# China Backs Open AI as DHS Bans DeepSeek Oct. 20 [2026] > Source: > Published: 2026-10-26 09:54:12+00:00 Chinese AI labs keep handing their models to the world for free, and Washington keeps moving to keep those models out of government systems. That split sharpened on October 20, 2026, when Reuters reported that the U.S. Department of Homeland Security ordered civilian federal agencies to remove DeepSeek products and any third-party services built on the model. The question behind the headline is simple: why does China favor open AI models like DeepSeek while others stay cautious? The answer starts with what DeepSeek itself has said. Its founder, Liang Wenfeng, told [Reuters](https://www.reuters.com/technology/artificial-intelligence/deepseek-share-some-ai-model-code-doubling-down-open-source-2025-02-21/) that the company decided to make all of its models [open-source](https://shattered.io/best-open-source-llm-october-2026-kimi-glm-deepseek/), unlike OpenAI. DHS, for its part, said the models posed “multiple threat vectors” and warned that use by U.S. government employees or vendors could send sensitive data to China. Our earlier report on the [DHS ban on DeepSeek under a 2018 law](https://shattered.io/dhs-bans-deepseek-federal-systems-2018-law-2026/) covers the legal mechanics. This piece focuses on the strategy behind the open-weight approach and the reasons others hesitate. One caution before the analysis. The search record behind this story does not contain a single official Chinese government statement that explains the preference for open models in terms of cost, innovation, geopolitical influence or software diffusion. Those motives are the reasoning of analysts and company founders, and this article labels them that way. It also does not show that “others” are universally wary. What it documents are specific concerns from some governments and organizations. Don't miss new tech stories on Google Add Shattered once in the Google app and our stories show up in your news suggestions. [Add now](https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F) ## What Happened on October 20: The DHS Order in Plain Terms Reuters reported that DHS ordered the removal of DeepSeek products, and of third-party services using the model, from civilian federal-agency IT systems on October 20, 2026. Some outlets and earlier coverage on this site date the filing a day earlier, so treat the exact calendar day with a little care. The substance is the same in every account. According to law-firm alerts and trade-press summaries of the notice, the exclusion took effect immediately. Agencies can request exceptions, but those requests face review and need written approval. The order covers federal civilian agencies and related systems. It does not amount to a general ban on consumer or private-sector use of DeepSeek, a distinction that gets lost in many social-media summaries. The Congressional Research Service describes DeepSeek as a developer of open-weight large language models, including DeepSeek and DeepSeek-R1. It characterizes the federal response as a requirement that executive-branch departments and agencies bar products made by the China-headquartered company and remove anything already deployed. That framing matters because it puts the open-weight nature of the models at the center of the policy story, not at its edge. ## Why China Favors Open-Weight AI: The Short Answer Chinese developers and the analysts who follow them tend to cite four motives for releasing weights. First, open release accelerates adoption inside China, because universities, startups and enterprises can run a model locally instead of renting access from a foreign cloud or a closed American API. Second, it lowers the cost of using AI, since users can self-host and tune the model for their own hardware. Third, it builds an ecosystem of derivative models, tools and services around Chinese-origin technology. Fourth, it extends influence abroad, because developers who build on a model tend to follow its roadmap. None of these motives comes from a published Beijing policy memo in our record. They are the working explanation of researchers and the stated positions of at least one lab founder. The fresh research behind this story also found no quantified measure of how much domestic diffusion open weights have produced. Anyone who cites a precise adoption figure for China open-weight AI should name the survey it came from. The practical effect is easy to see even without a statistic. A lab that publishes weights lets thousands of outside engineers find bugs, build tooling and port the model to new chips. That free labor compounds. A lab that keeps weights private has to pay for every improvement itself, and it has to persuade customers to trust an API they cannot inspect. ## Liang Wenfeng’s Case for Open Source, in His Own Words The clearest primary-source explanation comes from DeepSeek’s founder. DeepSeek is based in Hangzhou, and Reuters reports that its parent company is a quantitative hedge fund controlled by Liang. That ownership structure gives the lab unusual freedom to pick a strategy without pressure from outside investors. Liang has used that freedom to argue for openness, repeatedly and bluntly. In an interview cited by the [Cato Institute](https://www.cato.org/blog/chinas-march-imitation-innovation-case-deepseek), Liang said: “We won’t go closed-source. We believe that establishing a robust technology ecosystem matters more.” The sentence frames openness as an ecosystem bet, not a charity project. It also reads as a direct answer to the question of why a hedge-fund-backed lab would give away its main asset. In a separate interview published by [ChinaTalk](https://www.chinatalk.media/p/deepseek-ceo-interview-with-chinas), Liang said: “Open source, publishing papers, in fact, do not cost us anything.” He added: “We will not change to closed source.” The first line is the more revealing one. His argument is that sharing research does not drain the lab’s competitive position, because the real moat sits in the team’s ability to keep producing new work. A third quote, which Reuters carried in its February 2025 report on DeepSeek’s code release, adds a cultural motive: “Having others follow your innovation gives a great sense of accomplishment.” Reuters noted at the time that this approach contrasted with the more closed strategies common among major U.S. AI companies. Taken together, the three statements describe a lab that treats influence and ecosystem growth as the prize. ## DeepSeek’s Open Record: From R1 to Five Repositories DeepSeek’s openness did not start with a slogan. It started with a model. The company publicly released its open-source R1 reasoning model in January 2025, and Reuters reported that the system rivaled Western models while costing less to develop. That combination made the release a global story and gave every other Chinese lab a template to copy. On February 21, 2025, DeepSeek said it would make five code repositories public the following week. It called the planned release “small but sincere progress” and promised to share the code “with full transparency.” The wording matters less than the timing. The company chose to follow a high-profile model launch with more disclosure, not less. | Date | Event | Source named | |---|---|---| | January 2025 | DeepSeek releases open-source R1 reasoning model | Reuters | | February 21, 2025 | DeepSeek announces five code repositories for the following week | Reuters | | Reuters interview (date not stated in our record) | Liang Wenfeng says all DeepSeek models will be open-source, unlike OpenAI | Reuters | | October 20, 2026 | DHS orders DeepSeek products removed from civilian federal systems | Reuters | | October 20, 2026 | Order covers third-party services built on DeepSeek, effective immediately | Law-firm alerts, trade press | | October 20, 2026 | Agencies may seek exceptions with written approval | Law-firm alerts, trade press | The table shows how little distance separates the two ends of the story. A strategy that won DeepSeek global attention in early 2025 became, by late 2026, the stated reason a major government cut the company out of its networks. Openness created the reach, and the reach created the scrutiny. ## The Cost and Diffusion Logic Behind Open Weights Cost is the argument most analysts reach for first. Open weights let a company run a model on its own servers, quantize it for cheaper hardware and tune it for a narrow job. DHS materials recognize that DeepSeek models can be deployed in self-hosted and fine-tuned forms, which is precisely the property that makes them attractive to engineers and worrying to security officers. Price pressure is already visible in how developers talk about these models. Our comparison of [DeepSeek and Claude API pricing](https://shattered.io/deepseek-vs-claude-flash-1017-vs-claude-5-5-api-pricing-2026/) shows how far apart the input-token rates sit between an open-weight lab and a closed American one. Open weights also let a third-party host undercut the original lab, which pushes prices down across the whole market. There is a hardware angle too. Efficient models can spread across modest infrastructure and reduce dependence on scarce high-end computing. That point comes from analysts, and the fresh research found no current compute figure to back it with a number. Still, the logic is intuitive. When the best chips are hard to buy, a model that runs well on weaker ones carries more strategic value. ## Ecosystem, Standards and Soft Power Beyond cost, open weights buy something a closed API cannot: a community. When developers build tools, adapters and fine-tunes on a model, they also build habits, file formats and documentation that tie the next project to the same family. Chinese firms can use that effect to establish technical conventions and attract developers who might otherwise default to American stacks. The effect extends to government-to-government competition. Our analysis of the [US-China AI blocs, WAICO and Pax Silica](https://shattered.io/us-china-ai-blocs-waico-pax-silica-29-founders-35-warned-2026/) describes how Washington and Beijing now court the same middle powers. A freely downloadable model works as a quiet recruiter in that contest. A ministry that can run a Chinese model on its own hardware has less need to ask anyone’s permission. Be careful with the strong version of this claim. The search record does not show that Beijing designed open weights as a diplomatic tool. It shows that open release has side effects that favor Chinese visibility. Strategy and consequence are different things, and the evidence supports the second more firmly than the first. ## Why the U.S. Government Stays Cautious The caution on the American side rests on governance, not on hostility to open code. DHS said the DeepSeek models posed “multiple threat vectors” and raised the risk that government employees or vendors could pass sensitive data to China. The order’s own logic points at jurisdiction. The concern, as summarized by law-firm alerts, is that data tied to a company operating in China could become reachable by Chinese authorities under Chinese law. ### Data exposure and visibility Once weights are downloadable, an agency has less visibility into fine-tuning, deployment, data flows and derivative systems. A closed model run by a U.S. vendor at least has a contract, an audit trail and a legal entity to sue. A fine-tuned copy of an open model, passed through three subcontractors, can have none of those. That is the core open-weight worry in the federal view. ### The third-party reach of the order The order reaches beyond the DeepSeek app. It also covers products and services built using DeepSeek implementations. That clause closes a workaround, because a vendor could otherwise route inference through a DeepSeek model behind a different brand name. We covered the signing timeline in our report on the [DeepSeek ban and the 60-day window](https://shattered.io/dhs-deepseek-ban-mullin-signed-60-day-window-2026/), and the third-party clause is where procurement teams will feel the most work. ## Not a Simple East-West Split It would be wrong to say China supports openness while the United States opposes it. U.S. companies and researchers also develop and use [open-weight models](https://shattered.io/import-ai-476-jack-clark-china-open-weight-models-2026/). The DeepSeek case shows something narrower. A government can restrict the use of a particular model inside its own systems when provenance, jurisdiction, data access and geopolitical risk outweigh the benefits of openness. The same applies in reverse. Chinese regulators and companies have their own limits on data, content and security review. A Chinese lab publishing weights does not mean Chinese institutions have dropped all caution about AI. The record on that point is thin, and this article does not claim to settle it. The South China Morning Post framed the security side of the question as weapon or armour, and we examined that framing in [our look at China’s open AI and the cybersecurity debate](https://shattered.io/china-open-weight-ai-weapon-or-armour-cybersecurity-debate-2026/). One side sees cheap, downloadable capability that lowers the barrier to misuse. The other sees inspectable tools that defenders can run without sending data to a third party. Both camps point at the same models. ## Open Weights vs Closed APIs: A Competitive Comparison The table below contrasts the two approaches as design choices. It is our analysis of the trade-offs described in the sources, and it contains no benchmark or price figures. | Dimension | Open-weight approach (DeepSeek model) | Closed-API approach (OpenAI model) | |---|---|---| | Access | Download weights, run anywhere | Call a hosted endpoint | | Customization | Fine-tune and quantize locally | Limited to vendor-offered options | | Data path | Stays on the user’s hardware if self-hosted | Travels to the vendor’s servers | | Auditability | Community can inspect behavior and code | Vendor controls disclosure | | Government risk view | Provenance and derivative systems hard to track | Contract and single legal entity easier to govern | | Ecosystem effect | Wide third-party tooling and forks | Tooling centered on vendor platform | Each column hides a real irony. The open approach gives users more control and, for the same reason, gives regulators less. The closed approach protects the vendor’s secrets and gives governments a single throat to choke. Which trade a buyer prefers depends on who the buyer is and where the data lives. For a side-by-side view of how three open-weight labs differ in practice, see our [DeepSeek vs Qwen vs Mistral comparison](https://shattered.io/deepseek-vs-qwen-vs-mistral-open-weight-models-2026/). ## Historical Context: Supply-Chain Tools Meet AI Washington has used supply-chain exclusion tools against vendors tied to rival states before, most often in telecom equipment and security software. Law-firm alerts from firms such as Crowell and Haynes Boone place the DeepSeek order in the same family of orders under the Federal Acquisition Supply Chain Security Act. The notable change is the target. This time the product is a model, not a router or an antivirus agent. Models raise a harder inventory problem than hardware. A router has a serial number. A model can be copied, fine-tuned and embedded inside another product without any visible change on the box. Agencies that want to comply with the order have to ask every vendor whether its software calls DeepSeek anywhere in the chain, and vendors have to know the answer. On the Chinese side, the historical pattern runs the other way. Open release has been a way to win mindshare when direct sales into Western enterprises face headwinds. DeepSeek’s January 2025 R1 launch is the clearest example. A model that anyone could download reached engineers directly, bypassing every procurement process that might have slowed a commercial rollout. ## Market Impact: Vendors, Cloud Hosts and Developers The immediate market impact lands on federal contractors. Any company that sells software to civilian agencies now has to check whether DeepSeek models sit inside its products. The order says exceptions need written approval, so vendors that cannot swap out the model quickly will join a queue. Expect compliance teams to spend the next several weeks mapping their dependencies. ### What changes for developers Developers working outside government contracts face no new legal restriction from this order. The order applies to federal civilian agencies and related systems, so private-sector use stays legal. Still, enterprise security teams read federal orders as signals. Some will tighten their own policies, especially in regulated industries that sell to the government. Open-weight models also keep their appeal for local use. People running models on their own machines, such as the setups in our [llama.cpp and Open WebUI tutorial](https://shattered.io/llama-cpp-server-open-webui-gguf-12-steps-2026/), keep the data on their own hardware. That fact is exactly why open weights have a loyal following and exactly why agencies worry about uncontrolled copies. For wider coverage of the sector, visit our [AI and machine learning section](https://shattered.io/ai-machine-learning/). ## What Remains Unconfirmed Several claims circulate around this story without firm support. We separate them below so readers can tell what the record establishes from what it does not. | Claim | Status | |---|---| | Reuters reported the DHS removal order dated October 20, 2026 | Confirmed by Reuters report | | DHS cited “multiple threat vectors” and data-transfer risk to China | Confirmed by Reuters report | | Liang Wenfeng said all DeepSeek models will be open-source | Confirmed by Reuters interview | | Beijing has an official policy that favors open models for cost, innovation or influence | Unconfirmed, no single official statement found | | “Others” universally avoid open models | Unconfirmed, only specific concerns documented | | Release dates and specs for the newest Qwen, Kimi, GLM or MiniMax models | Not established by our fresh research | | Any quantified measure of domestic adoption from open weights | Not found | The search results behind this story also include material dated October 2026 that cannot be independently confirmed from the record alone. We rely on the Reuters report and the notice itself for the DHS order, and we avoid assigning specifications to other Chinese models, because the fresh research did not return verifiable data on them. ## Five Predictions for Open-Weight AI These are our editorial forecasts, not sourced facts. We have kept them free of invented numbers. - **Other agencies and states follow.** Expect procurement offices and some allied governments to review their own DeepSeek exposure, especially inside vendor contracts. - **Vendors publish “no DeepSeek” attestations.** Software suppliers that sell to government will likely add written statements about which models sit in their stacks. - **Chinese labs keep releasing weights.** Liang’s stated position, and the ecosystem logic behind it, give little reason for DeepSeek to reverse course. - **Western firms lean on provenance.** American and European open-weight releases will advertise their origin, audit trail and licensing as selling points. - **Security tooling for model supply chains grows.** Scanners that detect which model sits inside a product will become a procurement checkbox. The takeaway: openness and caution are both rational responses to the same property of open weights. Chinese labs see free distribution as a growth engine. U.S. agencies see free distribution as a loss of control. Neither side is confused about what the other is doing. ## FAQ: China, DeepSeek and Open AI Models ### Why does China favor open AI models like DeepSeek? Analysts point to faster domestic adoption, lower costs, ecosystem building and wider influence. DeepSeek’s founder, Liang Wenfeng, has said the company will not go closed-source and believes a strong technology ecosystem matters more. The search record does not contain a single official Chinese government statement on the reasons. ### What did DHS order on October 20, 2026? Reuters reported that DHS ordered DeepSeek products and third-party services using the model removed from civilian federal-agency IT systems. DHS said the models posed “multiple threat vectors” and could transfer sensitive data to China. ### Is DeepSeek banned for everyone in the United States? No. The order applies to federal civilian agencies and related systems. It is not a general ban on consumer or private-sector use. ### Are open-weight models the same as open-source models? Not exactly. Reuters describes open-source models as models whose base code is publicly available for developers to use and modify. Open-weight models publish the trained weights so users can run and adapt them, though the full training data and code may not be released. ### Do all other countries and companies avoid open models? No. Our record documents specific concerns from some governments and organizations, not a universal position. U.S. developers also build and use open-weight models. ### When did DeepSeek release R1? DeepSeek publicly released its open-source R1 reasoning model in January 2025, according to Reuters. On February 21, 2025, it said it would publish five code repositories the following week. ### What should vendors do about the DHS order? Map every product and service that calls a DeepSeek model, including through third parties. Agencies can request exceptions, but those requests need review and written approval.