Weapon or armour? The contentious debate over China’s open-weight AI

Weapon or armour? The contentious debate over China's open-weight AI

TL;DR

Unitree's sub-$20,000 humanoid robots and DeepSeek's MIT-licensed language models have staged the same argument in two domains simultaneously: China's AI sector is releasing systems capable enough for Western professionals to build real products on, and many are doing exactly that. The efficiency gains are documented and real. Whether the cumulative strategic effect of that dependency resembles a gift or a hook remains genuinely unresolved — and anyone telling you they have it figured out is working with assumptions they have not disclosed.

Key Takeaways

• Unitree Robotics announced the G1 humanoid robot at approximately $16,000, according to the company's product page, a price point that places commercial-grade humanoid hardware within reach of research labs and startups that would otherwise wait years for Western alternatives
• DeepSeek released DeepSeek-R1 under an MIT open-source license in January 2025, per the model's GitHub repository, allowing commercial use, fine-tuning, and redistribution with no royalties — the most permissive terms any frontier-class AI lab has offered
• Alibaba's Qwen series has accumulated millions of downloads on Hugging Face, establishing it as one of the most widely used open-weight alternatives to GPT-4-class APIs outside the United States
• DeepSeek reported training costs for DeepSeek-V3 at approximately $5.6 million in its technical paper — a figure widely repeated but challenged by independent researchers who noted it excluded amortised infrastructure, failed experiments, and alignment work
• US federal agencies have opened formal review processes on whether Chinese open-weight AI models should be covered by export-control frameworks similar to those governing chips and semiconductor equipment, with no settled policy outcome as of this writing, according to Reuters reporting
• Both DeepSeek and Qwen models are trained under Chinese cybersecurity regulations requiring that outputs do not contradict "socialist core values" — a constraint built into pre-training that persists in released weights and is not fully removable through fine-tuning alone

From Unitree to DeepSeek: why Hangzhou keeps going open

Start with geography, because it does real analytical work.

Unitree Robotics is based in Hangzhou. So is DeepSeek. So is the Alibaba AI research team that produces Qwen. This is not coincidental: Hangzhou runs on private capital and engineering culture that differs measurably from the state-adjacent Beijing AI scene where Baidu and defence-adjacent institutes operate. The city's relationship with the central government is closer to Shenzhen's commercial pragmatism than to Beijing's policy priorities.

This matters for the weapon-or-armour question because the motivations behind Unitree releasing the G1 humanoid at $16,000 — with an SDK that researchers can actually use — are more legibly commercial than political. Unitree wants to dominate global robotics hardware the way DJI dominated drones. Open ecosystems build moats when third-party developers start depending on your interfaces. That playbook is familiar. It is what GitHub did for Microsoft, what Android did for Google.

DeepSeek's MIT licensing follows the same commercial logic. If Western AI startups build their products on DeepSeek-R1, they depend on DeepSeek's continued model development, its architectural choices, and ultimately on a company operating under Chinese law. The licence is free; the dependency is not.

This is not evidence of a coordinated state strategy. It is evidence of companies making rational commercial decisions that carry strategic implications as a side effect. That distinction matters because the appropriate Western response to "commercial competition" differs substantially from the response to "state-directed technology transfer." Conflating the two produces bad decisions in both directions.

What the benchmarks actually show — and what they don't

DeepSeek-R1, released under MIT licence in January 2025, benchmarked competitively with OpenAI's o1 on mathematical reasoning and code generation. That result held across independent evaluations by researchers at several European universities and by technical reviewers at Hugging Face. The reasoning chain quality is not fabricated for benchmarks; you can inspect the outputs directly.

What is less settled is the training cost claim. DeepSeek reported approximately $5.6 million in compute for DeepSeek-V3 in its technical report. This number spread faster than any AI statistic since GPT-4's parameter count. Independent researchers noted that the figure counted H800 GPU hours during the final training run, not the full R&D cycle, pre-training experiments, alignment work, or the cost of the infrastructure itself. The honest read: DeepSeek almost certainly trained V3 for substantially less than the $100 million-plus that comparable Western frontier models cost. But $5.6 million is an undercount by an unknown amount, and citing it at face value is credulous.

For Unitree's hardware, the verified claims are more straightforward. The G1 has been benchmarked in locomotion and dexterity tasks by research labs who purchased units — the motion control performance is real and reproducible. The unverified claims are longer-horizon: roadmap materials have suggested expanded manipulation capabilities, but those are forward-looking projections and should be treated accordingly.

What Qwen and DeepSeek share, and what neither the boosters nor the skeptics discuss adequately: both model families are trained under Chinese content regulations that require filtering outputs contradicting official state positions on Taiwan, Xinjiang, Tiananmen, and political history. The weights are open. The training values embedded in those weights are not cleanly separable by fine-tuning alone. Multiple researchers have documented that content restrictions on topics sensitive to Chinese regulators persist after aggressive RLHF overrides. This is not a backdoor in the technical sense. It is a constraint. The difference is real — and the constraint is also real.

What this changes for Western founders and consultants

The useful frame is not "should I trust a Chinese AI company." That question generates more heat than decision-making clarity. The more productive question is: what specifically am I building, and what does the dependency structure look like in eighteen months?

If you are building an internal tool — a coding assistant, a document summarizer, a RAG system over proprietary data — DeepSeek-R1 or Qwen 2.5 running on your own infrastructure is a defensible choice. The model weights are local. A Chinese company cannot push a model update to weights you already downloaded and self-host. You absorbed the risk at the moment of download, and you can assess that risk concretely.

If you are building a consumer product with inference running through a Chinese API service, you are exposed to different risks: service availability, policy changes in your jurisdiction, potential future export controls, and reputational exposure in markets that are becoming more sensitive to AI provenance. None of these are hypothetical edge cases.

For robotics specifically, Unitree hardware is already in research labs across the US and Europe, and the debate over restricting sales has not produced restrictions. If you are a team evaluating humanoid hardware for physical AI development, the context around China's broader AI hardware stack — including how Huawei and domestic chip manufacturers are closing the gap with Nvidia, as tracked in this breakdown of China's AI chip market share — informs whether the cost advantage in hardware is a durable feature or a transitional one. A $16,000 humanoid versus a $200,000 humanoid changes the set of labs that can run embodied AI experiments by an order of magnitude. That is not a subtle competitive shift.

China's open-weight AI landscape: what you're choosing between

The geography column is doing real work here. Hangzhou's two frontier open-weight labs (DeepSeek, Alibaba Qwen) and its dominant robotics hardware maker (Unitree) have each chosen openness. Beijing's labs have largely not. If you are deciding whether to build on Chinese AI, you are mostly deciding whether to build on Hangzhou's output — and that is a more specific, more evaluable question than "Chinese AI" as a monolithic category.

When NOT to build on Chinese open-weight AI

Don't use it if your outputs regularly touch politically sensitive content categories. Open-weight Chinese models carry training-level restrictions that may produce inconsistent or filtered outputs in use cases you did not anticipate. A risk management tool for a regulated institution that processes documents referencing Taiwan, Hong Kong, or certain historical events could fail in ways that are not easy to debug. Test specifically for your edge cases before committing.

Don't assume self-hosting eliminates all regulatory risk. Export controls on AI models are an active policy area in the US and EU. A model that is legal to download today could face restrictions if your jurisdiction extends chip-style export controls to software. If your architecture is difficult to swap mid-product, a future restriction creates a rebuild cost, not just a compliance headache. Keep the coupling loose by design.

Don't treat open weights as equivalent to open data provenance. You can inspect what DeepSeek-R1 does. You cannot inspect what data it was trained on, or what the full reinforcement signal was. "Open-weight" is a licensing term, not a transparency guarantee. In healthcare, legal, or regulated finance applications where data lineage matters for compliance, that gap needs a documented response before you sign off on the architecture.

Don't use API-dependent Chinese AI services for production applications without a fallback. Doubao, Kimi, and Ernie run API-only. A service disruption, policy change, or regulatory action in your market takes your product down with it. The open-weight options — DeepSeek, Qwen, InternLM — do not carry this risk if you are self-hosting. The licensing and the deployment model are separate decisions; make both deliberately.

Where this is heading

Export control friction will increase before it decreases. The US has already moved to restrict Chinese AI chip imports through successive rounds of controls. Extending similar logic to model weights — particularly in dual-use applications like robotics perception, autonomous systems, and computer vision — is a policy conversation that has been formally started. The timeline is unclear. The direction is not.

Physical AI closes the loop in a way software alone does not. Unitree's humanoid, Fourier Intelligence's humanoid platforms, and the broader Chinese robotics hardware wave matter to this debate because open-weight software and accessible-hardware robotics create a compounded capability transfer. A Western startup using DeepSeek for AI reasoning and Unitree hardware for physical manipulation is building on an integrated Chinese-developed stack. That is a different risk profile — and a different strategic dependency — than using either in isolation.

Hangzhou's open strategy is deliberate market capture, not idealism. Ecosystem lock-in through developer dependency is a proven commercial playbook. How durable that lock-in proves depends on whether Western alternatives — Llama, Mistral, and publicly funded European initiatives — can stay competitive on capability per dollar. At the moment, they are not clearly ahead.

Regulatory divergence will create real architecture splits. European AI regulation, US export controls, and Chinese data law are moving in different directions simultaneously. A global product built today on a single model family may face a point where the same weights are legally complicated in one jurisdiction and not another. Architectural flexibility — the ability to swap models at the inference layer — is risk management, not over-engineering.

FAQ

Is Unitree's G1 humanoid actually usable for commercial AI development right now? For research and prototyping, yes — multiple labs have published results using Unitree hardware for locomotion, navigation, and simple manipulation. For production deployment in environments with liability exposure — healthcare, public spaces, industrial settings — the honest answer is not yet, and this applies to all humanoid platforms, not just Unitree. The constraint is software maturity and safety certification, not price.

Can I fine-tune out the political content filters in DeepSeek or Qwen? Partially. Researchers have found that RLHF fine-tuning can reduce, but not reliably eliminate, content restrictions baked in during pre-training. If your application involves content that Chinese regulators would flag — and the list of flagged topics is broader than most Western users expect — plan for residual filtering behaviour and test specifically for it before deployment.

Does using DeepSeek mean my data is exposed to Chinese government access? If you are running open weights locally or on your own infrastructure, no. The weights are static files with no telemetry channel. If you are using the DeepSeek cloud API, the same data exposure questions apply as with any third-party API service, plus the additional consideration that DeepSeek operates under Chinese data law, which grants authorities broader access than most Western regulatory frameworks permit. Self-hosting resolves the data transmission question; it does not resolve the dependency or content-filtering questions.

What is the actual security risk from open-weight Chinese AI, compared to the policy rhetoric? Current evidence does not support the backdoor framing that circulates in policy circles. Open weights are inspectable, and researchers have not found evidence of covert exfiltration capabilities in DeepSeek or Qwen models. The documented risks are different: content filtering constraints built into training, uncertain data provenance, and commercial dependency on a company operating under a legal regime that may not align with your interests long-term. These are real risks. They are not the same as the espionage framing, and treating them as equivalent produces bad decisions in both directions.

Is there a meaningful difference between DeepSeek (Hangzhou) and Baidu Ernie (Beijing) in terms of strategic risk? Yes, and the distinction is consequential. DeepSeek is a private company with no documented state ownership, operating in Hangzhou's private-capital ecosystem; its open licensing is a commercial strategy. Baidu operates with significant state-adjacent relationships and government contracts under different regulatory pressures. Both companies operate under Chinese law. Their corporate structures and motivations are meaningfully different. "Chinese AI company" is not a homogeneous risk category, and treating it as one is an analytical shortcut that leads to both over-restriction and under-caution in the wrong places.

What is the practical 'so what' for a Western founder in 2026? Build on open-weight Chinese models for internal tooling and research where self-hosting is feasible and your content categories do not intersect with Chinese regulatory sensitivities. Avoid API-only Chinese services for production-critical applications without a documented fallback. On the hardware side, Unitree is a real option for physical AI research at a price point that changes what is accessible — evaluate it on the merits, with clear eyes on the supply chain and export control trajectory. Keep your model layer swappable. The biggest strategic error is either refusing to engage with the capability on ideological grounds, or embedding so deep that a future policy change becomes a crisis.