US removes DeepSeek from most civilian government systems as AI rivalry hardens
The US has ordered DeepSeek technology out of most civilian government systems, escalating a fight over artificial intelligence security, supply chains and strategic dependence on Chinese software. The Department of Homeland Security moved to bar the model across agencies that handle major federal functions, including the Treasury, State Department and Justice Department. DeepSeek was also prohibited in any technology built with its models, widening the restriction beyond direct use of the company’s own tools.
One sentence in the order carried more weight than its bureaucratic language suggested: the ban already took effect at the end of July.
The decision arrives at a moment when open-source AI has become both a commercial advantage and a security headache. DeepSeek rose quickly because it is free to access and relatively inexpensive to deploy, a combination that has made it attractive to startups and to major technology firms looking to lower costs. In plain business terms, it offered performance at a price point that undercut many closed systems from US rivals such as Google’s Gemini, Anthropic’s Claude and OpenAI’s Codex.
Security fears, without a public technical explanation
DHS did not spell out the exact security concerns behind the ban. That silence leaves the order anchored more in strategic distrust than in publicly disclosed technical defects, but the policy logic is easy to see: federal agencies are being told to avoid a system built by a Chinese firm at a time when Washington sees AI as a critical domain of competition. The White House has already pressed the argument that rival firms in China have been engaged in “malicious exploitation” of American AI companies, a claim that has shaped how policymakers now frame the entire market.
The exclusion of the Department of Defense from the order shows how selective the response is. Civilian agencies are being pulled away from DeepSeek, while the military branch remains outside this specific directive. That split suggests the government is treating the model as a risk in day-to-day administrative and regulatory environments first, rather than imposing a total federal shutdown across every branch at once.
The practical problem is simple: once a model is woven into software products, the exposure is no longer limited to a single chatbot window. It can sit inside coding tools, customer platforms, automated workflows and cloud services used by federal workers or their contractors. A restriction on the model therefore reaches farther than a ban on consumer use alone.
DeepSeek’s rapid rise made it hard to ignore
DeepSeek’s popularity has grown sharply since its public release in 2025, with demand high enough to knock its servers offline at times. Its free chatbot has also become one of the most downloaded apps in Apple’s App Store rankings, giving it visibility far beyond specialist developer circles. For many users, the attraction has been straightforward: lower cost, easier access and performance that online users say rivals far more expensive models on routine coding and software tasks.
Those cost differences have been central to the shockwaves around the company. DeepSeek said last year that it spent about $6 million to train its debut model, while the costs tied to OpenAI’s GPT-4 were roughly $100 million. Those figures, whether taken as exact accounting or broad comparison, rattled investors because they suggested the economics of frontier AI could shift faster than expected. Tech stocks moved sharply on the news, reflecting how quickly model efficiency can turn into market anxiety.
For startups, the attraction is obvious. Open-source systems let smaller companies build products without paying premium fees for every query or workflow. A low-cost model lowers the barrier to entry, but it also creates a new dependency: if the model becomes embedded deeply enough, companies can inherit the geopolitical risk of the provider as well as the technical advantages.
US firms and policymakers are converging on the same warning
American AI companies have not stayed neutral in the argument. Anthropic and OpenAI have both said that their models were effectively copied by some Chinese firms, including DeepSeek. Those claims sharpen the political case for tighter controls because they frame the issue not only as foreign competition, but as appropriation of American research and engineering effort.
Michael Kratsios, the White House tech lead, said earlier this year in an internal memo that “malicious exploitation” of American AI firms had been taking place, primarily by rival firms in China. That language matters because it links product competition to national security rather than to ordinary market rivalry. When the same models that help write software also help structure federal workflows, the line between commercial tool and strategic vulnerability becomes thin.
Trump has also made clear that he views China as the central US rival in the race to dominate AI, a contest he sees as pivotal to the country’s future. The current ban fits neatly into that view. It treats AI not as a generic productivity tool, but as infrastructure that can transmit advantage, dependency and surveillance risk across government systems.
DeepSeek’s availability to US tech firms had helped normalize it in the market. Snowflake, Vercel and OpenRouter have all incorporated DeepSeek into offerings sold to customers, which means the company’s influence has already spread well beyond its own branded products. Pulling it out of government systems may prove straightforward on paper, but the wider task is harder: identifying every downstream product that now depends on it.
The policy fight is shifting from access to trust
The bigger issue is no longer whether a low-cost Chinese model can compete. It clearly can. The question is whether the US government, and eventually the wider US technology sector, is willing to keep using systems tied to an adversarial state while the strategic stakes around AI keep rising.
DeepSeek’s latest version was released last week, and online users have praised it for handling common coding tasks especially well. That timing may make the US ban feel abrupt, but the decision reflects a deeper pattern: as AI systems become cheaper, faster and more deeply embedded in software pipelines, governments are moving from curiosity to restriction. The unresolved problem is how far those restrictions will extend once commercial dependence is already in place.