Last month, NVIDIA’s CEO launched an initiative advocating for an open AI by publishing the three-page “Open Weights and American AI Leadership” letter. Soon afterward, major AI companies signed the letter, except for one of the big players in the field: Anthropic, along with a few other companies. Why Nvidia—the GPU manufacturer leader—specifically is bankrolling this campaign has generated a lot of discussion and speculation!

The letter asks the US government and American policymakers to embrace open-weight AI rather than restrict it. The signatories of the letter are convinced that open-weight AI is the guarantee of broader access to AI, increased competition, less dependence on a few AI providers, and the strengthening of AI security through wider testing and scrutiny.

The letter builds on the history of computer software and how the open source philosophy changed the software landscape and the Internet, to claim something similar for AI.

From Open Source Software to Open AI

The Free Software movement, launched in the early 1980s by Richard Stallman’s GNU Project, pushed for a transparent ecosystem where developers around the world could study, modify, and share software. The Open Source movement emerged later, in 1998, reframing the same underlying freedoms in more business-friendly terms. Together, the two are now commonly referred to as FOSS. Richard Stallman, the central figure of the Free Software movement, defined it as software where the users have the freedom to run, edit, contribute to, and share the software, making it explicitly not linked to “fees” but rather to the liberty of the user.

Without FOSS, the Internet ecosystem would be totally different. Free and Open Source Software support most of the underlying Internet infrastructure and major protocols and services. In SAC132 report, “The Domain Name System Runs on Free and Open Source Software (FOSS),” published by ICANN’s Security and Stability Advisory Committee on September 25, 2025, the SSAC highlights the critical role that FOSS plays in the operation, security, stability, and resilience of the Domain Name System (DNS), underlining the extent to which the global DNS ecosystem depends on openly developed and widely deployed software. It has been proven that the availability of source code for FOSS implementations has helped to identify and address vulnerabilities, often more quickly than in proprietary systems.

Can Open-Source Principles Be Applied to AI?

The question is to what extent open-source principles can be applied to AI?

The open source initiative makes no difference between open source software and open source AI in the sense that the latter is an AI system made available under terms and in a way that grants the freedoms to use, study, modify and share the system for others to use for any purpose. However, a system is not like software. In fact, an AI system and more specifically a machine learning system, includes the AI model that consists of the model architecture, model weights and the inference code for running the model. In order to reproduce, verify and modify an AI system, one also needs access to the data that was used in training the AI model. Data is much more significant in building an AI system unlike software, where the source code is the primary artifact.

Without sufficiently detailed information about the data used to train the AI model, we can not adequately assess whether the system contains bias and determine its source. The code should also be made available for anyone who wants to use or modify it for any purpose.

Open Weights Are Not Open Source

The signed letter does mention open-weight AI rather than open source AI, and the distinction is significant!

Certainly, the open-weight initiative remains an important milestone towards opening up AI systems for everybody and for all purposes; however, it still lacks the requested level of transparency many researchers and practitioners are advocating for.

The Security Case for Closed Models

This triggers us to ask the next question: Why do some labs refuse to open their platforms?

Most frontier AI labs cite security concerns as a primary reason for keeping their models closed, while also acknowledging the need to protect their business interests and competitive advantages. Once the model weights are available, safeguards can be dropped or deleted. It becomes hard to track accountability when the AI system is used in a bad way. A recent report published by SaferAI, a nonprofit organization publishing reports on risks of AI systems, shows that GLM-5.2, Zhipu AI’s open-weight flagship model did not refuse any of the tested offensive-security or biological tasks, reinforcing the position of those rejecting the call for an open-weight AI.

Supporters of opening AI systems argue that making the model weights available will help defenders prepare for potential cyberattacks.

Open to Whom?

Considering the ongoing claim for an open-source or an open-weight AI, which entity or country will actually benefit from it?

Opening an AI system is not the same as making the software source code open. The infrastructure layer in AI plays a decisive role, as many scholars have already raised this concern¹. The high costs of hardware and infrastructure requirements, including GPUs, electricity and cloud platforms, constitute a barrier to widespread adoption and to AI sovereignty, even using open approaches.