Despite security concerns, US companies might adopt Chinese open-source models like GLM because they can be hosted on US hardware with no data leakage. The immense cost savings and ability to maintain full control over the stack make them a practical alternative to expensive, risky frontier models.

Writer built its flagship model on China's GLM 5.2, an open-weight model. The CEO argues this is a non-issue for enterprises, as the model's weights are open (MIT license) and all post-training, hosting, and security are handled by a US company on US infrastructure, neutralizing geopolitical and security concerns.

AI company Clay uses powerful open-weight models from overseas for sensitive applications. They manage security concerns not by avoiding the models, but by hosting them with stateside inference providers like Base 10 and Fireworks, keeping the data and processing within the U.S. and under their control.

The key distinction between open-weight and closed models is access. Open models provide both the software runtime and the crucial parameter "weights" for self-hosting. Closed models restrict access to one or both, typically offering functionality only through a managed API.

To mitigate risks of sharing sensitive data with cloud AI, use tools like LM Studio. These applications allow you to download and run powerful open-source models directly on your laptop, ensuring that your financial statements or insurance policies are analyzed without ever leaving your device.

Most nations' sovereign AI strategies will not involve creating frontier models from scratch. Instead, they will adopt the best open-source models, customize them with local data and values, and run them on-premise for national security.

Sending proprietary enterprise data to external foundational models is a critical mistake that 'leeches' value and intellectual property. The correct, secure approach is to bring AI models into a company's own air-gapped or on-premise environment to maintain data sovereignty and control.

Hosting a foreign model on your own infrastructure does not eliminate security threats. Attackers can embed hidden triggers, like specific character sequences, during training. When an external user inputs that trigger, it can jailbreak the model, causing it to exfiltrate sensitive company data.

With open-weight models, the user has full control, transparency, and access, mitigating risks of bias or manipulation from the creator. This is fundamentally different from using a foreign-hosted API, where you send them your data and they control access, making provenance a critical security concern.