DeepSeek’s improving performance gives businesses a reason to evaluate its models, but federal procurement restrictions change the calculation for government suppliers. The compliance question now reaches beyond the tool a company buys to the underlying model it uses. For organizations serving both public and private customers, that distinction calls for clearer separation and stronger oversight.
Federal Restrictions Follow the Model Into Supplier Products
On October 17, DeepSeek released DeepSeek-V4.1-Flash-1017, claiming stronger coding capabilities and quicker execution of complex tasks. Two days later, the U.S. Department of Homeland Security disclosed a summer order requiring civilian federal agencies to exclude and remove the company’s technology, including versions delivered by other suppliers.
Homeland Security Secretary Markwayne Mullin issued the order under the Federal Acquisition Supply Chain Security Act of 2018. It became effective upon issuance. Its reach includes the State Department, Treasury Department and Internal Revenue Service (IRS).
For suppliers, the third-party coverage is particularly consequential. An application can carry another vendor’s branding while still depending on DeepSeek underneath. Procurement review therefore needs to establish model ancestry, including fine-tuned versions, rather than stop at the name on an invoice.
Trade lawyers at Sheppard Mullin explain that these restrictions follow the model’s developer rather than its hosting location. Running DeepSeek in an American data center does not make it eligible for covered federal work.
The civilian order excludes specified national security systems, which operate under separate authorities. Suppliers should not read that boundary as permission: defense and intelligence restrictions require their own assessment.
Contract Obligations Make Discovery Time-Sensitive
The defense market already has a separate statutory restriction. Section 1532 of the fiscal 2026 defense authorization law prohibits Pentagon use of AI from DeepSeek and High-Flyer, its parent. It also prohibits contractors and subcontractors from using that technology when performing defense contracts. Only the defense secretary can authorize a case-specific waiver.
Law firm WilmerHale interprets the provision as extending to AI systems domiciled in China and other adversary nations, beyond the named companies. That interpretation warrants legal review before a supplier assumes another Chinese model is acceptable for defense work.
Contract language turns these restrictions into operational duties. FAR 52.204-30 covers both suppliers and subcontractors, incorporates new orders through contract modifications, and sets short deadlines after covered technology is discovered.
| Contract pressure point | Operational implication |
|---|---|
| Supply-chain order checks at least every three months | Assign responsibility for reviewing the federal contracting portal and assessing changes. |
| Reporting within three business days | Give employees and subcontractors a clear escalation route. |
| Remediation explanation within 10 more days | Prepare a replacement and containment process before an issue arises. |
| Coverage of technology used during performance | Review internal tools as well as customer-facing deliverables. |
The exception process also illustrates why public disclosure is a poor trigger for compliance work. Agencies had 60 days to submit written requests supported by a compelling justification for review by the Cybersecurity and Infrastructure Security Agency. Measured from July 24, that period ended around Sept. 22. The October disclosure did not reopen it.
A supplier that waits for broad publicity may already be behind the applicable timetable.
Commercial Use Remains Possible, With a Different Risk Boundary
The DHS restriction does not place DeepSeek on the Commerce Department’s Entity List, and no federal rule currently prohibits a U.S. business from using it solely for private customers. A company with government contracts can retain a commercial deployment if it keeps that deployment separate from federal work.
Export controls create another consideration. Sending controlled technical information to DeepSeek’s China-hosted app, website or API can amount to an export requiring a license. Downloading model weights or using a U.S.-based cloud avoids that particular transfer issue. It does not resolve procurement eligibility.
Consider a hypothetical software supplier using DeepSeek for a private-customer product while delivering a separate service to a civilian agency. Hosting both services domestically would not, by itself, establish compliance. The supplier would need to check whether shared employee tools, subcontractor applications or backend components introduce the restricted model into federal performance.
Separation therefore needs evidence. Distinct product labels are insufficient if the underlying workflows still overlap.
Lower Inference Costs Do Not Settle the Purchasing Decision
DeepSeek’s commercial appeal remains credible enough to justify testing. Its MIT-licensed weights allow businesses to run and adapt the model themselves, while reported performance gains suggest a stronger candidate for software-development tasks.
DeepSeek reports a DeepSWE v1.1 score of 77.0, compared with 74.2 previously and 74.0 for Claude Opus 5. Its Terminal-Bench 4.0 result increased from 31.2 to 42.3, although Claude Opus 5 scored 51.8. These vendor comparisons do not establish superiority across enterprise workloads.
Artificial Analysis separately scored V4.1-Flash-1017 at 45, versus 39 for its predecessor, and recorded average evaluation spending falling from $0.49 to $0.26 per task.
Those results support evaluation, not an automatic purchase decision. A mixed public-private supplier should also account for separate testing, oversight and the potential expense of replacing a model. Self-hosted and U.S.-cloud deployments have their own cost profiles.
Pending Expansion Makes Portability a Practical Control
The Senate Armed Services Committee’s fiscal 2027 defense-bill version would add ten Chinese AI companies to the Pentagon restriction and include derivative models. It would require removal guidance from the defense secretary and prohibit contractor use 90 days later. These are proposed measures, not current obligations.
Derivative coverage would make rebranding or fine-tuning an unreliable route around restrictions. Separately, the White House considered limits on Kimi and Qwen during the summer. Current civilian rules do not name Qwen, Kimi or GLM, but suppliers should distinguish present permission from long-term dependency risk.
Organizations can prepare without treating every proposal as enacted law:
- Require vendors to identify base models, derivatives and hosting locations.
- Review employee tools and subcontractor software alongside delivered products.
- Test an alternative against actual workloads, including accuracy, latency and cost.
- Use explicit model versions and design integrations that allow replacement without a product rebuild.
For federal-facing teams, the immediate priority is identifying and removing covered DeepSeek dependencies, with counsel and contracting officers involved where scope is uncertain. Commercial teams can preserve their options, but should document the separation and budget for maintaining it. Model provenance belongs in procurement records and change controls before the next restriction forces an urgent review.
