Behind the U.S. Government Procurement Ban: Why Does DeepSeek Affect All Chinese Model Companies?

A procurement rule recently announced by the U.S. Department of Homeland Security requires federal civilian executive agencies to stop using DeepSeek-related products and extends the restriction to third-party products that integrate the DeepSeek model underneath. This means that even if a U.S. software company uses its own brand, as long as its product calls the DeepSeek model, it may affect its eligibility to supply government departments. For Chinese large-model companies that are accelerating their overseas expansion, this is far more than a ban targeting a single company—it is an industry signal worth unpacking in depth.

The Substance of the Ban: How Government Procurement Transmits to the Commercial Market

The rule was originally issued on July 24, with a public pre-release text on October 19 and scheduled publication in the Federal Register on October 20. It is not a temporary response to DeepSeek^Ad's recent model upgrades, but rather an exercise of administrative authority under the U.S. Federal Acquisition Supply Chain Security Act of 2018, requiring no separate congressional legislation. Its direct scope is procurement and information systems of federal civilian executive agencies; it does not comprehensively prohibit U.S. companies and individuals from using DeepSeek.

The key change lies in the transmission mechanism. Government procurement qualification is often regarded as a certification label of corporate credibility. A company serving both government and private clients can choose to continue using DeepSeek in its private business, but the government version must switch models. Compliance requirements in contracts also propagate along subcontracting relationships, and companies need to assess the cost of maintaining two versions, the expense of retesting, and the uncertainty of future adjustments.

This path is similar to the commercial impact of domestic government procurement blacklists: government orders themselves may not be the main source of revenue, but the spillover effect of qualification certification will change companies' technology choices. The logic by which U.S. companies choose Chinese models is being reshaped by government procurement rules.

Federal Acquisition Supply Chain Security Act (FASCSA) is U.S. legislation passed in 2018 that grants the Federal Acquisition Security Council the power to review and exclude "high-risk" suppliers in the supply chain, enforceable without separate congressional legislation. The Act establishes three types of exclusion measures: exclusion from a single agency, government-wide exclusion (decided by a government-wide council), and exclusion targeting specific products. The Department of Homeland Security's invocation of this Act means its action is administrative enforcement within an existing legal framework, not a newly created trade barrier. This also explains why the rule can take effect quickly and extend to third-party integrated products: FASCSA's review logic focuses on "supply chain origin" rather than the product itself. As long as the underlying technology involves a restricted supplier, even if repackaged or hosted, it may still fall within the scope of restriction. For overseas-expanding companies, understanding the nature of this legal tool helps predict the boundaries and replicability of restrictions.

Under Cost Pressure, Why U.S. Companies Turn to Chinese Models

The real driver of U.S. companies' adoption of Chinese models is that, as AI enters core business, cost begins to determine whether a product can be profitable. According to a case cited by Bilibili creator Liu Dian in a video, the U.S. AI assistant Lindy disclosed that after completing evaluation, prompt adjustment, and deployment optimization, the company migrated most of its hosted agent model traffic to DeepSeek, reducing some inference costs by about 90% while retaining some high-capability model calls. The AI programming software Goeser also confirmed in a technical report that Goeser 2 continues training based on Kimi K2.5, using Chinese open-weight models to develop its own programming model. Another platform publicly disclosed that it used Tongyi Qianwen in customer service business procurement.

The common thread in these cases is: reducing daily operating costs while gaining more control over product development. Customer service, email processing, information extraction, and routine programming tasks do not require the most expensive top-tier models; companies match task capability and price. An agent may call models multiple times to complete a task, and the difference in per-call cost will be amplified into a significant profit difference across a large customer base and continuously running business. Media data show that in September, DeepSeek accounted for 58% of token usage on a certain AI gateway but only 6.6% of overall output, indicating that high call volume and high cost-effectiveness can coexist.

This is affecting the pricing space of U.S. model companies. Top models can still command a premium for the hardest tasks, but customers with daily stable tasks already have more choices. Competition brought by Chinese models means that model companies' capability advantages must continuously translate into prices customers are willing to pay.

Open-weight model refers to a model whose parameters can be publicly downloaded and that allows companies to deploy or fine-tune it themselves, unlike closed-source models that only offer API access. Chinese models such as Kimi K2.5 adopt an open-weight strategy, enabling U.S. software companies like Goeser to continue training their own programming models based on them, thereby achieving product differentiation while maintaining cost advantages. The open-weight model reduces companies' dependence on external APIs and facilitates localized deployment and data processing, which is attractive to application companies that need to control costs and data sovereignty. However, this is precisely the focus of government procurement review: the traceability of the underlying technology origin. Open weight does not equal completely free use under an open-source license; companies still need to comply with the licensor's terms and disclose model origin in compliance scenarios. For Chinese model companies, open weight is an effective means of expanding overseas adoption, but it also requires them to establish clear licensing and traceability mechanisms.

The Replicability of Supply Chain Review Logic

This order does not automatically ban other Chinese models; further restrictions still require corresponding decisions. But some of the judgments raised by the Department of Homeland Security involve the supplier's jurisdiction, data processing, and software supply chain assessment, and these logics are replicable.

The traceability of DeepSeek implementations in third-party products shows that government procurement looks at the underlying technology origin; a U.S. brand or U.S. hosting does not automatically eliminate restrictions. In the video, Liu Dian judges that what is worth watching next is whether model origin review will be incorporated into more procurement scenarios, whether it will spread to more models, and whether companies should adjust product design accordingly in advance.

There is another easily overlooked fact: overseas adoption of Chinese models is also creating revenue for U.S. companies. A platform disclosed that more than 70% of output related to Chinese models on its gateway occurs through U.S. service providers. The chain of Chinese teams developing models, U.S. service providers running them, and U.S. application companies purchasing them has become quite common. Interests within U.S. companies are not aligned: application companies want to reduce costs, hosting platforms earn service revenue, and foundation model companies face pressure from replaced workloads.

AI gateway is an intermediate layer between applications and multiple model APIs, responsible for routing requests, managing token usage, monitoring cost and performance, and enabling multi-model switching. The AI gateway data mentioned in the article—DeepSeek accounts for 58% of token usage but only 6.6% of output—reflects the metering characteristics of gateways: token usage measures the total input and output processed by the model, while output share reflects the share of final generated content. High token usage and low output share indicate that DeepSeek is heavily used for long-context processing, multi-turn calls, or background tasks, scenarios that do not demand high single-output quality but are cost-sensitive. Gateway data therefore becomes an important window for observing enterprise model selection, and also explains why U.S. service providers can benefit from Chinese model traffic: gateway hosting and routing services themselves create revenue, regardless of model origin.

Directions for Chinese Model Companies

The ban itself does not provide a final answer, but it reveals where competition is taking place. A model being downloaded means the technology begins to spread; a model entering companies' daily tasks means industrial relationships begin to take shape. For Chinese model companies, two things must be done at the same time: continue to provide technology that customers are willing to adopt, and turn that adoption into reliable service, stable revenue, and long-term cooperative relationships.

The real question is not whether a particular ban is severe, but whether model origin review will become a routine variable in overseas commercial systems. Companies choose models based on cost and performance, while government rules add conditions to different choices; this tension will persist. The DeepSeek case shows all Chinese model companies that overseas competition takes place not only on training grounds and leaderboards, but also in every customer deployment and renewal decision.

One-line Summary

U.S. government procurement bans transmit to the commercial market through supply chain compliance, and Chinese models going overseas face normalization of review.

Paragraph Summary

The U.S. Department of Homeland Security, under the Federal Acquisition Supply Chain Security Act, requires federal civilian agencies to stop using DeepSeek products and extends restrictions to third-party products integrating DeepSeek. Although the ban does not directly prohibit commercial use, government procurement qualification and contract compliance requirements propagate along the supply chain, forcing companies to assess the cost of dual versions. The core driver of U.S. companies turning to Chinese models is cost: after Lindy's migration, some inference costs fell by about 90%, and DeepSeek accounted for 58% of token volume on a certain gateway but only 6.6% of output. The jurisdiction, data processing, and software supply chain assessment logic involved in supply chain review is replicable and may spread to more Chinese models. Overseas adoption of Chinese models also forms a chain that benefits U.S. service providers, and U.S. internal interests are not aligned. For Chinese companies, technology adoption must be translated into reliable service and long-term cooperation to cope with model origin review becoming a routine variable.