The Main Players in the US-China AI War are Changing: Why Meta and NVIDIA are Pursuing Chinese Companies with "Open Weights"

When people hear about the competition in generative AI, many might think of American companies like OpenAI, Anthropic, and Google.

However, understanding the AI market in 2026 requires more than just this perspective.

Rapidly gaining presence now is "open-weight AI."

On August 10, Meta released a 30B-scale AI model called "Muse Glimmer." Following this, NVIDIA announced "Nemotron 3.5 Lightning" on August 11.

Both models emphasize not just providing AI as a cloud-based API, but allowing developers and companies to acquire the models, run them on their own computers or servers, and customize them for their specific needs.

The simultaneous actions of these two companies carry significance beyond mere product competition.

This is because Chinese companies already have a significant presence in the field of open-weight AI.

Chinese AI models such as DeepSeek, Alibaba's Qwen, and Moonshot AI have penetrated the global developer community, not only with their performance but also by allowing models to be downloaded and used in private environments.

The AI competition is no longer a simple race to create the best-performing chat AI.

It is evolving into a competition over who can provide the models that developers worldwide will use as a foundation, essentially a competition over the AI ecosystem itself.

From "Using AI" to "Owning AI"

The term "open-weight" is still not widely understood.

In services like ChatGPT and Claude, users typically access the operating company's servers and use AI through APIs or web services.

They do not possess the AI model itself.

In contrast, open-weight models allow the download of a large number of parameters, known as "weights," formed through AI learning.

Therefore, if one has a PC or server with sufficient computing power, AI can be run internally without sending data to external AI services each time.

This difference is significant.

For companies, it's not just about cost.

There is a lot of data, such as internal documents, customer information, R&D data, source code, and financial information, that must be handled cautiously when sending to external services.

If models can be operated in-house, the potential to process such information with AI without external exposure expands.

Additionally, it becomes easier to conduct additional training and tuning tailored to the company's operations.

In other words, open-weight AI embodies the idea of "owning AI as part of the company's IT infrastructure" rather than "renting AI."

Why Meta Returned to an Open Strategy

For Meta, an open AI strategy is not new.

The company became one of the representatives of the open model market with its Llama series.

Developers worldwide developed derivative models and tools based on Llama, allowing Meta to wield strong influence over the entire AI ecosystem without monopolizing the AI API market itself.

However, Meta later revised its AI strategy and focused on providing closed models with the Muse series.

Therefore, there is some caution about the recent return to open-weight, with concerns that the policy might change again.

The recently released Muse Glimmer, with approximately 30 billion parameters, is designed with the intention of running continuously as a local AI agent.

According to Meta, it is intended to be used on Macs, general PCs, and single consumer GPUs, with local coding support, tool invocation, and AI agents as the main applications.

More importantly, Meta has indicated that it plans to offer an open-weight version of the more advanced "Muse Spark 1.2."

If a model close to the current cutting-edge class is indeed released, the positioning of US-made open-weight AI could change dramatically.

Aaron Levie, CEO of Box, also highly appreciates this move.

What he focuses on is not just the addition of a new AI model.

It is the clear indication by American companies of their intention to continue providing open models with near-cutting-edge performance.

NVIDIA Targets the "AI Agent Era"

Unlike Meta, NVIDIA is approaching open AI from a different angle.

The Nemotron 3.5 Lightning announced by the company is also approximately 30 billion parameters in scale, but it does not use all parameters simultaneously.

It employs a mechanism called Mixture-of-Experts, limiting the parameters active in a single operation to about 3 billion.

This aims to maintain the capabilities of large-scale models while reducing computational costs.

NVIDIA emphasizes the application of "constantly operating AI agents."

As AI agents become more prevalent, unlike traditional chat AI that answers a single question at a time, AI will invoke models dozens or hundreds of times.

For example,

"Research competitors,"
"Gather materials,"
"Analyze data,"
"Write reports,"
"Compose emails,"
"Input into systems,"

entrusting these tasks to AI will generate a large amount of inference processing behind a single operation.

If the highest-performing, expensive large models are used each time, costs will rapidly escalate.

Therefore, NVIDIA envisions assigning only advanced judgments to large models, while distributing the numerous processes like tool execution and verification tasks to lightweight models like Nemotron 3.5 Lightning.

This suggests that the future AI market may not converge into "one giant AI," but rather move towards using multiple AIs selectively.

Why Chinese Companies Are Strong in Open-Weight AI

The presence of China cannot be ignored here.

The impact of DeepSeek on the world was not merely that a Chinese company developed a high-performance AI.

By making high-performance models developed at relatively low cost widely available, it challenged the assumption that "cutting-edge AI is created only by large American companies and provided as cloud APIs."

Subsequently, Alibaba's Qwen and Moonshot AI have also gained strong influence in the global open model market.

Especially in the world of developers using local LLMs, "performance," "required memory," "inference speed," "license," and "ease of quantization" are more important than the model's nationality.

If a model performs well, runs on a local GPU, and can be freely modified, Chinese models will be used.

This highly pragmatic developer culture has supported the global spread of Chinese models.

This is the issue for American companies.

Once a model becomes the standard foundation for developers, tools, libraries, training data, know-how, and derivative models accumulate around it.

This is similar to an operating system.

An "ecosystem" is formed that is difficult to overturn by temporarily winning in performance comparisons.

The recent moves by Meta and NVIDIA also have a strategic meaning of wanting to increase the options for US-made models again before Chinese companies establish a decisive advantage in this area.

Mixed Reactions on Social Media

The reaction from the developer community to the recent announcements is not entirely welcoming.

On X, Aaron Levie perceives the open-weight release of Muse Spark 1.2 as a significant event, and there are positive reactions from those supporting open AI, including people related to Hugging Face, welcoming Meta's change in direction.

Among AI developers, there is particularly high expectation for the point that "prominent open-weight models are emerging again from American companies."

On Hacker News, there are comments expressing a sense of anticipation similar to the "era of Llama 3," and posts favorably evaluating the chat performance and text generation of Muse Glimmer tested on Mac.

On the other hand, there are also many cautious opinions.

One of the most common points of discussion is trust in Meta.

Despite Meta's strong promotion of open AI with Llama previously, it later made some models closed.

Therefore, on Hacker News, there are questions like "Why did they stop open models once?" and "How long will this policy last?"

Additionally, there are concerns about whether they will truly release competitive cutting-edge models while releasing smaller models.

This is why the release of Muse Spark 1.2 is attracting attention.

Will the promise of open-weight even for high-performance models be fulfilled, or will the strategy be limited to releasing only some models?

The seriousness of Meta's open strategy is likely to be judged by the next release.

Local AI Users Have Already Started Performance Verification

On Reddit's local LLM community "LocalLLaMA," posts attempting to run Muse Glimmer and Nemotron 3.5 Lightning have been appearing since the announcement.

Regarding Muse Glimmer, while some users are testing it for local coding purposes, harsh evaluations are also emerging through actual use, such as "tends to overthink" and "might be too restrictive."

For Nemotron 3.5 Lightning, discussions are ongoing about how practical it can be in VRAM environments of 16GB or 24GB, and how usable the quantized version will be.

Interestingly, Qwen is frequently mentioned as a comparison.

For American companies, Chinese AI is no longer a "potential future competitor."

It is a real benchmark that developers compare new American models against the moment they try them.

This symbolizes the current structure of the AI competition.

"Open Means Safe" or "Open Means Dangerous"?

There is another major debate regarding open-weight AI.

It concerns security and safety.

While making models available to everyone can promote research and corporate use, it also allows malicious individuals to exploit them.

Concerns about misuse for cyberattacks, fraud, and disinformation generation lead to the idea that certain restrictions on the release of powerful AI models are necessary.

On the other hand, there is an argument that "it is more dangerous for only a few large companies to control AI models."

Being able to research the internals of models, independently verify their safety, and use them without relying on external services can also be a security advantage.

The important thing is not to judge safety solely based on whether something is open or closed.

It is necessary to consider the scope of information disclosed, the model's capabilities, usage licenses, security evaluations, and the management systems of the organizations implementing them.

The Biggest Meaning for Japanese Companies is "Where to Place Data"

How will this US-China open AI competition affect Japan?

First is the cost of implementing generative AI for companies.

Currently, many companies use AI services or APIs provided by overseas companies.

While there is a significant advantage of easy implementation and high performance, API fees increase as AI is used extensively.

Especially as AI agents enter daily operations and models are called dozens of times for a single task, managing inference costs becomes a critical issue for corporate IT.

If more high-performance models that can run locally become available,

advanced judgments can be made by cloud-based cutting-edge AI,

a large amount of routine processing can be handled by open models on company servers,

and processes involving confidential information can be managed by completely on-premises AI.

This allows for such differentiation.

The second is data sovereignty.

In fields like finance, healthcare, manufacturing, R&D, and administration, where to process confidential data becomes extremely important.

Being able to place models in their own environment becomes a significant option for Japanese companies.

It's not just about "American-made or Chinese-made."

The important issue is "whose server the data is sent to, who manages the model, and whether company operations can continue even if the service is suspended."

Can Japan Aim to Be Not Only a "Country That Creates AI" but Also a "Country That Masters AI"?

The Japanese government also positions AI utilization as an important policy.

In the Basic Plan for Artificial Intelligence Phase II, decided by the Cabinet in July 2026, important themes include not only AI research and development but also implementation in industry and administration, AI agents, robotics, and governance.

The "DX Trends 2026" published by IPA also shows that AI adoption is rapidly expanding among Japanese companies.

On the other hand, continuing the competition to develop foundational models on the same scale as the giant AI companies in the US and China requires enormous capital and computational resources.

Therefore, the expansion of open-weight models is not necessarily bad news for Japan.

This is because it becomes easier for Japanese companies to use high-performance models released by foreign companies and optimize them for specific fields such as Japanese language, manufacturing, finance, healthcare, and administration.