Open Weights, Closed Power. American AI Leadership

Downloadable AI models promise autonomy, competition and lower costs. They also shift power towards whoever controls compute, standards and deployment, turning openness into an instrument of industrial and geopolitical influence.

The Signal : A letter that changed before the ink dried

On 24 July, an industry letter appeared with NVIDIA, Microsoft, Meta, Mistral, and IBM among its supporters. OpenAI does not appear in the first version. It does appear in the copy now hosted by NVIDIA. That change captures an awkward moment in the AI market: even companies whose strongest systems remain behind controlled interfaces need a credible position on open models.

The letter presents open-weight AI as a foundation for American prosperity, competition and technological sovereignty. Yet “open” can describe very different forms of access, and every company supporting the document holds an interest in where the boundary gets drawn.

Open weights are not open source

A trained AI model contains billions of numerical parameters called weights. When a developer releases them, others can download the model, run it on their own infrastructure, fine-tune it and examine its behaviour without sending every request to the original provider. This gives organisations practical control over cost, latency, privacy and customisation.

Under the Open Source Initiative’s definition, a genuinely open-source AI system should also provide the information and code needed to study and reproduce its creation: architecture, training and inference code, data provenance, processing methods and relevant parameters. Many products marketed as open source disclose final weights while hiding much of the training data and development process.

The distinction matters in boardrooms. A bank can host an open-weight model internally and keep customer data under direct control. Its risk team may still struggle to explain the model’s training provenance, embedded biases or possible exposure to copyrighted material.

Downloadability improves operational sovereignty. It does not automatically create auditability.

America’s two-layer strategy

The United States increasingly treats open models as a strategic asset. Its policy direction supports their use by startups, researchers, public institutions and organisations handling sensitive information. American policymakers also recognise their international influence: widely adopted US models can shape technical standards and development practices far beyond the country’s borders.

At the same time, Washington continues to restrict access to the advanced chips required to train and operate the most capable systems. This creates a two-layer strategy. The model may travel freely, while the United States retains leverage over the computing infrastructure beneath it.

The weights move across borders. Much of the economic gravity stays close to the infrastructure provider.

NVIDIA, OpenAI and Meta define openness differently

NVIDIA approaches open models as deployable infrastructure. It supports models that organisations can download, customise and run across data centres, workstations and cloud environments. Wider model adoption also increases demand for the hardware and software needed to operate those systems efficiently.

OpenAI follows a tiered approach. It releases selected open-weight models for local use, while keeping its most capable frontier systems behind controlled services. Access expands at one level of capability and becomes more restricted at another.

Meta uses open-weight distribution to build a broad developer ecosystem. It encourages companies to modify, host and integrate its models across competing clouds and hardware platforms. Meta does not need to control every deployment directly; it gains influence when developers adopt its architecture, tools and conventions as a common foundation.

The three companies therefore place value in different parts of the market. NVIDIA benefits from the infrastructure required to run open models. OpenAI retains control over access to its strongest systems. Meta seeks scale through widespread adoption of its model ecosystem.

China exports an ecosystem, not only models

China has made open models part of its industrial strategy. Companies such as DeepSeek and Alibaba allow businesses and researchers around the world to download, adapt and run their models.

The influence of these models grows when organisations start building products around them. Developers learn how they work, companies invest in compatible tools, and teams create processes based on their strengths and limitations.

Over time, switching to another system becomes harder. The organisation may have trained staff, adapted software and built services around the Chinese model.

What began as a low-cost technical choice can gradually become a long-term dependency.

This gives China a way to extend its technological influence despite its disadvantage in advanced computing infrastructure. Its models can still shape how developers build applications, how companies organise their AI systems and which technical standards become widely adopted.

For the United States, this creates a difficult choice. Restrictions may address genuine concerns about security and intellectual property. Still if American technology becomes too expensive, difficult to access or heavily controlled, companies in emerging markets may turn more quickly to Chinese models that they can download and use at lower cost.

In trying to limit China’s influence, Washington could unintentionally help Chinese AI gain a stronger position in parts of the world where affordability and access matter most.

Europe’s sovereignty gap

Europe recognises the opportunity presented by open models, but it still faces a serious infrastructure constraint. European regulation gives some flexibility to freely available general-purpose models, while public programmes are expanding computing capacity and support for European AI development.

This approach fits Europe’s need for multilingual and sector-specific systems in manufacturing, energy, finance, healthcare and government. Even so, technological sovereignty remains incomplete when European applications may use American or Chinese models, run on hardware built around US-designed processors, and depend on non-European cloud providers.

Regulation can support trustworthy deployment. It cannot provide the computing capacity, patient capital and distribution networks required to build globally competitive AI companies.

Europe’s strongest opportunity lies in developing specialised systems between frontier research and large-scale commercial deployment. This requires treating models, datasets, computing infrastructure and public procurement as parts of the same industrial strategy. Funding isolated research projects will not create strategic autonomy if successful European companies must later rely on foreign infrastructure to scale.

What this means for SMEs

For small and medium-sized enterprises, open-weight models can offer lower costs, greater control over data and more flexibility than closed services. But they can also create new dependencies on cloud infrastructure, specialised skills, hardware and external integrators.

Before deployment, SMEs should ask which parts of the system they truly control, how easily they could change provider, who is responsible after fine-tuning, and whether they have the resources to monitor, secure and maintain the model over time. Open weights can expand choice, but only sound governance can turn that choice into real autonomy.

The governance problem moves downstream

Open-weight models give organisations more freedom to adapt AI to their own needs. That freedom also changes who must take responsibility when the system causes harm.

Consider an insurance company that modifies an open-weight model to help employees assess claims. The original developer created the foundation model, but the insurer chooses the training data, adds internal rules and decides how employees should use the output.

Several months later, the model begins rejecting certain claims more often than expected. The cause may lie in the original model, the company’s fine-tuning data, an internal software change or the way employees interpret its recommendations.

The organisation cannot simply point back to the model provider. By adapting and deploying the system, it has influenced how the model behaves and how its decisions affect customers.

This makes internal governance essential. By taking greater control over the system, the organisation also assumes a larger share of its accountability. Open weights increase control. They also move more responsibility into the organisation that chooses to use them.

The next frontier: governable openness

The coming contest will produce layers of access: downloadable weights, disclosed architectures, transparent data summaries, reproducible training pipelines, staged capability releases and trusted deployment environments.

Policymakers will need to address the risk created by a system and its use while recognising that responsibility changes across the model’s lifecycle. The original developer may control training. A distributor may alter the model. An enterprise may fine-tune it. A public authority may use its outputs to make consequential decisions.

A workable framework will require model registries, verifiable version histories, common evaluations, incident reporting and clearer duties for developers, distributors, fine-tuners and deployers. High-risk capabilities may justify conditional or staged releases. Lower-risk models should remain easy to adapt, especially where local deployment protects sensitive information and strengthens competition.

Leaders should therefore ask more than whether a model carries an open label. They should examine which parts of the system they can understand, move, govern and replace.

Openness will shape the next AI economy. It will not distribute power automatically. Countries and companies that release models can still control the chips, standards, platforms and expertise around them.

The organisations that benefit most will treat open weights as a transfer of capability and responsibility, then build the governance needed to carry both.