Why Silicon Valley is Rallying Behind Open-Weight AI
Nadella, Pichai, Huang and Altman have joined a push against broad US restrictions as increasingly capable Chinese models gain users.
A push to shield open-weight artificial intelligence from broad US restrictions has drawn support from many of the technology industry’s most influential executives, as Washington weighs how to respond to the growing reach of Chinese models.
The statement, titled Open Weights and American AI Leadership, was published on Microsoft’s website on Friday, 24 July, with 25 signatories. They included Microsoft, Nvidia, Meta, IBM, Palantir, Dell Technologies, Hugging Face, Mistral, Mozilla and the Linux Foundation.
By Monday, 27 July, the number of organizations listed on the website had risen to 77. Google and OpenAI, both absent from the original document, had joined, along with AMD, Cisco, Cloudflare, Cohere, GitHub, Scale AI and SpaceX. Anthropic, Amazon and Elon Musk’s xAI were not among the listed signatories.
The additions turned what began as a coalition led by chipmakers, enterprise technology companies and open-model developers into a broader industry show of support.
Open-weight models make their trained parameters available for download, allowing users to inspect and run them on their own infrastructure. The extent to which they may be modified or redistributed depends on the license. Once seen mainly as an option for researchers and smaller developers, they are now being promoted as technology that companies and governments should be able to operate and adapt themselves.
Nvidia chief executive Jensen Huang used his first post on X to back the statement.
“The world needs both frontier closed models and frontier open models,” Huang wrote.
Microsoft chief executive Satya Nadella described open-weight models as “essential to a healthy AI ecosystem”.
Google chief Sundar Pichai said he was “very happy to support this on behalf of Google”, pointing to the company’s open-source contributions and its release of Gemma open-weight models.
Sam Altman also backed the initiative. OpenAI’s leading commercial models remain proprietary, although the company also offers its gpt-oss family of open-weight models.
“I want the US to win in AI both in open source and proprietary models,” Altman wrote. OpenAI was later added to the signatory list.
Musk wrote that the statement had his “full support” and that “Jensen is right”. SpaceX subsequently appeared among the signatories, though xAI did not.
The China Question
China is not mentioned in the document. Even so, the statement comes amid an intensifying debate in Washington over Chinese AI models and the terms on which they should be allowed into the US market.
Chinese developers including Moonshot AI, Alibaba, DeepSeek and Z.ai have produced models that can be downloaded and adapted. Their combination of improving performance and lower prices has attracted developers and companies looking for alternatives to paid services from US laboratories.
Moonshot had said it planned to release the full weights for its Kimi K3 model by 27 July.
Days before the industry statement appeared, White House science adviser Michael Kratsios accused Moonshot of using Anthropic’s Fable model to help develop Kimi K3.
Treasury secretary Scott Bessent said sanctions and trade restrictions could be considered. Moonshot had not publicly answered the allegation, which some researchers questioned.
US officials and American AI companies have focused much of their criticism on distillation, in which the outputs of one model are used to train or improve another.
Distillation itself is a standard development technique. Anthropic argues that it becomes an attack when companies use false accounts, proxy services or other methods to collect responses at scale in breach of access restrictions and commercial terms. The company said in February that three China-based laboratories had generated more than 16 million exchanges with Claude through about 24,000 fraudulent accounts.
The industry statement draws a line between the technique and the way it is used. It calls distillation a widely employed method for model development, evaluation and validation and urges policymakers not to treat every instance as misappropriation.
Unlawful extraction should instead be addressed through “targeted legal and commercial frameworks”, it says, rather than restrictions that could also curb legitimate research and development.
That distinction is central to the coalition’s argument. Companies that violate contracts or intellectual-property laws could still face penalties without Washington restricting downloadable models as a class.
The Businesses Behind the Argument
The executives supporting open weights have framed their case around competition, security and US technological leadership. The commercial logic is never far away.
Nvidia gains when more models are trained and deployed. Whether a system comes from Meta, Microsoft, OpenAI or a Chinese laboratory matters less to the chipmaker than how much computing power it consumes.
Huang put that case plainly in an interview with Axios.
“Free AI should be great for hardware,” he said. “Free AI should be great for chips. Free AI should be great for data centers.”
Microsoft and Google can also support both open and proprietary systems. Each develops its own models while running a cloud platform that hosts models from numerous outside providers.
Microsoft Foundry offers systems from OpenAI, Anthropic, Meta, Mistral, DeepSeek and others. Google’s Model Garden similarly includes Google models, partner systems and open-weight models. Greater model choice gives customers more reasons to buy computing capacity from Azure or Google Cloud, whichever model they select.
Nadella has separately warned that companies may lose control of the institutional learning created through prompts, corrections, evaluations and workflows when they depend on outside AI providers.
He calls this the “reverse information paradox”, in which customers pay for a service and then supply the proprietary information needed to make it useful.
Running an open-weight model on a company’s own infrastructure can provide greater control over its data, customization and accumulated knowledge. That option is particularly relevant to banks, governments, defense organizations and other users handling sensitive information.
Meta has a different calculation. Mark Zuckerberg has long argued that companies should be able to “control our own destiny and not get locked into a closed vendor.” He has also said Meta wants Llama to become an industry standard.
Wider adoption of Llama gives Meta influence over the model ecosystem without requiring the company to charge users for every query.
For Dell, IBM, Palantir, ServiceNow and other enterprise suppliers, much of the opportunity begins after a model is released. Customers still need hardware, cloud capacity, security, software and consulting to install, govern and adapt it.
OpenAI is keeping a foothold in both markets. It continues to protect its most advanced commercial models while offering downloadable gpt-oss systems. Altman can therefore support an industry containing open and proprietary models without committing OpenAI to release the weights of its flagship proprietary systems.
Open Weight Is Not Open Source
An open-weight model makes its trained numerical parameters available for download, use and modification. It does not necessarily disclose the training data, source code, data-processing methods or full development recipe.
The Open Source Initiative says weights alone are not enough for an AI system to qualify as open source. Its definition also requires the information and code needed to study how the system was created and to modify it meaningfully.
The industry statement is therefore asking for something narrower than complete transparency. It wants companies to retain the freedom to distribute trained models even when much of the development process remains private.
That freedom carries risks. Once weights have been released, their creator cannot reliably recall them, trace every copy or restore safeguards removed by users.
The statement acknowledges this problem but argues that closed models are not inherently safe. They can also be breached, misused or fail in ways that outsiders cannot identify. Open models, supporters say, allow a wider group of researchers and security specialists to examine weaknesses and develop defenses.
A downloadable system can nevertheless be stripped of restrictions and distributed far beyond its creator’s reach. Wider scrutiny must therefore be weighed against the loss of control that follows the release of model weights.
The endorsements do not amount to a demand that every frontier model be released. Huang explicitly called for both open and closed models, while Google and OpenAI continue to keep their flagship systems proprietary even as they support separate open-weight families.
The policy debate is also a fight over the structure of the AI market. Closed-model laboratories make money by controlling access. Nvidia sells the computing power whichever model wins. Microsoft and Google can host both. Meta benefits if Llama becomes a widely used standard.
The coalition’s warning to Washington is that broad restrictions could protect a few US model providers while pushing developers toward open systems produced elsewhere. Its members agree on the need for open weights, even if they expect to profit from very different parts of the market.
