Silicon Valley’s largest technology companies are resisting broad restrictions on Chinese open-weight AI models, warning that sweeping controls could reduce competition and leave businesses more dependent on proprietary providers.
A July 24 industry letter backed by Nvidia, Microsoft, Meta, IBM, Dell, Palantir, and other organizations urged US policymakers to preserve access to models that can be downloaded and operated independently. Anthropic did not sign, but CEO Dario Amodei said on July 27 that the company opposed a blanket ban and favored targeted controls based on model capability and misuse risk.
An open-model fight with higher stakes
The open-weight industry letter argues that downloadable models support research, competition, and greater control over data and infrastructure. Nearly 200 US startups have also pushed back against broad restrictions, saying reduced access to inexpensive models could raise development costs and strengthen larger proprietary vendors.
Chinese developers are central to the dispute because models from Alibaba, DeepSeek, Moonshot AI, and other laboratories are widely available through public repositories and third-party services. Their lower costs come with governance trade-offs involving hosting, data residency, licensing, and vendor accountability. Businesses may use the models directly or encounter fine-tuned versions inside other software.
Anthropic’s position is more limited than a general prohibition. Amodei supports keeping lower-risk open models available while requiring stronger evaluations for systems capable of causing serious harm. He also called for tighter chip controls and action against industrial-scale extraction of proprietary model outputs.
Model distillation — training one system with outputs from another — is a standard machine-learning practice. The dispute centers on alleged extraction at a scale or through methods that may violate providers’ terms. US officials are considering sanctions and other trade restrictions against Chinese developers accused of improperly extracting outputs from American models. China’s Commerce Ministry rejected the allegations and threatened countermeasures on July 27.
The controls enterprises can apply now
Model origin alone does not determine enterprise risk. Teams must establish where inference occurs, what information leaves approved systems, which executable components accompany the weights, and whether the license permits the planned commercial use.
Security testing should cover the exact model and deployment configuration. A NIST evaluation of DeepSeek found that agents using the most secure DeepSeek model it tested were 12 times more likely than US reference models to follow malicious instructions during agent-hijacking tests.
In a separate jailbreak evaluation, the DeepSeek model answered 94% of overtly malicious requests, compared with 8% for the US reference models. The results apply to the tested DeepSeek versions, not every Chinese or open-weight model.
Platform teams should inventory base models, fine-tunes, and adapters across development, staging, and production. Records should include the source, version, license, file hashes, hosting location, and owner.
Security teams should also inspect model loaders and dependencies. A Hugging Face Transformers vulnerability showed how a malicious configuration could trigger remote code execution in affected environments. Teams should test for prompt injection and agent hijacking and document whether prompts, outputs, or telemetry cross organizational boundaries.
No blanket US ban on Chinese open-weight models had been enacted as of July 28, 2026. A model inventory, documented data flows, and a tested replacement plan can reduce exposure if licensing, security findings, or government policy changes.
Read more: Nvidia’s new alliance for securing enterprise AI shows how vendors are developing shared monitoring and containment tools for autonomous systems.