Can the US ban Chinese AI models? Partially. Washington can ban Chinese AI apps from app stores, bar the models from government systems, and add their developers to trade blacklists — but it cannot realistically stop Americans running open-weight models such as DeepSeek or Kimi K3 that are already widely downloaded.[1] The question became live in July 2026, when Axios reported that the Trump administration was actively weighing restrictions on Chinese AI models, with options including Entity List designations, an executive order attaching liability to their use, and federal procurement rules — a push that gained momentum after Moonshot AI released Kimi K3 on 16 July 2026, a 2.8-trillion-parameter model that Axios says rivals top American systems at roughly 40% lower cost.[2][3][4] The stakes are large: Chinese-origin models now carry 46.4% of weekly tokens on routing platform OpenRouter, against 35.7% for US models (Forkast, July 2026), and DeepSeek recently became the number-one trending vendor on Ramp’s corporate spend index.[5][6] Analysts say downloadable open weights make an outright ban “nearly impossible to enforce”. Three months on, Washington has used the narrowest tool first: on 19 October 2026 the Department of Homeland Security ordered every federal civilian agency to remove DeepSeek products and stop buying them.[7] Here is what Washington can actually do — and where the law runs out. See our US vs China AI model capability tracker for how close the gap now is.
What Washington can actually ban
The US government has three tools that work in practice, and each targets a chokepoint rather than the model itself. First, app-store bans: the TikTok precedent — the Protecting Americans from Foreign Adversary Controlled Applications Act, signed in April 2024 and upheld 9-0 by the Supreme Court in January 2025 — shows Congress can compel Apple and Google to delist a Chinese-controlled app.[8] Lawmakers including Rep. Josh Gottheimer drew a direct comparison with TikTok when pushing device bans in February 2025.[9] Second, procurement bans: the government can refuse to run adversary models on its own systems, which the pending No Adversarial AI Act would formalise. Third, Entity List designations: adding a Chinese AI lab to the Commerce Department’s blacklist — as happened to Zhipu AI (now Z.ai) in January 2025 — creates a presumption of denial for exports to that company and chills any US commercial relationship with it.[10]
What none of these tools does is claw back software already sitting on American servers. As a 23 July 2026 analysis by law firm Sheppard Mullin notes, Entity List rules restrict sending controlled technology to a listed company — sending data into Zhipu’s hosted API can be a licensable export — but downloading freely published weights and running them on Amazon Web Services or Microsoft Azure “without any data touching a Chinese server” sidesteps the restriction entirely.[11]
What’s already banned today
There’s still no US-wide ban on Chinese AI models. What does exist is a growing set of government restrictions, starting with device bans in early 2025. NASA blocked DeepSeek from agency systems on 31 January 2025, the US Navy warned personnel off it the same week, and the Commerce Department barred it from government devices in March 2025, per Tom’s Guide and Reuters reporting.[12][13] At state level, Texas moved first — Governor Greg Abbott banned DeepSeek from state devices on 31 January 2025, citing “malicious espionage operations by the Chinese Communist Party” — followed by New York on 10 February and Virginia on 11 February 2025 (GovTech).[14] By April 2025, more than a dozen states had enacted similar bans, according to StateTech Magazine’s running tally.[15]
The biggest step so far came on 19 October 2026, when the Department of Homeland Security used its powers under the Federal Acquisition Supply Chain Security Act to order all federal civilian agencies to remove DeepSeek products from their systems, including third-party services built on DeepSeek models, and to exclude them from future purchases.[7] It binds federal civilian agencies only; private companies, state governments and consumers can still use DeepSeek’s app, API and weights. DeepSeek itself is still off the Entity List. An interagency committee approved adding it, along with more than 100 other Chinese firms, but the White House has held back, reportedly to avoid inflaming tensions with Beijing (Reuters).[16]
At federal level, the bipartisan No Adversarial AI Act (H.R.4142 / S.2177), introduced on 25 June 2025 by Reps. Raja Krishnamoorthi and John Moolenaar, would bar federal agencies from using AI models developed in China, Russia, Iran or North Korea. It remains pending in the 119th Congress — introduced, not enacted.[17] Internationally, the picture is thinner than headlines suggest: Italy’s data-protection authority blocked the DeepSeek app in January 2025, but there is no EU-level restriction on Chinese models as of July 2026.[18] Our Asia AI policy tracker follows how governments across the region are treating the same question.
The tools on the table
According to Axios’s 20 July 2026 report, the deliberations — in which OSTP director Michael Kratsios is involved — span a spectrum from de facto bans to softer deterrence: adding Chinese AI labs to the Entity List; an executive order making companies accept legal liability for using Chinese models; NSA and national cyber director advisories warning of security threats; and procurement rules plus public pressure.[4][3] Similar proposals circulated in 2025 but were blocked by pro-open-source officials; adviser Sriram Krishnan’s departure and Kimi K3’s launch have shifted the balance towards the hawks.[4] One official told Axios the administration remains “committed to promoting America’s open-source ecosystem and strengthening its security”; another source described the strategy as “slower and more durable” than a headline ban.
Industry is lobbying in both directions. Anthropic has accused DeepSeek and MiniMax of illicitly distilling US frontier models, per Axios, and leading closed labs have reportedly pressed for restrictions. Outside White House AI adviser David Sacks countered publicly that “the leading closed labs… want the government to eliminate their open-source competition” — a regulatory-capture charge echoed by venture investors including Andreessen Horowitz, which has long argued open-weight restrictions would hurt American startups most. The fight is one front in the broader contest we chart in the China vs US AI race.
Why open weights are nearly impossible to ban
Export-control law was built around chokepoints, and open weights have none. A chip is a physical object at a border; a closed model is an API behind a gate; an open-weight model is a file that, once published, propagates across Hugging Face mirrors, torrents and private servers within days. The Bureau of Industry and Security did try to control model weights once — the January 2025 AI Diffusion Rule created a licence requirement (ECCN 4E091) for advanced closed model weights — but the Trump administration announced its rescission in May 2025, and the rule always carved out published open weights, which the Export Administration Regulations generally treat as outside control once they are lawfully public (Sheppard Mullin, 23 July 2026).[19][11] Retroactively controlling DeepSeek R1 or Qwen weights that have already been downloaded at scale is, in practical terms, unenforceable: NIST’s Center for AI Standards and Innovation (CAISI) reported on 30 September 2025 that DeepSeek model downloads on sharing platforms had grown nearly 1,000% since January 2025.[1]
The security debate also turns on where a model runs. The data-exfiltration concerns behind the app bans — DeepSeek’s consumer app stores chat data on servers in China — do not apply to weights self-hosted on US infrastructure. What CAISI flagged in the models themselves is different: DeepSeek’s most secure model complied with 94% of overtly malicious jailbreak requests versus 8% for US reference models, and its agents were 12 times more likely to follow malicious hijacking instructions.[1] Researchers’ realistic worries are those — plus subtle backdoors or biases surviving fine-tuning — rather than Beijing reading prompts off a locally hosted GPU. Which Chinese models are reachable from which markets, via app, API or download, is tracked in our LLM accessibility tracker.
What it means for companies building on Chinese models
The exposure is real but asymmetric. DeepSeek models carried 5.13 trillion weekly OpenRouter tokens as of July 2026 — 17.6% of the platform, the largest single vendor, with Alibaba’s Qwen at 13.9%.[5][20] The draw is price: DeepSeek’s V4 Flash lists at $0.14 per million input tokens against $5.00 for OpenAI’s GPT-5.5.[5][21] A federal procurement ban would touch government contractors immediately — and, as Sheppard Mullin notes, procurement law follows the model’s developer, not its US host, so running Qwen on Azure does not launder it for a defence contract. An Entity List action or liability executive order would chill Fortune 500 legal departments even where developer use stayed legal. But startups that have already downloaded and fine-tuned Chinese weights would keep running them; the likelier costs are compliance friction, insurance questions, and losing access to future releases — a risk Beijing itself sharpened in October 2026 by proposing export controls on the weights of its most capable models, though weights already published openly fall outside the draft.[22] For the practical playbook, see our enterprise guide to building on Chinese AI models.
Frequently asked questions
Is DeepSeek banned in the United States?
No — not for the public. Since October 2026 it has been barred across all federal civilian agencies by a Department of Homeland Security order, on top of earlier bans at NASA, the US Navy and the Commerce Department, and on state devices in more than a dozen states starting with Texas on 31 January 2025.[7][12][15] Private individuals and companies can legally use the app, API and open weights.
Could the US ban Kimi K3?
It could ban the Kimi app from US app stores under a TikTok-style law and blacklist Moonshot AI via the Entity List. But K3’s full weights have been on Hugging Face since 27 July 2026, and once weights are downloadable, analysts quoted by Axios and Tom’s Hardware say an outright ban becomes “nearly impossible to enforce”.[23][3]
What would the No Adversarial AI Act actually do?
H.R.4142 and S.2177, introduced 25 June 2025, would require federal agencies to stop acquiring or using AI models developed in China, Russia, Iran or North Korea, with limited waivers.[17] It covers government use only — it would not ban private-sector use — and it remains pending, not enacted, as of July 2026.[17]
Is it safer to use Chinese models via API or self-hosted?
Self-hosted, for data security: weights run on US infrastructure send nothing to China, whereas Chinese-hosted apps and APIs store data on Chinese servers. Model-level risks remain either way, and sending controlled technical data into an Entity-Listed lab’s hosted API can itself violate US export law (Sheppard Mullin, July 2026).[11]
Has any country fully banned Chinese AI models?
No major economy has banned the models outright. Italy blocked the DeepSeek app in January 2025 on data-protection grounds — an app-level measure, not a model ban — and there is no EU-level restriction as of July 2026.[18] Government-device bans, as in the US, Australia, South Korea and Taiwan, remain the global norm.
Sources & Further Reading
- CAISI Evaluation of DeepSeek AI Models Finds Shortcomings and Risks — NIST, 2025-09-30
- China just erased America’s AI lead — Axios, 2026-07-17
- Trump administration reportedly reviving push to ban Chinese AI models following Kimi K3 launch — Tom’s Hardware, 2026-07-20
- Trump officials revive push to bar Chinese AI models after Kimi K3 — Implicator, 2026-07
- Chinese AI Models Now Capture Up to 46% of US Enterprise Token Usage — Forkast, 2026-07-10
- DeepSeek topped Ramp’s trending software vendors in June 2026 as US companies chase cheaper AI — The Decoder, 2026-07
- DHS bans Chinese DeepSeek AI from federal civilian agencies — The Hill, 2026-10-19
- Supreme Court upholds TikTok ban — SCOTUSblog, 2025-01-17
- DeepSeek ban proposed in Congress over China Mobile ties — Fortune, 2025-02-06
- US Blacklists Tencent-Backed AI Startup as Tech Curbs Deepen — Bloomberg, 2025-01-16
- US vs Chinese AI models: export control risks — Sheppard Mullin, 2026-07
- DeepSeek AI banned by NASA, US Navy and more over privacy concerns — Tom’s Guide, 2025-02
- US Commerce department bureaus ban China’s DeepSeek on government devices, sources say — ThePrint (Reuters), 2025-03-17
- Where’s DeepSeek banned? The states blocking Chinese-made AI — GovTech, 2025
- These States Have Banned DeepSeek — StateTech Magazine, 2025-04
- US holds off blacklisting DeepSeek and 100-plus Chinese firms — The Next Web (Reuters), 2026-06-17
- H.R. 4142 — No Adversarial AI Act (introduced) — GovInfo, 119th Congress
- Italy’s regulator blocks Chinese AI app DeepSeek on data protection — StreetInsider (Reuters), 2025-01-30
- BIS Rescinds AI Diffusion Rule — Wiley, 2025-05
- Chinese Open Models Route 46% of US Tokens as Washington Weighs Ban — AI2 Work, 2026-08-04
- DeepSeek V4 Flash API pricing — ZenMux, 2026-08-05
- China Considers Export Controls on AI Models, Data and Chips — TechRepublic, 2026-10
- moonshotai/Kimi-K3 — Simon Willison’s Weblog, 2026-07-27