U.S. Bans DeepSeek: What Small Businesses Should Know

The Ban, In Plain English

On Monday, the Department of Homeland Security made public an order requiring every federal civilian agency — Treasury, State, Justice, and the rest — to pull DeepSeek’s AI models out of their IT systems and stop buying anything built on them. That last part matters more than most coverage is letting on: the order doesn’t just cover DeepSeek’s own apps. It covers third-party products and services that use DeepSeek technology underneath, even if the vendor’s name on the invoice is a US company.

The stated reasons are what you’d expect: “multiple threat vectors,” sensitive data potentially flowing to Chinese authorities, and cybersecurity risk. The original exclusion letter was signed back in June; this week’s Federal Register notice just drags it into daylight. China’s Foreign Ministry fired back on Tuesday, calling the move an “overstretching of the concept of national security.”

Here’s the part small business owners keep asking us: is DeepSeek now illegal for private companies in the US? No. This is a government supply-chain restriction under the Federal Acquisition Supply Chain Act — it binds federal procurement, not private use. If you were using a DeepSeek-powered tool this morning, nobody is coming for your laptop.

Why This Still Matters If You’re Not a Federal Contractor

Three reasons this shouldn’t scroll past your feed:

1. “Built using DeepSeek” is invisible to most buyers. The feds are now required to audit their tool stack for model provenance — which model actually powers that transcription app, that chatbot, that analytics add-on? Most small businesses have never done that audit once. If a vendor rebrands a DeepSeek-derived model, you’d never know from the pricing page. We hit exactly this problem when we looked at gray-market AI API resellers — provenance is the security question nobody asks.

2. Government customers will pass this down. If you sell software or services into any federal agency — directly or through a prime contractor — expect flow-down clauses requiring you to certify your stack is DeepSeek-free, including embedded models. Agencies can request exceptions, but you don’t want your contract riding on one.

3. It’s a template, not a one-off. This is the second DeepSeek restriction in under a year (a December 2025 law already blocked it from intelligence systems). The mechanism — supply-chain exclusion by vendor origin — can be reused for any foreign model vendor. The smart move is building a process, not a DeepSeek-specific panic.

The 30-Minute Stack Audit We Run

When a restriction like this lands, we run the same quick audit across our tools and our clients’ tools. It takes about half an hour:

• List every AI tool touching customer data — chat assistants, transcription, CRM enrichment, support bots, content generators.
• Find the model behind the marketing. Check the vendor’s docs, model card, or privacy policy. If it says “proprietary model” with no named provider, email support and ask directly which foundation model powers it and where inference runs.
• Check the data path. Where do prompts and uploads get processed and stored? A model can be fine; a data route through a jurisdiction you’re not comfortable with is the actual risk.
• Note the swap cost. For each tool, note what it would take to replace it in a week. That’s your real risk register — not fear, just a plan.

When we ran this on our own stack last year, two tools turned out to route inference through resellers we couldn’t verify. Both were replaced within a week — not because of any ban, but because we couldn’t answer the provenance question. That’s the standard worth copying.

Should Small Businesses Drop DeepSeek?

Honest answer: it depends on what you do with it. Running an open-weight DeepSeek model locally on your own hardware — where no data leaves the building — is a very different risk profile from pasting client data into a hosted endpoint. That distinction between open weights you run yourself and hosted services you feed data to is exactly why we keep watching open-weight models you can run on your own machines.

Our line, and the one we recommend to clients: no customer data goes into any hosted AI tool whose underlying model and data jurisdiction you can’t verify. That rule survives whatever ban comes next.

The Takeaway

The DeepSeek order changes almost nothing for private small businesses today — but it hands you a useful forcing function. If the federal government thinks auditing model provenance is worth doing across thousands of systems, one afternoon spent auditing your own dozen tools is cheap insurance. Do the audit, note the swap costs, and you’ll never have to scramble when the next ban headline names a vendor in your stack.

Want help figuring out which AI tools in your business are safe to keep — and which to replace? SQUIDBOT maps your stack, flags provenance and data-path risks, and builds the swap plan. Take a look here.