Recommended Reading
What This Article Covers
In the United States, the debate over how far open-weight AI models should be allowed is rapidly intensifying.
One of the direct triggers of this debate is Kimi K3, developed by the Chinese AI company Moonshot AI.
On July 22, Michael Kratsios, Director of the White House Office of Science and Technology Policy (OSTP), said that the U.S. government had obtained information indicating that Moonshot AI had used distillation of Anthropic’s Fable model in the development of Kimi K3. He also claimed that Moonshot AI had conducted large-scale distillation using U.S. AI models and had even built an internal system capable of switching among multiple access methods to avoid detection. Based on currently available information, these allegations have not been independently verified, so it is important to distinguish them as claims made by the U.S. government.
Kratsios also separately raised allegations that Moonshot AI had obtained NVIDIA GB300 servers subject to U.S. export controls on China and had accessed GB300 systems in Thailand. This article focuses on the distillation issue, which is more directly connected to the open-weight regulation debate.
As the rapid development of Chinese open-weight models has become intertwined with intellectual property concerns, the United States has begun discussing possible restrictions or sanctions on Chinese open models. AI Times reported that Chinese models, including Kimi K3, have been rapidly catching up with leading U.S. companies through open-weight strategies, prompting discussions within the U.S. government about possible restrictions or sanctions on Chinese open models.
In response, parts of the U.S. technology industry began pushing back against broad regulation of open-weight models.
On July 24, 25 companies and organizations issued a joint letter titled “Open Weights and American AI Leadership,” urging the U.S. government to avoid broad restrictions on open-weight AI. They argued that improper technology extraction should be addressed separately, but that normal distillation techniques or open-weight models themselves should not be broadly restricted because of such concerns.
The market response has also been rapid.
The number of companies and organizations signing the letter doubled from 25 to 50 in just two days. The currently published official list of signatories also includes 50 organizations.
As of 5:30 p.m. EDT on July 26, 2026, Anthropic has not joined the letter. The official list includes OpenAI and Google, but Anthropic is not listed.
Ultimately, two different perspectives are emerging in this debate.
One view emphasizes the need to protect AI technologies and competitive advantages developed through substantial investments of time and money. The other argues that technology must spread to more companies and research institutions in order for new businesses and services to emerge and for the overall AI market to grow.
To understand this difference, it is first necessary to understand what open-weight AI actually means.
First, Open-Weight AI Is Different From Open Source
Open-weight and open-source models are sometimes discussed as though they mean the same thing, but they are not exactly identical concepts.
Open source generally refers to a model in which software source code is made available so that users can examine and modify its internal structure.
By contrast, open-weight AI makes the weights produced through AI training available so that users can download the model, run it directly, or adjust it for their own purposes.
This does not necessarily mean that the training data, the full training process, or every piece of code used to build the model must also be disclosed.
In other words, an important characteristic of open-weight AI is that users can take the AI model itself and run it on their own infrastructure. The joint letter also describes open-weight models as models that can be downloaded, examined, modified, and run on an organization’s own infrastructure.
With closed models, users generally access AI through services or APIs provided by the developer.
With open-weight models, companies or institutions can instead download the model and operate it directly on their own servers or in a cloud environment of their choice.
Why Has the Open-Weight Regulation Debate Intensified Now?
The important point in this debate is not simply that high-performing AI models have emerged in China.
The question has become how those models were able to close the gap with leading models so quickly.
This is where distillation comes in.
Put simply, distillation is a method of using the outputs of an already high-performing AI model to teach another AI model.
If a high-performance AI model is asked to solve many problems and its outputs are then used to train another model, the new model can improve certain capabilities more quickly than if it had to build all of those capabilities from scratch.
However, distillation itself does not automatically amount to technology theft.
The joint letter also describes distillation as an AI development technique widely used for model improvement, evaluation, and validation. At the same time, it argues that if value is improperly extracted from a closed model, that conduct should be distinguished from normal distillation and addressed through legal or commercial measures.
Therefore, the issue currently being raised in the United States is not simply:
Should distillation be allowed or prohibited?
Rather, the question is closer to:
Where does legitimate technological learning end, and where does the improper extraction of capabilities created through another company’s substantial investment begin?
Why Do Different Groups in the Same AI Market View Regulation Differently?
This difference becomes easier to understand through an analogy involving people.
Imagine that a company spends a great deal of time and money developing one highly skilled expert.
That expert does not simply possess a large amount of knowledge.
Through extensive work experience, the expert has also developed ways of making judgments, sequences for solving problems, and accumulated practical know-how.
From the company’s perspective, that person’s capabilities are the result of long-term investment and an important source of competitive advantage.
Now imagine that another company asks this expert to solve a very large number of problems and then uses those results to train another person to a similar level in a short period of time.
The company that originally developed the expert may want to restrict such behavior.
That is because another company could rapidly catch up with a capability that required substantial time and investment to build, while spending far less itself.
A similar conflict of interest can arise with AI.
Developing a powerful AI model requires enormous computing resources, data, research personnel, and time.
If a competing model can rapidly acquire similar capabilities by making large-scale use of the outputs of an already developed AI, the party that made the original investment may see regulation as necessary to protect its technology and the value of that investment.
This is where a more regulation-friendly perspective emerges.
Others Place Greater Importance on Expanding the Market as a Whole
Now consider the same expert from a different perspective.
Suppose an expert who developed inside one company moves to another company or starts a new business.
That may be a loss for the original company.
But from the perspective of the overall industry, a new company may emerge, along with new products and services.
Other people may also develop new skills and businesses based on the knowledge and experience that the expert brings with them.
In other words, capabilities that were once concentrated within a single company can spread across an industry and expand the market itself.
A more regulation-skeptical view of open-weight AI is closer to this logic.
If more companies and research institutions can directly use AI models, even startups and universities that do not have the capital to develop models from scratch can make use of high-performance AI.
They can also adapt those models to specific industries and tasks and build new products and services on top of them.
The joint letter argues that open-weight AI can expand competition not only among model developers but also across cloud computing, semiconductors, applications, and services.
The difference between the two sides is therefore not simply whether they like openness or dislike it.
The more regulation-friendly perspective places greater importance on protecting the technology and competitive advantages created through early investment.
The more regulation-skeptical perspective places greater importance on the possibility that technology will spread across companies, create new businesses and services, and expand the overall market.
Of course, actual companies do not fall perfectly into two separate camps.
Their positions can differ depending on how powerful a model is, what types of distillation should be allowed, and where the line between legitimate use and technology extraction should be drawn.
Ultimately, the Question Is Whose Servers AI Models Will Run On
The open-weight debate is also connected to the question of who will operate AI.
Today, many generative AI models are operated directly by the companies that developed them.
Users do not own the AI models themselves. Instead, they access services or call APIs.
In simplified form, the structure looks like this:
AI developer’s model → developer’s servers → service/API → user
If open-weight AI expands, another structure becomes possible.
AI model → downloaded by a company → company’s own servers or cloud → direct operation
In this model, companies are not limited to relying on the services of a particular AI provider. They can select models and operating environments based on their business needs, costs, and data-management requirements.
The joint letter also emphasizes that one advantage of open-weight AI is the ability to reduce dependence on individual vendors and allow organizations to exercise more direct control over their data and accumulated AI capabilities.
If More Companies Operate AI Directly, the Direction of Money in the Market May Also Change
This difference does not end with a technical choice.
It can also affect where money flows within the AI industry.
When AI developers operate models themselves, growing usage primarily expands the market for APIs and subscription services.
However, if more companies begin operating AI models directly, a broader set of industries becomes necessary.
Companies need semiconductors and servers to run AI models, as well as data centers and cloud infrastructure to operate them.
They also need platforms and enterprise software to connect AI with their existing systems.
In other words, the AI market could expand from a structure centered on:
Models → APIs and services
to a broader structure involving:
Semiconductors → servers → cloud → data centers → AI platforms → enterprise AI services
The joint letter likewise argues that when more organizations can directly build, modify, and deploy AI, competition can expand beyond model development into cloud computing, chips, applications, and services.
From this perspective, the increase in signatories from 25 to 50 in a short period of time does not necessarily have to be interpreted only as agreement with a particular philosophy of open AI.
Different companies may also see different business opportunities in the expansion of the open-weight ecosystem.
That Does Not Mean Open-Weight AI Is Automatically Safe
Those opposing broad open-weight regulation do not deny the risks.
Once model weights are released, the original developer may have difficulty taking them back.
If someone copies or modifies a model for another purpose, the original developer may also be unable to track and control every use of it.
The joint letter itself acknowledges the risk that once weights are released, they leave the original developer’s control and that modified models can be difficult to track or reverse.
At the same time, closed models are not automatically safe either.
If powerful AI capabilities become concentrated in a small number of companies, security incidents or system failures may also concentrate risk within a small number of providers.
Supporters of open-weight AI also argue that researchers and developers need to be able to examine models in order to identify vulnerabilities and develop safeguards.
Ultimately, the practical debate is less about choosing between complete openness and complete control than about determining what level of model should be released, under what conditions, and which types of conduct should be restricted.
DANA NOTES Commentary
The current open-weight debate is not simply about deciding whether open-weight AI is good or closed AI is good.
Companies that develop AI have legitimate reasons to protect capabilities created through enormous investments of time and money.
At the same time, if those capabilities spread to more companies and research institutions, new AI companies and services may emerge, allowing the overall market to grow more quickly.
So the important question does not end with whether technology should be protected or allowed to spread.
The more important question is:
Where should the boundary be drawn between legitimate technological diffusion and improper technology extraction?
At the same time, if open-weight AI expands, competition in the AI industry may no longer be determined by model performance alone.
Competition will also involve who can operate the model, what infrastructure it runs on, and what new services and industries can be built on top of it.
The current U.S. debate over open-weight regulation therefore goes beyond the question of whether AI technology should be open or closed. It is becoming an industry-structure debate over how to protect the value of technologies developed first without preventing the broader AI market from continuing to grow.
