Nvidia, Microsoft, and Meta compete across the artificial intelligence industry, but they increasingly agree that U.S. businesses and developers should have greater access to open-weight AI models. 

What Open-Weight AI Means

An AI model’s “weights” are the numerical settings it develops as it learns from training data. The weights store patterns the model has learned, such as which words commonly appear together and what type of response is likely to fit a prompt. They are not a list of memorized answers. Instead, they function more like millions or billions of adjustable dials that guide the model’s responses. When a company releases an open-weight model, developers can download the settings, run the model on their own computers or servers, and modify it for specific use cases.

ChatGPT, which is not open-weight, provides a familiar comparison. OpenAI, the company behind ChatGPT, controls ChatGPT’s models and operates them on its own systems. OpenAI also offers separate open-weight models called gpt-oss, but those models are not available in ChatGPT. Examples of well-known open-weight AI models include Meta’s Llama, DeepSeek, and Qwen.

Open-weight AI is not necessarily the same as open-source AI. A fully open-source system generally provides the source code, model weights, and other information users need to study, modify, and redistribute the technology. An open-weight model makes the learned numerical settings available but may keep its training data, development process, or portions of its code private. In other words, users receive the trained model but not necessarily the complete recipe used to create it. In practice, openness exists on a spectrum. Models developed by companies such as DeepSeek and Alibaba have attracted users partly because their weights are available under relatively flexible terms.

Open Models Can Expand the Market

Open weights are especially valuable to businesses, software developers, researchers, and government agencies that want more control over how an AI model operates. Training an advanced model from the beginning requires powerful chips, large amounts of data, and specialized talent, putting it beyond the reach of many organizations. An open-weight model allows an organization to begin with an existing system, customize it, and run it on its own computers or cloud systems. This can lower costs, reduce dependence on the model’s original developer, and make it easier for new competitors to enter the market.

Most consumers do not need direct access to model weights. Downloading and operating an open-weight model can require suitable hardware and technical knowledge. However, consumers may benefit indirectly when open weights lead to more competing products, lower prices, stronger privacy options, and applications designed for particular languages, communities, or needs.

A business may not need a frontier model, one of the newest and most advanced systems available, for routine tasks such as classifying documents or analyzing internal records. A smaller open-weight model may be cheaper, faster, and easier to customize. Switching between models is becoming easier because developers are building software systems called harnesses. A harness connects an AI model to the tools, data, and applications it needs to complete a task. A harness can allow a company to replace one model with another as prices, performance, or business needs change. The underlying model remains important, but the complete product also depends on the software, data, tools, and controls surrounding it.

That shift increases competition. AI developers must compete not only on benchmark results, which are scores from tests used to compare model performance, but also on cost, reliability, security, and compatibility. Cloud providers compete to host the models. Chipmakers compete to supply the computing infrastructure. Software companies compete to build applications around them. Open-weight AI can therefore enlarge the overall market, even when it creates more rivals for individual model developers.

Why Nvidia Wants More Models

Nvidia’s support for open-weight AI reflects its position in the AI supply chain. The company sells the chips and computing systems used to train and operate many kinds of models. More models, applications, and AI users can create more demand for computing infrastructure.

A less expensive model does not necessarily reduce Nvidia’s opportunity. It could make AI affordable for more organizations, increasing the total number of models running in workplaces, data centers, and cloud platforms. As AI moves from experimentation to widespread business use, the market may focus less on which company owns the single best model and more on how many useful applications companies can build.

Chinese developer Moonshot AI illustrates this possibility. The company’s Kimi K3 model reportedly narrowed the performance gap with leading U.S. systems while offering an open-weight alternative. Its arrival renewed attention to models that developers can download, adapt, and substitute for more expensive proprietary products. Political concern followed, but the model also demonstrated how quickly new competitors can enter the upper levels of the market.

For Nvidia, the threat posed by a new model may be balanced by the computing demand it generates. More capable models can encourage developers to create more AI applications, which then require chips, servers, networking equipment, and cloud services.

Why Microsoft Benefits from Choice

Microsoft has invested heavily in proprietary AI, but it also operates a major cloud platform and sells software to organizations with widely different needs. That gives the company reasons to support both closed and open models. Microsoft can benefit from model variety because its cloud and software businesses can serve as the infrastructure through which customers access that variety.

Open weights can also reduce customer concerns about dependence on one supplier. A business may hesitate to build an important process around a model when pricing, access rules, or performance can change. The ability to move among models gives customers more bargaining power and flexibility. Microsoft can position its platform as the place where those choices are managed rather than insisting that every customer use one model.

This strategy resembles other technology markets in which a company benefits by supporting a broad ecosystem. The platform may become more valuable as the number of compatible products increases.

Why Meta Wants an Open Ecosystem

Meta has fewer incentives than some rivals to sell access to a closed model as its primary source of revenue. Its main businesses depend on advertising, digital services, and large online platforms. Releasing model weights can encourage outside developers to improve the technology, create complementary products, and make Meta’s approach more influential.

If many organizations build around Meta-supported models, those models can help establish industry practices and technical standards. Meta may gain access to a larger developer community, faster feedback, and improvements created outside the company. An open ecosystem can also weaken competitors whose business models rely on charging customers for access to proprietary models.

Security Is Both a Risk and a Selling Point

Opponents of open-weight AI worry that widely available models could be misused. Developers may remove safeguards, create harmful applications, or use stronger systems to support cyberattacks. U.S. policymakers are also debating the national security implications of Chinese open-weight models and the possibility that foreign developers used outputs from American models to train competing systems.

Supporters argue that openness can also strengthen cybersecurity. Independent researchers can inspect models, test their weaknesses, and create defenses without waiting for the original developer. Companies can run models on their own infrastructure, which may provide greater control over sensitive data.

Nvidia and more than 30 organizations have formed the Open Secure AI Alliance to develop and share technologies for securing AI software and agents. The group’s work includes model scanning, vulnerability research, secure coding processes, identity controls, and systems that make AI agents easier to test, trace, audit, and govern. 

The alliance shows that the debate is not simply a choice between openness and safety. Companies are trying to make security part of the open ecosystem. Their argument is that defenders need access to powerful tools because cybercriminals will continue improving their methods regardless of whether responsible organizations share defensive technology.

A Fight over the Structure of the AI Industry

The disagreement over open-weight AI is ultimately a disagreement over how the AI economy should be organized. Closed-model companies can maintain control over their technology, charge for access, update safeguards centrally, and prevent customers from seeing proprietary methods. Open-weight supporters favor a more distributed market in which businesses can operate, modify, and combine models.

Nvidia, Microsoft, and Meta do not have identical interests. Their support for open weights therefore combines public-policy arguments with competitive strategy. They can promote lower costs, wider access, innovation, and cybersecurity while also strengthening the parts of the AI market in which they hold an advantage.

The future will likely include both open and closed systems. Some organizations will pay for frontier models when they need maximum performance and centralized support. Others will choose open-weight models for customization, privacy, control, or lower costs. The most successful companies may not be those that force every customer into one approach. They may be the companies that profit no matter which model a customer selects.

In the Classroom

This article can be used to discuss the role of AI and technology in the economy (Chapter 1: The Dynamics of Business and Economics).

Discussion Questions

  1. How does an open-weight AI model differ from a fully open-source AI model?

  2. Why can open-weight models lower barriers to entry for startups and smaller organizations? 

  3. What security benefits and security risks are associated with open-weight AI?

This article was developed with the support of Kelsey Reddick for and under the direction of O.C. Ferrell, Linda Ferrell, and Geoff Hirt.