Summary of Key Points
This article focuses on the concept of "AI deployment sovereignty": Satya Nadella, CEO of Microsoft, has raised the "reverse information paradox" to highlight the core drawbacks of closed-source AI (companies that use AI must also contribute data for free), advocating for a shift towards more localized and autonomous AI solutions. Domestic open-source models such as Kimi K3 and GLM-5.2 have reached the global forefront, directly challenging the Western control systems that rely on cloud services and chips. The Atlantic Council uses the "trust-intelligence-power triangle" framework to analyze how, with AI models running locally, intelligence is no longer a scarce resource. The new focus of competition has shifted to computational sovereignty, model traceability, and trust governance, rendering existing Western regulatory mechanisms ineffective. Consequently, adjustments are needed in legislation, trust systems, and security assessments.
1. The Reverse Information Paradox: Is Closed-Source AI Really "Working for Someone Else"?
Nadella's concept highlights the issue with closed-source AI: when using models like ChatGPT, companies not only pay subscription fees but also provide their business data (such as customer needs and internal processes) to the service providers. Each query or correction actually helps train the models, yet the resulting intelligence belongs to the service providers. For example, if a factory uses closed-source AI to optimize production processes, it must transfer its data to the AI company, which may then use that data to improve its models and potentially sell them to competitors. Nadella argues that companies should retain the value they create by keeping the models on their own hardware (such as company servers or employees' computers), thus achieving autonomy.
2. Why Have Domestic Open-Source Models Become the Best Option for "Ownable AI"?
The main difference between open-source and closed-source models is that open-source models have their source code and parameters available for download, allowing them to be run locally without relying on the cloud. Domestic open-source models are now competitive:
- Performance: Kimi K3 (with 2.8 trillion parameters) ranks fourth globally, only behind OpenAI and Anthropic's top closed-source models; GLM-5.2 is the leading open-source model, and Alibaba's Qwen3.8-Max also plans to make all its weights available.
- Hardware Support: This year, Microsoft and NVIDIA introduced hardware (e.g., RTX Spark accelerators) capable of running 120 billion-parameter models on laptops, with over 30 mainstream laptops set to support this, making local model execution feasible for both individuals and enterprises.
- National Strategy: The Shanghai World Artificial Intelligence Conference identified open-source AI as a historic opportunity, with China aiming to become a key provider of international AI products and promoting it through the World Artificial Intelligence Cooperation Organization (WAICO). This marks a shift from being a follower to a co-maker of rules.
3. The Failure of Western Regulations: Why Are Chip and Data Controls Ineffective?
Western AI regulations relied on two main strategies: restricting high-end chip exports and preventing data transfer to Chinese servers. However, open-source models have rendered these measures ineffective:
- Data Control: Open-source models run on users' own hardware, eliminating the need to transfer data to the cloud. For example, Germany's ban on DeepSeek was due to concerns about data leakage, but this issue doesn't arise with local deployment.
- Chip Control: Chinese open-source models are stored on overseas servers (e.g., Frankfurt, Dallas), so Western chip export restrictions don't apply. Additionally, China uses multilateral platforms like WAICO to distribute these models, bypassing chip restrictions. In short, Western regulations were designed for centralized cloud systems, but AI has moved towards local deployment, rendering them obsolete.
4. The New Focus of Competition: The Trust-Intelligence-Power Triangle
The Atlantic Council's "trust-intelligence-power triangle" framework outlines three dimensions of AI competition:
- Power: Resources like chips, power grids, and export licenses (where the West holds an advantage, though China is constrained by chip restrictions).
- Intellectual Property: Models, data, and talent (formerly dominated by Western models, but now open-source models have made intelligence less scarce).
- Trust: The reliability of models (e.g., whether they contain backdoors or have transparent data sources; the West lacks a trust framework for open-source models).
With intelligence no longer being a scarce resource, the focus has shifted to computational sovereignty and model trust. For example, Western open-source teams cannot use local closed-source models due to regulatory restrictions, forcing them to rely on Chinese models like Kimi K2.5.
5. Western Responses: How to Fill the Gap in Autonomy?
The Atlantic Council suggests three approaches:
- Legislative Changes: Western leading labs (e.g., OpenAI) prohibit using their models to train smaller models, forcing Western teams to use Chinese models. Legislation should allow for the "reasonable use" of training data to enable the advancement of Western open-source models.
- Building Trust Systems: All open-source models should disclose their origins and modifications, along with any security vulnerabilities. Microsoft's Foundry Local platform already has foundational features that could become industry standards with mandatory disclosure requirements.
- Enhanced Security Assessments: Instead of focusing on the cloud, security assessments should now target the models themselves, ensuring they are secure before being used in government or sensitive tasks.
In summary, future AI competition will not be about which model performs best but about who can provide safe and autonomous access to AI solutions. The West must address its shortcomings in open-source models to prevent China from dominating core AI infrastructure.