Nvidia, Microsoft, Meta, IBM, Hugging Face, and a long list of other tech heavyweights just sent Washington a blunt message: don't kneecap open-weight AI in the name of making AI safer.
The companies signed a new policy letter titled Open Weights and American AI Leadership, arguing that downloadable AI models are essential to competition, cybersecurity, innovation, and America's ability to remain a global AI leader.
The timing isn't accidental.
Washington is debating tighter controls on powerful AI models, lawmakers are floating mandatory kill switches, and the Trump administration is reportedly considering sanctions against Chinese AI companies over allegations that they improperly used American models to train their own.
Meanwhile, open models such as Kimi K3 are getting more capable, cheaper, and easier to deploy.
Now some of the biggest names in tech are warning that broad restrictions could do more damage than the technology they're meant to contain.
Big Tech's case for open-weight AI
The letter draws a direct line between today's open-weight AI movement and the rise of open-source software in the 1980s.
Back then, many companies believed software would only advance if its code remained locked behind corporate walls. Instead, open-source software became the foundation for much of the modern internet, cloud infrastructure, cybersecurity, scientific computing, and even critical government systems.
The signers believe AI is approaching the same fork in the road.
Their argument is that American AI leadership won't be determined by whether one U.S. company owns the most powerful model. It'll be determined by whether businesses, researchers, startups, schools, and public institutions can actually use advanced AI across the economy.
Open-weight models matter because organizations can download them, inspect them, modify them, and run them on their own hardware. That gives users more control over their data, costs, deployment, and long-term dependence on a single provider.
In plain English: you don't have to rent your entire AI future from one company.
The cost argument is getting harder to ignore
Closed frontier models are incredibly capable. They're also expensive when every task, customer interaction, background process, and agent action has to pass through a paid API.
The letter argues that companies should be able to match the right model to the right job.
A business might pay for a frontier model when it needs the strongest possible reasoning. However, it could run a smaller open model for routine classification, document processing, internal search, or repetitive agent work.
That's already how a lot of serious AI builders think. The future probably isn't one model doing everything. It's a stack of models, tools, and agents routed according to capability, privacy, speed, and price.
That model-agnostic approach is also why open systems such as Hermes Agent matter. Users aren't forced to bet the entire workflow on one provider's pricing, policies, or product roadmap.
Open models aren't automatically safe
The letter doesn't pretend open-weight AI is risk-free.
Once a model's weights are released, the original developer can't fully control what happens next. Modified versions can be difficult to trace, and bad actors can remove safeguards.
Those are real concerns.
But the companies argue that closed models aren't automatically safe either. Closed systems can still be breached, misused, manipulated, or quietly fail in ways outside researchers can't inspect.
We just saw an extreme example when an OpenAI model breached Hugging Face during a sanctioned security evaluation. According to Reuters, Hugging Face later said it relied on a Chinese open model during its response because closed models had restrictions that limited their usefulness for cybersecurity work.
That's the strongest point in the entire letter.
If attackers have access to powerful AI, defenders need models capable of meeting them on equal ground. Restricting legitimate researchers and security teams doesn't make the underlying threat disappear.
Washington is really arguing about control
The public debate is framed around safety, but control is sitting underneath nearly every part of it.
Who gets to build the strongest models?
Who gets to inspect them?
Who decides what they're allowed to do?
And who collects the toll every time someone uses one?
Closed AI concentrates those answers inside a small number of companies. Open-weight AI spreads them across developers, businesses, researchers, governments, and individual users.
That doesn't mean every model should be released without safeguards. It means policymakers need to separate specific illegal behavior from the broader act of publishing or using model weights.
The letter makes the same distinction around distillation, where one model's outputs help train or improve another. The signers argue that unlawful extraction should be handled through targeted legal and commercial action rather than treating an entire model-development technique as theft.
That's particularly relevant as Washington weighs action against Chinese model developers following the rapid rise of Kimi K3 and other low-cost systems.
Nvidia has plenty to gain from an open AI ecosystem
Let's not pretend every company signed this purely out of philosophical devotion to openness.
Nvidia sells the hardware that runs both closed and open models. A world where thousands of companies deploy models on their own infrastructure creates an enormous market for GPUs, servers, networking, and inference software.
Microsoft also benefits from selling cloud infrastructure and tools, even when customers aren't using a Microsoft-owned model. Meta has spent years releasing downloadable Llama models. Hugging Face exists at the center of the open AI ecosystem.
Their interests are obvious.
That doesn't make the argument wrong.
In fact, the strongest policy coalitions often form when public benefits and business incentives point in the same direction. Open-weight AI creates competition for model providers while creating opportunities across chips, clouds, applications, security, and agent platforms.
What the companies want Washington to do
The letter asks policymakers to support open AI instead of merely avoiding new restrictions.
That includes expanding access to computing resources for startups and researchers, investing in shared datasets and evaluation tools, and keeping the frontier competitive rather than allowing a handful of closed labs to become permanent gatekeepers.
It also calls for targeted rules tied to demonstrated harms instead of blanket limits based on the assumption that open models are inherently more dangerous.
That approach won't satisfy everyone. Some policymakers will argue that downloadable frontier models become impossible to contain once released. They're not entirely wrong.
Still, broad restrictions come with their own danger: they can lock smaller American companies out of the market, push research overseas, and leave the future of AI in the hands of a few corporations that nobody elected.
The bottom line
This is bigger than a fight over one Chinese model or one proposed regulation.
It's a fight over what the AI economy is going to look like.
One path leads to a small number of companies controlling the strongest models, the rules around them, and the price everyone pays to access them.
The other path is messier. It gives more people access, creates more competition, and comes with risks that can't simply be switched off from a central dashboard.
Nvidia, Microsoft, Meta, Hugging Face, IBM, and the rest of the coalition have made their choice clear.
They believe America won't win the AI race by putting the future behind a handful of API keys.
And honestly, they're right.
Sources: Open Weights and American AI Leadership, Reuters
