform Across Scenarios

“Brain–body bifurcation” will not happen everywhere at once. Ranked by data sensitivity and price-performance weighting, the speed and depth of decoupling follow a clear gradient.

Tier 1 — Government, defense and critical infrastructure: already decoupled. This tier never needed a commercial ban: NDAA Section 163, ASRA and 1260H are themselves the boundary. China mirrors it through procurement catalogs and its indigenous-substitution logic. The split here was a fait accompli in 2026.

Tier 2 — Home and consumer service: the fastest commercial split. A humanoid at home is a 24/7 sensor on legs: audio, video, spatial maps, family routines — far more sensitive than a car cabin. The “G1 phones home every five minutes” research pre-emptively destroyed trust in this tier. Legislatively it is the cheapest place to regulate — substitutes exist, vendor lobbying is weak, and the “Chinese robot in your living room” narrative is maximally mobilizing — so expect two parallel home markets from day one. The 4x price gap between 1X’s Neo ($20,000) and Unitree’s R1 (~$4,900) will be partly rationalized as a privacy premium.

Tier 3 — Commercial and public venues (hotels, retail, campus logistics, security patrol): the gray middle. Speed depends on proximity to public safety. Patrol and security robots will be regulated first — US police and prison use of Unitree gear has already been called out in Congress — while greeter and guide robots may stay open for years. This buffer zone is Chinese vendors’ main channel for keeping a presence in Western markets.

Tier 4 — Industrial: the hardest and slowest to decouple. Three structural reasons: production-line data is comparatively controllable and deliverable fully offline and on-premise; hard real-time requirements favor edge deployment, blunting the “cloud AI dependence” pretext; and factories buy ROI, not ideology — price-performance weighting is enormous, exactly where the Chinese supply chain is irreplaceable. Industry will be the last shared market and the longest-held overseas position for Chinese vendors. Two erosion vectors to watch: the expanding definition of “critical infrastructure” — energy, ports and semiconductor fabs get regulated first, with the eviction of ZPMC port cranes on national-security grounds a ready-made script — and the built-in conflict between offline-deployment promises and the fleet-learning business model: vendors promise data never leaves the factory, yet the flywheel needs it back. Regulators will eventually exploit that contradiction.

Bottom line: drone decoupling required one side to rebuild a hardware chain; humanoid decoupling requires both sides to rebuild a full stack of hardware + data + models + compute — advancing as a gradient from the government end through homes and commerce toward industry, not as one clean cut. The Shanghai model likely cannot be replicated for humanoids; instead, both sides will replicate their own Shanghai models in third markets — Europe, Southeast Asia, the Middle East. The race for time concentrates in Tier 4: Chinese vendors must lock in installed base and scenario data before industry gets regulated, while US policymakers face the constraint that every step toward the industrial tier costs domestic manufacturing more.

6. The China Side: Symbolic Impact, Three Lines of Response

Near-term, the restrictions barely register. Unitree sells almost nothing to the US government; three weeks after its 1260H listing, China’s securities regulator approved its STAR Market IPO (July 3 — an RMB 4.2 billion raise at a ~RMB 40 billion valuation, on 2025 revenue of RMB 1.69 billion and profit of RMB 278 million); six weeks after the listing, its H1 Pro launched in Europe, with North American channels slated for August. AgiBot secured A-share capital access by taking control of Shangwei New Materials and is reported to be preparing a 2027 Hong Kong IPO. The price curve keeps diving: the R1 from $5,900 to about $4,900, the G1 from $16,000 to about $13,500, with a BOM now under $9,000.

Chinese players’ response runs on three lines:

• Capital re-circuits domestically. Financing and listings have moved wholly to Shanghai and Hong Kong; the retreat of dollar capital is fully offset by RMB funds and local-government vehicles — city robotics and embodied-AI funds now exceed $26 billion combined.

• Market diversification. Europe first — Chancellor Merz toured Unitree’s Hangzhou plant in February, a pointed contrast with Washington — then the Middle East and Southeast Asia. The US market is maintained through low-exposure channels such as AliExpress and third-party distribution.

• De-Americanizing the supply chain. Roughly 20% of Unitree’s procurement is imported content — Nvidia Jetson, Intel RealSense, power devices. This is the genuine soft spot should Washington extend chip controls to robotics compute, and the layer Chinese firms are racing to substitute.

The real risks are mid-term: a revived GUARD-style commercial ban; allied follow-on — 1260H listings create compliance and reputational pressure for European and Japanese buyers and financiers, and the EU AI Act’s high-risk regime (phasing in from August 2026) hands Brussels a “restrict without banning” lever; and the hardening of the data-security narrative into a global market-access bar — every UniPwn-style incident deepens the “Chinese robot equals backdoor” label. Jamestown’s analysts already liken Unitree to “Huawei in an earlier technology cycle.” The analogy is itself the risk warning.

7. Mid-to-Long-Term Outlook: Three Scenarios Under Brain–Body Bifurcation

Baseline (most likely): the control ladder ratchets slowly; the market forks brain from body. The military procurement ban lands this year; a federal-wide ban follows within one to two years; commercial bans and component curbs are repeatedly deferred by supply-chain reality. By roughly 2035, two parallel ecosystems: China supplies most of the world’s (ex-US) machines and nearly all value-for-money hardware — the “body” — while the US and its allies hold models, simulation and high-end compute — the “brain” — and incubate domestic machines in security-premium niches. TrendForce and McKinsey both point to ~2035 for ecosystem bifurcation.

Accelerated decoupling: the November 2026 truce collapses, or the Section 232 decision hits components; China reactivates rare-earth and battery controls and extends them to robot components; Washington counters with ICTS rules and FCC listing. Lose-lose, but asymmetric: China loses a small export market; the US loses the only supply chain that supports its production timelines — Optimus’s and Figure’s cost curves and delivery schedules take the first hit — and is forced into CHIPS-scale robotics industrial policy that Congress has so far declined to fund.

Détente (least likely, not zero): a repeat of November 2025 — rare-earth licenses traded against tariff and control relief; commercial markets stay open; competition returns to product and cost. Chinese share keeps compounding, which Washington cannot politically tolerate for long. The scenario is inherently unstable.

Three watch points:

• The overdue Section 232 robotics ruling. Parts tariffs equal self-harm; finished-robot tariffs equal a precision strike on China. The choice will reveal Washington’s real strategy.

• The November 2026 expiry of the bilateral truce — do the rare-earth and tariff cards return to the table?

• The NDAA conference — whether Section 163 survives, and whether ASRA or GUARD language rides along, determines where 2027 starts.

8. What It Means for Industry Readers

For US builders (Tesla, Figure and peers): do not plan on a Shanghai model. The realistic hardware play is two-step — bridge on near-shored Chinese Tier 1 capacity (Sanhua in Mexico and Tuopu in Thailand are already moving) to hold the cost curve, while betting on vertical integration and non-Chinese sourcing (Japanese reducers, allied magnets) to hedge the 2027 1260H component clause and the 232 ruling. On the AI layer, accept model forking as a premise: treat non-China Asia — Japan, Korea, Southeast Asia, India, the Middle East — as the incremental source for the data flywheel, not China as base-case market.

For Chinese OEMs: plan the US market as “gradually closing” over three to five years; Europe, the Middle East and Southeast Asia are the real overseas theaters. Data compliance — on-premise deployment, verifiably offline operation, third-party security audits — is no longer a bonus but the ticket to Europe. Whoever first produces credible independent security certification locks in channels before allied restrictions arrive.

For Chinese component suppliers: a delicate window. US builders must keep buying Chinese actuators, reducers and magnets for the foreseeable future, with buyers absorbing the 301 tariffs — but the 2027 1260H component clause and the 232 ruling hang overhead. Near-shoring (Mexico, Thailand) plus customer diversification is table stakes; and watch the naming risk — a 1260H listing means US defense-adjacent supply chains will proactively purge you after 2027.

For investors: the near-term winners are not OEMs but de-risking beneficiaries — non-Chinese magnets (MP Materials and peers), Japanese reducers (Harmonic Drive, Nabtesco), US actuator startups, and the compute layer that sells to both sides. Nvidia is simultaneously the foundation of the US ecosystem and Unitree’s hardest-to-replace supplier — a straddle whose window may close if chip controls extend to robotics. At the OEM layer, valuation stories have already forked: Chinese vendors cash out in domestic capital markets (Unitree’s IPO approval sent more than 30 A-share robotics names limit-up), while US vendors price policy moats plus AI premium (Figure at $39 billion; Agility listing near $2.5 billion). The fork in pricing logic is itself a leading indicator of decoupling.

Method note: based on public information as of July 25, 2026. Primary sources include congressional bill texts (GovInfo), DoD 1260H announcements, FCC public notices, compliance analyses by WilmerHale, Covington and Jones Day, industry research from Omdia, TrendForce, Morgan Stanley, McKinsey, Bank of America and SemiAnalysis, and reporting by IEEE Spectrum, SCMP, Caixin and TIME. Where source figures diverge (e.g., Unitree’s 2025 shipments: 5,500+ company-reported vs. 4,200 per Omdia), both are noted.