Nvidia and Microsoft Ask the US Not to Restrict Open AI Models

⏱️ Lectura: 11 min

Nvidia and Microsoft, two companies competing for control of AI infrastructure, agreed this week on the same request to Washington: don’t impose broad restrictions on open AI models. The request, reported by Fox Business, comes just as Chinese models like Qwen, DeepSeek, and Kimi increasingly dominate open AI adoption outside China.

📑 En este artículo

1. TL;DR
2. What happened
3. Context and history
4. What open models are and why they worry Washington
5. How to check if a model affects you
6. Impact and analysis
7. What’s next
8. Frequently Asked Questions
   1. What does it mean for an AI model to have open weights?
   2. Why would the US government consider restricting its publication?
   3. What was the AI Diffusion Rule?
   4. Why do Nvidia and Microsoft, who compete with each other, agree on this?
   5. Would Chinese open models like Qwen or DeepSeek be subject to the same rules?
   6. How does this affect me if I only use open models to develop software?
9. References

The joint stance is unusual: Nvidia sells the chips that train and run these models, while Microsoft operates Azure and is OpenAI’s largest investor, whose business partly depends on keeping its own frontier models closed. Both agree that broadly restricting open weights would hurt the US more than its competitors.

TL;DR

• Nvidia and Microsoft asked Washington to avoid broad restrictions on open AI models, according to Fox Business.
• The debate reopens whether publishing a model’s weights amounts to a controlled export under the EAR regime.
• The January 2025 AI Diffusion Rule was rescinded in May 2025 after industry criticism.
• Chinese open models like Qwen, DeepSeek, and Kimi already compete on equal footing and aren’t under US jurisdiction.
• Nvidia already faced chip export restrictions to China during 2025.
• Microsoft operates Azure AI Foundry, which hosts third-party open models alongside OpenAI’s own.
• RAND and CSET have documented since 2023 the risk of securing frontier model weights once released.

What happened

The Fox Business report, republished by MSN, indicates that Nvidia and Microsoft conveyed to US administration officials their concern about a possible expansion of export controls to open AI models. Until now, the EAR (Export Administration Regulations) control regime had focused on hardware: training GPUs, accelerators, and semiconductor manufacturing equipment. The current discussion, according to the report, points to something different: whether publishing the weights of a large-scale AI model should also be treated as an export of sensitive technology.

For Nvidia, the logic is straightforward: the more restrictions placed on the open ecosystem, the less GPU demand there is to experiment with, fine-tune, and deploy models outside a handful of closed labs. For Microsoft, the calculation is more nuanced: the company operates Azure AI Foundry, a catalog that hosts both its own models and third-party open models, and much of its enterprise customer base combines open models with OpenAI’s closed models. A broad rule treating the release of weights as a controlled export would require licenses, destination reviews, and potential penalties for noncompliance, exactly the kind of friction Microsoft and Nvidia already faced when they opposed the January 2025 AI Diffusion Rule.

The newsworthy detail isn’t the stance itself (the industry has spent years asking for more surgical rules), but that two companies with such different incentives publicly agree on the same argument: broadly restricting open models doesn’t slow China down, it just cedes ground.

Context and history

This debate didn’t come out of nowhere. In January 2025, in its final days, the Biden administration published the AI Diffusion Rule, which divided the world into three tiers of access to AI chips: allies with nearly unrestricted access, a middle group with quotas, and a third group, including China, with heavily restricted access. The industry, led by Nvidia, called the rule excessively broad: it penalized allied countries and legitimate commercial customers alongside the actors that actually worried national security.

The following administration rescinded that rule in May 2025, promising a more bilateral, country-by-country approach. But the underlying problem (how to control the spread of AI capabilities without slowing domestic innovation) didn’t disappear: it just shifted. First it was a discussion about chips. Now, with Chinese-origin open models like Qwen, DeepSeek, and Kimi gaining global adoption, the question has moved to the model weights themselves.

Researchers at institutions like RAND and CSET (Center for Security and Emerging Technology) have documented since 2023 the so-called problem of “securing model weights”: once a weights file is published, replicating it across thousands of servers takes hours, not months. Unlike a physical GPU, which can be tracked at customs, a file of several hundred gigabytes can be uploaded to a mirror in another country with a single command line. That asymmetry is the technical core of what Nvidia and Microsoft are now raising.

What open models are and why they worry Washington

Before continuing, two terms that are often used as synonyms, but aren’t, need to be separated. An open-source model publishes weights, training code, and sometimes the data. An open-weight model only publishes the file with the trained parameters, under a license that can be permissive (Apache 2.0, MIT) or restrictive (like Llama’s community license, which imposes usage conditions). Most of the frontier models discussed in this debate are, strictly speaking, open-weight, not fully open-source.

The following table summarizes where each relevant model family stands regarding a possible expansion of export controls:

Model	Origin	License	Open weights	Exposed to an expanded EAR control
Llama (Meta)	United States	Custom Community License	Yes	Yes, due to its US origin
OLMo (Allen Institute)	United States	Apache 2.0	Yes	Yes
Qwen (Alibaba)	China	Apache 2.0	Yes	No, outside US jurisdiction
DeepSeek	China	MIT	Yes	No
Kimi K2 (Moonshot AI)	China	Modified MIT-type license	Yes	No

The asymmetry is obvious: a rule requiring an export license to publish US-origin weights wouldn’t reach Qwen, DeepSeek, or Kimi, which are published by Chinese companies outside US jurisdiction. This is exactly the argument that, according to the report, Nvidia and Microsoft brought to Washington: a broad restriction only holds back US labs, while Chinese models keep circulating freely.

The following diagram summarizes, in simplified form, how a model would be evaluated today under the EAR regime if it were extended to open weights:

flowchart TD
A["Trained model"] --> B{"Are the weights published on Hugging Face or GitHub?"}
B -->|"Yes"| C{"Is the originating company under US jurisdiction?"}
B -->|"No, API access only"| D["Treated as a service, not a technology export"]
C -->|"Yes"| E["Would require a BIS export license"]
C -->|"No, origin outside the US"| F["Outside the scope of the current rule"]

💭 Key point: Nvidia and Microsoft’s central argument is that an open-weight model, once published on Hugging Face or GitHub, gets replicated within hours across mirrors and forks. No export control can “un-invent” a file that already circulates in thousands of copies.

How to check if a model affects you

For a developer, the regulatory debate is abstract until it intersects with a model you’re already using. Before redistributing fine-tuned weights of a model, you should programmatically check its license and origin.

A quick first check uses the public Hugging Face API:

curl -s https://huggingface.co/api/models/meta-llama/Llama-3.1-8B | jq '.cardData.license, .tags'

This command returns the license field from the model card and its associated tags. If the model declares a permissive license (Apache 2.0, MIT) and has no explicit geographic restrictions, the compliance risk is low. If the license is a modified version like Llama’s, you should read the acceptable use terms.

For a more thorough check, including the list of weight files, the Python SDK is more convenient:

from huggingface_hub import model_info

info = model_info('Qwen/Qwen3-72B')
print(info.cardData.get('license'))
print([s.rfilename for s in info.siblings if s.rfilename.endswith('.safetensors')])

This script prints the declared license and filters the .safetensors files, which are the ones that actually contain the trained weights (unlike configuration or tokenizer files). This is the kind of check any team redistributing fine-tuned models should automate in their CI pipeline, not leave as a manual step before a release.

⚠️ Heads up: If you work at a company that fine-tunes and redistributes US-origin models, check the license and destination country before publishing derived weights: the current regulatory ambiguity doesn’t mean there’s no legal risk.

Impact and analysis

The outcome of this debate matters beyond Nvidia and Microsoft. If Washington opted for a broad rule on open weights, the first effect would hit US labs that currently publish openly: Meta with Llama, the Allen Institute for AI with OLMo, Mistral (headquartered in France, but with a commercial presence in the US), and smaller startups that rely on fine-tuning open models instead of training from scratch.

The industry’s underlying argument is that this doesn’t reduce risk, it redistributes it. A developer, university, or lab in a third country that currently uses Llama or an OLMo model would simply migrate to Qwen, DeepSeek, or Kimi, which already compete on equal footing on several benchmarks and aren’t subject to US controls. The net result would be less US influence over which architectures, which safety safeguards, and which moderation practices become the de facto standard in the global open ecosystem.

On the other side, those defending stricter controls (national security sectors, some biosecurity researchers) argue that an unrestricted frontier model can facilitate dual-use capabilities: from assisting in pathogen synthesis to generating malicious code at scale. The tension has no clean technical resolution: it’s fundamentally a public policy decision about how much marginal risk to accept in exchange for maintaining US leadership in the open ecosystem.

What is technically clear is the enforcement asymmetry: any rule that depends on the publishing company’s jurisdiction leaves out, by design, the Chinese labs that currently lead much of open model adoption.

What’s next

The US regulatory process for a measure like this follows a familiar path: a proposed rule is first published in the Federal Register, a public comment period opens (where companies like Nvidia and Microsoft, as well as universities and civil organizations, can file formal objections), and only afterward does the Bureau of Industry and Security (BIS) issue a final rule. That’s exactly the path the AI Diffusion Rule followed, and then reversed, between January and May 2025.

As of this report, there is no specific proposed rule text on open weights publicly available to analyze line by line. What exists is the signal that the debate is back on the agenda, and that two of the most influential players in the AI industry have already staked out a position before any formal draft exists. Watch the Federal Register and BIS announcements in the coming weeks to see whether the discussion turns into a concrete proposal.

📖 Summary on Telegram: See summary

Try it yourself: go to huggingface.co/api/models/ followed by the name of the model you’re using today and check the license field before deciding whether redistributing those weights exposes you to any restriction.

Frequently Asked Questions

What does it mean for an AI model to have open weights?

It means the company that trained it publishes the file with the trained parameters for anyone to download, run locally, or fine-tune. This differs from a fully open-source model, which also publishes the training code and sometimes the data.

Why would the US government consider restricting its publication?

Due to national security concerns over dual use: an unrestricted frontier model could facilitate tasks ranging from automated cyberattacks to assistance in sensitive domains like biosecurity, if it falls into the hands of hostile actors.

What was the AI Diffusion Rule?

A rule published by the Biden administration in January 2025 that divided countries into three tiers of access to AI chips. It generated strong industry opposition for being too broad and was rescinded in May 2025.

Why do Nvidia and Microsoft, who compete with each other, agree on this?

Because both would be exposed if export controls were expanded to open weights: Nvidia because fewer open models means less GPU demand for experimentation, Microsoft because its Azure AI Foundry platform hosts third-party open models alongside its own.

Would Chinese open models like Qwen or DeepSeek be subject to the same rules?

No. Since they’re published by Chinese companies, they fall outside US jurisdiction. This is precisely Nvidia and Microsoft’s central argument: a broad rule would only restrict US labs.

How does this affect me if I only use open models to develop software?

For most developers, in the short term, nothing changes: downloading and using an already published model isn’t in question. Compliance risk appears if your company fine-tunes and redistributes weights derived from a US-origin model to third parties in other countries.

References

• Fox Business via MSN: original report on Nvidia and Microsoft’s joint request.
• Bureau of Industry and Security (BIS): US Department of Commerce agency that administers the export control regime (EAR).
• Federal Register: official registry where proposed and final export control rules are published.
• Hugging Face Docs: documentation on model licensing and distribution on the Hub.
• RAND Corporation: institute researching the security of frontier AI model weights.

📱 Enjoying this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.

Imagen destacada: Foto de BoliviaInteligente en Unsplash