Open AI: Nvidia and Mistral press Washington

Washington faces the open-model dilemma

The rules decided in Washington on artificial intelligence could have effects far beyond the US market. According to TechCrunch, several companies in the sector, including Nvidia and Mistral, are asking US authorities not to adopt broad restrictions targeting AI models with “open weights” (open weight). This position comes as the United States assesses its response to the rise of Chinese models, notably those from Moonshot AI.

The debate is not simply a clash between innovation and regulation. It concerns a more specific technical and geopolitical question: how far can public authorities control the distribution of an AI model’s parameters without, in the process, weakening the companies, researchers and ecosystems they seek to protect? A model’s weights are the numerical values learned during its training. When they are published, a third party can download the model, run it on its own infrastructure, adapt it or specialize it, subject to the license that accompanies it.

This availability does not necessarily mean that the entire development process is transparent. An open-weight model may be distributed without its training code, data, optimization recipes, infrastructure or internal evaluations themselves being accessible. The distinction matters: in public debate, the terms “open source,” “open” and “open weight” are frequently used as synonyms, even though they refer to different degrees of openness. The US discussion concerns precisely this circulation of weights, because it makes redistribution and use difficult to govern once the files are published.

The growing visibility of Chinese models has lent new intensity to this debate. US authorities and players must balance several concerns: preventing dangerous uses, maintaining an industrial lead, avoiding a situation in which national regulation shifts global users toward foreign solutions, and preserving the attractiveness of the United States as a place for research and development. In this equation, models published by Chinese companies are at once a matter of competition, national security and technological dissemination.

For Mistral, this US sequence has particular significance. The French company has established itself as one of Europe’s leading players in generative AI by combining commercial offerings with the publication of certain open-weight models. Its participation in a pushback against broad restrictions gives the issue a directly European resonance. It is no longer merely a face-off between Washington, major US platforms and Chinese labs: US rules can alter the competitive conditions for a French company whose strategy partly relies on customers and developers being able to access models that can run in their own environments.

What Nvidia and Mistral are asking for

According to the TechCrunch article, the companies participating in the debate are urging the United States not to impose broad restrictions on open-weight models. The central point is the indiscriminate nature of a potential ban or broad control. The signatories do not necessarily challenge the principle that certain models or uses may call for security measures. Their concern is with an approach that would treat all open models as a homogeneous risk category.

This distinction is crucial from a regulatory standpoint. A model’s capabilities vary greatly depending on its size, training data, level of reasoning, added safeguards, ability to use tools, access to external systems and the way it is deployed. A rule based solely on the fact that weights are published could, according to its opponents, disregard these differences. Conversely, a policy focused on capabilities or uses would seek to target identified risks, but it would raise another difficulty: how can a technical system whose performance changes rapidly be assessed reliably and sustainably?

Nvidia is not a peripheral player in this debate. The US group is at the heart of the computing infrastructure used to train and run a large share of contemporary models. Its position on model openness must therefore be read in a context where hardware, software, deployment services and developer ecosystems are closely linked. The distribution of open-weight models fuels demand for computing for fine-tuning, inference and the integration of AI into specialized products or services. It also helps make the ecosystem of US chips and software tools a global benchmark.

For Mistral, the stakes are different but complementary. Founded in Paris in 2023, the company has grown in a landscape dominated by US companies with considerable financial and computing resources. Publishing weights for certain models is, in this context, a way to build a developer community, reach companies that do not want to send their data to an external API, and offer an alternative to models accessible exclusively as proprietary services. The ability to distribute and run models locally is particularly valued by organizations subject to confidentiality, sovereignty or compliance constraints.

Mistral’s position does not mean that open models are risk-free. Companies that publish them know that independent reuse reduces their direct control over subsequent deployments. But their argument, as it emerges from the controversy reported by TechCrunch, is that an overly broad response to Chinese competition could produce a paradoxical result: restricting Western providers without preventing foreign models from circulating globally.

This objection lies at the heart of competitive logic. A model whose weights are already widely distributed can be hosted, copied, adapted and redistributed across multiple jurisdictions. National controls retain significant influence over companies established in the country, supply chains and exports of controlled technologies. However, they become more difficult to apply when the targeted resource is a digital file available from players outside the relevant jurisdiction. A US policy that limited domestic publications would therefore have to demonstrate that it effectively reduces risk rather than transferring the distribution advantage to foreign competitors.

The controversy is less about the idea of security than about the instrument chosen: should the publication of weights as such be regulated, or the capabilities and uses that create a concrete risk?

The requests made to Washington come in a climate where AI decisions are increasingly tied to strategic competition with China. US controls on certain exports of advanced chips and manufacturing equipment have already shown that computing infrastructure is regarded as a lever of power. Open-weight models add a different layer to the problem: they concern software and the spread of capabilities, not merely access to the hardware needed to train the most ambitious systems.

Moonshot AI, Chinese competition and the precedent of technology restrictions

The reference to Moonshot AI in the case followed by TechCrunch illustrates the increased attention paid to Chinese companies capable of producing competitive models. Moonshot AI is known for its Kimi family of models, and its name appears in a context in which the performance, distribution methods and costs of using Chinese models are being scrutinized far more intensively in Washington than before.

This vigilance is explained by a shift in perception. For several years, large generative models were associated, in Western debate, with a handful of US labs and their technology partners. OpenAI, Google, Anthropic, Meta and Microsoft have structured a large part of the discussions around capabilities, computing costs and risks. But the proliferation of models developed in China has made the idea that US regulation alone would determine AI’s global trajectory less credible.

The difference between distribution strategies reinforces this observation. OpenAI primarily provides its most advanced models through controlled interfaces and services. Anthropic likewise follows a logic of models accessible via APIs and cloud partnerships, without publishing the weights of its leading models. Google combines closed products with families of models whose weights are partly available, notably Gemma. Meta, for its part, has largely helped establish open-weight models in the industrial debate with Llama, while subjecting their use to a specific license rather than a traditional open-source license.

The closed commercial model has several obvious advantages for labs: access control, the ability to monitor certain abuses, centralized updates, monetization of requests and protection of part of the technological advantage. In theory, it is also easier to suspend a customer, modify usage rules or deploy new filters. But it concentrates dependence on a few providers and their cloud infrastructure. For companies, government bodies or research institutions, that dependence can become a strategic, financial or legal issue.

Open-weight models offer a different trade-off. They allow an organization to choose its host, keep data in a controlled environment, fine-tune the model for a particular task and avoid complete lock-in to a single interface. In return, the original provider loses some of its ability to oversee subsequent uses. States therefore cannot examine the question of openness solely from a market perspective. They must take into account the type of control they wish to preserve and the risks they consider priorities.

The precedent of semiconductor restrictions sheds light on the situation without providing an automatic answer. Leading-edge chips are physical, concentrated in complex industrial chains and subject to export procedures. A model’s weights are digital, duplicable and likely to circulate at low cost once available. The means of regulation are therefore not the same. It is possible to control certain hardware transactions by relying on manufacturers, distributors, customers and customs authorities. It is far more difficult to permanently prevent the spread of a file that can be copied onto servers, peer-to-peer networks or private media.

This observation does not imply that all regulation of open models would be ineffective. Authorities can act before publication, impose testing obligations, make certain funding conditional, oversee associated services or define responsibilities for companies distributing systems with particular capabilities. But each mechanism must balance practical effectiveness, international compatibility and the cost to innovation. It is precisely on this ground that companies such as Nvidia and Mistral seek to exert influence: to prevent a response built amid geopolitical urgency from becoming a structural ban on openness.

Chinese competition also introduces a definitional difficulty. A model may be developed in one jurisdiction, trained with resources spread across several countries, offered through a foreign platform, and then adapted by an international community. A model’s nationality is not always as straightforward as that of its originating company. A policy that distinguished models according to their provenance would therefore have to specify the criteria it uses: place of incorporation, ownership control, location of computing, origin of data, publication of weights or access to users. These questions, often technical, nevertheless determine the real scope of a measure.

A battle over security, but also over the global spread of standards

Supporters of restrictions on open weights generally put forward a precautionary argument. If a highly capable model is distributed without access controls, it can be studied, modified or exploited by malicious actors. Safeguards put in place in a hosted version can be removed. Providers can no longer see requests or intervene afterward. This situation is very different from a cloud service where the model remains on its creator’s servers.

The reasoning is particularly sensitive for systems that could, based on their capabilities, facilitate cyberattacks, the creation of dangerous content or the automation of high-risk tasks. National security concerns therefore cannot be dismissed in the name of openness. On the contrary, they are why US authorities are examining the consequences of the rise of foreign models and the conditions for distributing domestic models.

But opponents of a broad restriction point to a problem of proportionality. Not all models have the same level of capability, and openness does not automatically produce the same risk in every case. A small model specialized in text summarization, a multimodal model intended for research, a code-generation system and an agentic model able to call tools do not have the same profile. Regulation that made no distinction could discourage publications whose scientific or economic benefits are substantial, while leaving intact the availability of models published outside the United States.

The question of measurement is central. To establish a threshold, regulators must choose indicators: the amount of computing used during training, number of parameters, test results, access to tools, performance on cybersecurity tasks or operational autonomy. None of these criteria is sufficient on its own. The number of parameters, for example, does not fully describe a model’s capabilities. Tests are useful, but their results depend on protocols and can quickly become obsolete. Training compute is an indicator more readily tied to infrastructure, but it does not directly measure a model’s risk after fine-tuning or adaptation.

The regulatory choice also has an economic dimension. Publishing weights is not merely an ideological gesture in favor of transparency. It is a distribution method. It enables integrators to build products, researchers to reproduce or assess certain work, and companies to create vertical solutions. In sectors where data are sensitive — healthcare, finance, industry, defense, public services — the ability to deploy AI in a controlled environment may be a more important purchasing criterion than the raw performance of a general-purpose model.

The US debate therefore concerns the ability of Western providers to establish de facto standards. When a model is widely adopted by developers, libraries, fine-tuning tools, evaluation systems and professional skills tend to develop around it. This creates network effects. If US and European companies are heavily restricted in distributing their weights while Chinese models remain accessible, technical communities may migrate to the available alternatives. The loss would not be merely commercial: it would affect methods, platforms and influence over security practices.

Nvidia has a particular interest in this balance. Open models can run on a variety of infrastructures, but their development and industrialization also rely on software and hardware stacks. The global spread of AI workloads supports demand for accelerators, libraries and computing environments. Mistral, for its part, has an interest in ensuring that the local deployment option remains viable and legitimate in international markets. The two companies can therefore converge against a broad restriction while holding very different positions in the value chain.

It would nevertheless be reductive to portray this convergence as a simple defense of the absence of rules. The discussion may lead to more targeted requirements: documentation, safety evaluations, disclosure of known limitations, reporting procedures or enhanced obligations for specified capabilities. The challenge is to prevent these requirements from becoming, in practice, a cost impossible for smaller players, academic labs or European companies to bear. Regulation that claims to be neutral can favor the best-funded companies if it requires procedures that only a few platforms are able to implement.

Why Europe and France are directly concerned

Mistral’s presence in the US debate is a reminder that Europe cannot regard Washington’s rules as an external issue. Even when a measure formally targets only US territory, it can influence investors, industrial partners, cloud providers, multinational clients and compliance practices. AI is a cross-border market: models circulate, companies serve clients in multiple jurisdictions and infrastructure is distributed across different countries.

In France, the question of technological autonomy has become structuring since the acceleration of generative AI. The ambition is not merely to have conversational interfaces in French. It concerns the ability to train, adapt, host and audit models suited to European languages, sectors and legal frameworks. Open-weight models can contribute to this ambition because they give organizations greater control over where they run, how data are processed and how the system is fine-tuned.

This option is particularly important for companies that cannot or do not want to depend exclusively on an external API. A bank, industrial company, hospital, public administration or provider of critical services may need to integrate the model into an internal architecture. This does not eliminate security, confidentiality or governance obligations. It does, however, allow the hosting conditions to be chosen and controls specific to the organization to be operated.

The European framework does not treat open AI as an entirely separate subject. The European regulation on artificial intelligence, commonly known as the AI Act, includes provisions concerning general-purpose AI models. It also provides specific treatment for certain components published under free and open-source licenses, with limits, notably when systemic risks are involved. The existence of this architecture shows that Europe is already trying to distinguish openness, transparency obligations and the level of risk rather than reducing the debate to an alternative between unconstrained publication and complete closure.

Mistral’s position makes a European tension visible: the continent wants to develop its own capabilities and avoid excessive dependence on foreign platforms, while applying a demanding regulatory framework. If the United States were to broadly limit the publication of weights, some European companies could theoretically benefit from greater competitive space. But this reasoning would be incomplete. They also depend on global collaborations, US components, international developer communities and clients who expect compatibility with prevailing market practices.

A fragmentation of rules could also increase costs. A company that publishes or distributes a model in several regions would have to determine whether it can offer the same weights, the same license, the same documentation and the same features everywhere. Legal, safety and compliance teams would take on a larger role in the publication decision. For large companies, these costs can be absorbed. For younger players, they can directly influence the choice between an open model, a closed service or abandoning a product line.

The issue also has a linguistic dimension. The quality of models in languages other than English depends on data, evaluations, tools and communities able to adapt them. Accessible models can facilitate local work on French, regional languages and the languages of French-speaking partners, provided that rights, data and uses are handled properly. Conversely, complete market concentration around a small number of global APIs can reduce the room for maneuver of players seeking to build solutions tailored to specific contexts.

For French and European decision-makers, the challenge is therefore twofold. They must maintain credible requirements regarding risks and fundamental rights, while avoiding confusing sovereignty with closure. Sovereignty can also involve the ability to examine a model, host it, modify it within an authorized framework and develop local skills. Open weights do not by themselves guarantee this autonomy: training the most powerful models requires data, skills and computing. But they can be one of the tools for strengthening it.

Toward regulation of capabilities rather than a blanket ban

The upcoming US decision will be watched far beyond the companies cited by TechCrunch. It will help define an international doctrine on the following question: should the publication of weights be presumed dangerous for the most advanced models, or can it remain possible under graduated conditions? The answer will influence investments, research strategies, licenses and the way labs choose to commercialize their systems.

A broad ban would have the apparent advantage of simplicity. It would send a strong political signal and reduce uncertainty about what is permitted. But its simplicity could be misleading. It would have to define the models concerned, publication arrangements, scientific exceptions, rules applicable to versions already distributed and the treatment of foreign models. Without these details, it could create uncertainty rather than reduce it.

A capability-based approach would be more nuanced, but more demanding. It would require robust evaluations, revisable thresholds and sufficient public expertise to keep pace with the technology. It would also have to distinguish theoretical risk from the risk actually enabled by a model. A system that performs very well on a language task does not automatically have the same implications as a model connected to tools, trained to act in real-world environments or optimized for specialized operations.

The pressure exerted by Nvidia, Mistral and other companies is precisely aimed at preserving this possibility for differentiation. Their message is strategic: in a market where Chinese models are accessible, broadly closing off Western alternatives may weaken the position of the United States and its partners without ensuring a proportionate security gain. The debate is therefore not only about the right to publish. It concerns democracies’ ability to organize a response that is technically applicable, economically sustainable and credible in the face of global competition.

For Mistral and the French ecosystem, the outcome will be a test of the place Europe can occupy between major closed labs and the rapid spread of models from China. If Washington favors a restrictive line, European companies will have to assess the consequences for their partnerships, clients and distribution choices. If the United States opts for targeted rules, the market for open-weight models could retain its role as a concrete alternative to proprietary APIs, with increased competition on quality, safety, cost and deployment conditions.

The most likely prospect is neither the immediate disappearance of open AI nor its complete lack of oversight. It is a shift in the debate toward publication arrangements, capabilities regarded as sensitive and the responsibility of organizations that develop or distribute models. In this phase, Mistral’s voice matters because it is a reminder that a US decision on open weights can reshape a global market. The regulation that emerges in Washington will help determine whether open models remain a space of technological plurality, or become a field dominated by players able to evade the constraints imposed on others.