China Weighs Restricting Top AI Model Exports as DeepSeek Develops Its Own Inference Chip

Beijing Considers Restricting Top AI Model Exports

According to Reuters, China's Ministry of Commerce has held intensive meetings over the past month, engaging in consultations with leading domestic AI companies to discuss whether to restrict overseas access to China's most advanced AI models. The companies reportedly involved include Alibaba, ByteDance, and startups such as 01.AI, while officials from the National Development and Reform Commission and national planning departments also participated.

You may not have noticed that the scope of these discussions is not limited to closed-source systems—it also covers open-weight models, precisely the products such as Alibaba's Tongyi Qianwen (Qwen), ByteDance's Doubao, and 01.AI's Yi series.

Open-weight models differ fundamentally from fully open-source models: the former publicly release the model parameters (weights), allowing users to deploy and fine-tune locally, but do not necessarily open up the training data or complete code. Understanding this distinction requires some background: the so-called "model weights" refer to the billions or even hundreds of billions of numerical parameters that a neural network solidifies after large-scale data training—these parameters are essentially the carriers of the model's "learned knowledge." From a technical architecture perspective, the weight files of modern large language models are typically stored in formats such as safetensors or GGUF, often reaching tens or even hundreds of GB, containing all the parameters of the Attention Layers, Feed-Forward Networks (FFN), and Embedding Matrix in the Transformer architecture. Publishing these weights means that any entity with sufficient computing power—from individual developers to sovereign nations—can deploy, fine-tune, and even modify the model in a completely offline environment, and the original developers can impose almost no technical-level restrictions on this.

This "limited openness" strategy has allowed Chinese AI companies to build powerful ecosystem influence within the global developer community. Publishing weights means anyone can run the complete model locally, without needing to make networked API calls or pay the original developers—a stark contrast to closed-source models that only provide API interfaces. Take Alibaba's Qwen as an example: its downloads on Hugging Face have reached hundreds of millions, forming a large number of downstream applications dependent on Chinese AI infrastructure. The core strategic value of open weights lies in this: once developers build applications based on a particular model, a technological path dependency forms—analogous to the "lock-in effect" of commercial software. It is precisely for this reason that these models are core assets for building China's AI reputation abroad: low price, excellent performance, and low deployment barriers.

Since DeepSeek exploded in popularity, its open-source models have been widely embraced by overseas developers on Hugging Face. Hugging Face is currently the world's largest AI model hosting and collaboration platform, functioning like the GitHub of AI, hosting over 1 million public models and hundreds of thousands of datasets. The platform was founded in New York in 2016, initially focusing on natural language processing tools, and has now evolved into a comprehensive community covering nearly all AI directions including vision, speech, and multimodal. The reason Hugging Face has become the central hub for global AI model distribution lies in the complete interoperability standard it has built: at its core is the transformers library, which provides a unified model loading interface for the PyTorch and JAX frameworks, allowing developers to load any hosted model with just three lines of code; the "Model Card" specification establishes a comparable foundation of trust through a standardized metadata format. The network effect of this infrastructure is extremely strong: the datasets, fine-tuning scripts, and evaluation code that developers accumulate on the platform form a collective knowledge base difficult for any single company's private platform to replicate. What Chinese models gain through this platform is not just download counts, but a strategic entry point into becoming part of the default toolchain of global AI developers. Within days of DeepSeek-R1's release, its download count exceeded one million, far surpassing the spread speed of its official channels. Its combination of low cost and high performance is extremely attractive—reportedly, some versions have performance approaching top American models at a fraction of the cost. This openness is precisely the key to China's AI going global. If China tightens the overseas release of open weights, it could mean such models are no longer uploaded to Hugging Face, or only "capability-neutered" versions are provided, directly affecting global independent developers' access to Chinese AI technology.

A Tiered Regulatory Framework Emerges

Current discussions lean toward establishing a regulatory system tiered by model capability:

• Basic open-source models: may only require lightweight filing procedures with regulators;
• High-performance models: may need to pass a formal security review before public release;
• The most sensitive frontier models: may be prohibited from public release, or restricted to use within China only.

In addition, the meetings raised two noteworthy proposals: first, classifying the leaking or theft of proprietary AI technology as endangering national security and imposing severe penalties; second, restricting foreign investment access to Chinese AI startups. Taken together, Beijing appears to be attempting to elevate frontier AI to the same status as other strategic technologies—with a focus on protection and control rather than mere market showcasing.

A Mirror Game About Weaponizing AI

This development forms a subtle contrast with the direction of U.S. policy. The United States has recently continued to tighten export controls on advanced AI models, with the legal basis primarily coming from the Export Administration Regulations (EAR) and the International Emergency Economic Powers Act (IEEPA). The EAR is enforced by the Bureau of Industry and Security (BIS) of the U.S. Department of Commerce. Originally designed to regulate the export of physical goods, its ECCN (Export Control Classification Number) system provides relatively clear identification for tangible goods, but faces significant legal interpretation challenges when applied to intangible software and model weights—the classification of model weights, as "intangible carriers of intellectual property," remains legally disputed to this day. The IEEPA grants the president broad economic control powers under a "national emergency," and has been frequently invoked in recent years to rapidly advance technology control measures.

The Biden administration's 2023 "Advanced Computing and Semiconductor Manufacturing Export Controls" rule uses compute thresholds (FLOP/s, or floating-point operations per second) as a core metric, setting caps on the capabilities of chips and models exported to China; the "diffusion rule" advanced in early 2025 further extends controls to third countries to prevent circumvention through transshipment. However, this measurement method itself has loopholes: as MoE (Mixture of Experts) architectures become widespread, the separation between "active parameter count" and "total parameter count" causes FLOP/s thresholds to lose accuracy—the compute consumed in training DeepSeek-V3 was far below what its parameter scale would imply, which precisely demonstrates the limitations of measuring model capability with a single compute metric. Part of the reasoning stems from concerns that cutting-edge systems could be misused in cyberattacks, military, or intelligence domains.

Notably, export controls on software and model weights are far more complex to enforce at a technical level than hardware—digital files can be transmitted instantly over networks, and how to conduct verifiable technical assessments of "model capability" remains a common challenge faced by regulators worldwide to this day. Unlike chips, once model weights are publicly released, they are difficult to "recall"—anyone can mirror-host them on overseas servers, rendering after-the-fact controls almost futile. If China establishes a similar tiered framework, it will face the same enforcement challenges.

And China reportedly harbors similar concerns: advanced U.S. cybersecurity AI models could be used to discover and exploit vulnerabilities in Chinese software. The two sides stand on opposite sides, yet harbor the same fear—that the other's models will be turned into weapons to attack their own systems. The features that give a model defensive capabilities are precisely the same features that make the opponent wary of its existence, evolving into an almost perfectly symmetrical mirror game.

Tightening Controls Is Not an Isolated Event

In fact, China's tightening of AI controls has continued for months. Regulators are reportedly investigating Chinese AI companies such as Manus that have relocated their operations overseas, scrutinizing whether they have engaged in circumventing export restrictions. In early June this year, the scope of government review had expanded to all cross-border transactions involving Chinese capital and technology.

Internally, Alibaba has reportedly required employees to stop using Anthropic's models and related agent tools, switching instead to the company's self-developed assistant. Regardless of whether this is directly related to the government meetings, it aligns with China's current overall strategic direction of building an "autonomous, controllable, closed ecosystem."

If the Policy Takes Effect, the Global AI Landscape Will Change

If the above framework ultimately takes effect, it will be a major policy shift. The global reputation of China's AI industry is largely built on a foundation of "openness"—companies like Alibaba proactively publish model weights for global developers to freely download and further develop. Even partial tightening would change how the world interacts with China's AI industry, and could prompt more international developers to default to American labs.

However, rational judgment must be maintained: nothing is settled yet, and the actual rules may not be as strict as rumored. Multiple sources lean toward believing that restrictions may only target unreleased frontier models, rather than existing released products. If so, the actual impact on the existing open-source ecosystem may be far smaller than headlines suggest. But the "open weights" dimension is undoubtedly the core focus most worth closely watching.

DeepSeek Quietly Develops Its Own Inference Chip

That said, another major piece of news comes from DeepSeek. According to three people familiar with the matter cited by Reuters, DeepSeek has quietly been developing its own chip for about a year, aiming to reduce its dependence on Nvidia and Huawei.

This chip is dedicated to inference (i.e., running already-trained models), not for training new models. Understanding this distinction is crucial: AI chip application scenarios fall into two major categories—Training and Inference. The training phase requires repeated iteration over trillions of text or image tokens, continuously adjusting model parameters through the Backpropagation algorithm, a process with extremely high requirements for High Bandwidth Memory (HBM) and high-precision floating-point operations (FP16/BF16)—the Nvidia H100 is equipped with 80GB of HBM3 memory and provides 900GB/s of NVLink interconnect bandwidth, precisely optimized for this "globally synchronized" parallel computing mode. It typically requires hundreds or even thousands of GPUs running in parallel for weeks or even months.

The inference phase is entirely different: model parameters are fixed, computation shifts to an "autoregressive generation" mode, generating only one token at a time, and the core bottleneck shifts from compute to the memory access efficiency of the KV Cache (key-value cache). The chip's task is to rapidly generate output for each user request, with core metrics shifting to Time to First Token, tokens generated per second (Tokens/s), and throughput per watt. Inference chips can typically adopt low-precision quantization formats such as INT8 or even INT4, greatly reducing memory footprint and power consumption while maintaining acceptable precision loss. The reason Apple's A-series chips can efficiently run medium-scale models on edge devices is precisely a reflection of this optimization direction. From a business logic perspective, inference is the ongoing cost after AI is productized, and becomes more critical at scale—OpenAI processes billions of queries daily, with inference costs making up the major portion of its operating expenses.

DeepSeek has been secretly recruiting chip engineers without posting any public job listings, and has already been in contact with foundry manufacturers and memory chip suppliers.

Why Enter the Hardware Game Now?

DeepSeek, known for its model development, entering the hardware game at this time is intriguing in its timing. The models it launched earlier shocked the industry with their extremely low operating costs, even rattling the U.S. tech stock market—and those models were trained using Nvidia H800 chips before export restrictions took effect. The H800 is a replacement version of the H100 that Nvidia released for the Chinese market with "neutered" interconnect bandwidth; with single-card compute power roughly comparable, multi-card communication efficiency drops significantly. DeepSeek engineers achieved surprising performance on constrained hardware through a series of algorithmic innovations (including the Mixture of Experts (MoE) architecture and efficient memory management mechanisms), and this experience may well be an important catalyst for its determination to take proactive action at the hardware level. Developing its own inference chip both avoids the extremely high technical barriers of training chips and directly targets the core pain point of large-scale commercial deployment, precisely to fill this technological gap: better controlling costs and the supply chain, and escaping the passive situation of "using whatever chips are available."

This trend is not unique to DeepSeek. OpenAI has released a self-developed chip in collaboration with Broadcom, and Anthropic is reportedly also considering the self-development route. More and more AI companies are seeking hardware autonomy rather than relying solely on renting compute.

First External Financing and Market Limitations

Notably, DeepSeek is opening external financing for the first time while developing its chip, reportedly raising about $7 billion at a valuation of nearly $55 billion—a major strategic shift for a company that had previously deliberately avoided external investment.

However, analysts also point out that, given Nvidia has already been shut out of the Chinese market, DeepSeek's self-developed chip may also be difficult to sell overseas, unless it can obtain support from more advanced manufacturing processes. Currently, China's most advanced domestic foundry capability comes mainly from SMIC (Semiconductor Manufacturing International Corporation), whose mass production process is around the 14nm to 7nm level. Constrained by the inability to obtain ASML's Extreme Ultraviolet (EUV) lithography machines, its yield and density lag significantly behind TSMC's true 7nm—while TSMC has already mass-produced 3nm and is advancing toward 2nm. This gap in manufacturing process will directly constrain the chip's performance in inference-per-watt (Power Efficiency), which has a direct impact on the electricity costs of data-center-scale deployment, and it will face an inherent disadvantage in performance and energy efficiency compared to Nvidia's Blackwell architecture products manufactured on TSMC's advanced processes. Therefore, it is unlikely we will see this chip outside China in the short term. Once successfully developed, it will join Alibaba's Hanguang series and Baidu's Kunlun chips to become yet another strong competitor to Huawei's Ascend, together building the "autonomous, controllable" foundation of China's AI computing power. After the news broke, Nvidia's stock price dropped at one point in pre-market trading.

Conclusion

From restricting open-source exports to developing its own inference chip, China's AI industry is simultaneously advancing an "autonomous and controllable" strategy on two fronts. The former concerns China's open image and market influence in AI globally, while the latter points toward deep control over the hardware supply chain. Regardless of the final policy intensity, one major direction is already clear: AI is increasingly being viewed as a national strategic asset, incorporated into a broader technology competition game. For global developers, future technology selection may face more geopolitical considerations.

Key Takeaways

Key Takeaways

Key Takeaways