Technical Deep Dive
The US strategy shift from targeting AI models to infrastructure represents a sophisticated understanding of the AI value chain. At its core, modern AI development depends on three interdependent layers: compute hardware (GPUs, TPUs, ASICs), software frameworks (training, inference, orchestration), and data pipelines (collection, curation, storage, transmission). The blacklist expansion targets companies operating at each layer.
Compute Hardware Layer: The US export controls have already restricted NVIDIA's A100, H100, and now the H200 and B100 series to China. However, Chinese companies like Huawei (Ascend 910B), Cambricon (MLU370), and Biren Technology (BR100) have developed alternatives. The blacklist now targets companies involved in advanced packaging (e.g., JCET, Tongfu Microelectronics) and semiconductor manufacturing equipment (e.g., Naura Technology, AMEC). This is a direct attack on China's ability to produce competitive AI chips domestically.
Software Framework Layer: The blacklist includes companies developing AI training frameworks and inference optimization tools. DeepSeek's exemption here is notable—its open-source model weights (e.g., DeepSeek-V2, DeepSeek-R1) have been widely adopted, but the company does not control the underlying frameworks like PyTorch or TensorFlow. However, Chinese alternatives like PaddlePaddle (Baidu), MindSpore (Huawei), and OneFlow (OneFlow Inc., now part of Zhipu AI) are gaining traction. The blacklist targets companies that could create a closed-loop Chinese AI software stack independent of US-controlled infrastructure.
Data Pipeline Layer: This is the most overlooked but critical dimension. The blacklist now includes companies involved in data center operations, cloud services, and cross-border data transmission. Alibaba Cloud, Tencent Cloud, and Huawei Cloud are already under scrutiny, but the new designations extend to smaller data center operators and fiber-optic cable companies. The goal is to prevent Chinese AI companies from accessing high-quality training data from global sources or using international data centers for model training.
Open-Source Implications: DeepSeek's open-source strategy has been a double-edged sword. On one hand, it has accelerated global AI research and adoption. On the other, it has made DeepSeek less of a direct threat to US dominance because its models run on hardware that the US can still control. However, the open-source ecosystem creates a dependency: if DeepSeek's models become the de facto standard for Chinese AI applications, and those applications begin to drive demand for Chinese-designed chips, the US could lose control over the hardware-software co-optimization loop.
| Layer | US Control Mechanism | Chinese Counterpart | Blacklist Status |
|---|---|---|---|
| Compute Hardware | Export controls on NVIDIA/AMD GPUs | Huawei Ascend, Cambricon, Biren | Heavily targeted |
| Advanced Packaging | Equipment restrictions | JCET, Tongfu, Huatian | Newly targeted |
| AI Frameworks | Open-source (PyTorch, JAX) | PaddlePaddle, MindSpore, OneFlow | Selectively targeted |
| Cloud Infrastructure | Data center access controls | Alibaba Cloud, Tencent Cloud, Huawei Cloud | Expanding coverage |
| Model Weights | Open distribution allowed | DeepSeek, Qwen, Yi, ChatGLM | Currently exempt |
Data Takeaway: The US is systematically closing off every layer of the AI stack except model weights, which are inherently difficult to control once open-sourced. This suggests a bet that controlling hardware and data will ultimately constrain model development more effectively than banning specific models.
Key Players & Case Studies
DeepSeek (Hangzhou): The company's exemption is the most analyzed aspect of this policy shift. DeepSeek has focused on efficient model architectures (Mixture-of-Experts, multi-head latent attention) that achieve competitive performance with fewer parameters. Its DeepSeek-V2 model, with 236B total parameters but only 21B activated per token, demonstrates a compute-efficient approach. This efficiency reduces its dependence on the most advanced chips, making it a less immediate target. However, DeepSeek's recent open-source release of DeepSeek-R1, a reasoning model that rivals OpenAI's o1, has raised its profile. The company's founder, Liang Wenfeng, has publicly stated that DeepSeek's goal is to democratize AI access, which aligns with the open-source ethos but also makes it harder to control.
Huawei: The blacklist expansion directly targets Huawei's AI ambitions. The company's Ascend 910B chip, while not matching NVIDIA's H100 in raw performance, has been adopted by major Chinese AI labs including Baidu and Tencent. Huawei's MindSpore framework and its cloud services (Huawei Cloud) create a vertically integrated AI stack that the US views as a direct threat. The blacklist now covers subsidiaries involved in chip design (HiSilicon), software (Huawei Technologies), and cloud infrastructure.
Baidu: The company's PaddlePaddle framework and ERNIE models have been central to China's AI ecosystem. Baidu's Kunlun chips, while less advanced than Huawei's, represent another domestic alternative. The blacklist targets Baidu's AI cloud division and its autonomous driving unit (Apollo), indicating that the US is concerned about AI applications in critical infrastructure.
Zhipu AI: This Beijing-based startup, backed by Alibaba and Tencent, has developed the GLM series of models. Its open-source ChatGLM-6B has been widely used in China. Zhipu AI was not initially on the blacklist, but its close ties to the Chinese government and its focus on enterprise AI applications make it a likely future target.
| Company | Primary AI Focus | Blacklist Status | Key Threat to US |
|---|---|---|---|
| DeepSeek | Open-source LLMs | Exempt (current) | Model efficiency, open-source adoption |
| Huawei | Full-stack AI (chips, framework, cloud) | Targeted | Vertical integration, domestic chip supply |
| Baidu | AI framework (PaddlePaddle), autonomous driving | Targeted | Ecosystem lock-in, critical infrastructure |
| Alibaba | Cloud AI (Qwen models), e-commerce AI | Targeted | Cloud dominance, data access |
| Cambricon | AI chips (MLU series) | Targeted | Hardware alternative to NVIDIA |
| Zhipu AI | Enterprise LLMs (GLM series) | Currently exempt | Government adoption, enterprise market |
Data Takeaway: The blacklist is not random—it systematically targets companies that control infrastructure layers (chips, cloud, frameworks) while temporarily exempting those focused on model development. This suggests a strategic bet that infrastructure control is more durable than model control.
Industry Impact & Market Dynamics
The selective blacklist creates a bifurcated landscape for China's AI industry. Companies with heavy reliance on foreign computing resources face immediate disruption, while those with domestic alternatives may gain competitive advantage.
Short-Term Impact: Chinese AI startups that rely on NVIDIA GPUs accessed through cloud providers or gray markets will face increased costs and uncertainty. The blacklist targets companies that facilitate GPU access, including distributors and cloud resellers. This could slow down the pace of model development for smaller players, potentially consolidating the market around larger companies with domestic chip access (Huawei, Baidu) or those with efficient model architectures (DeepSeek).
Medium-Term Impact: The blacklist accelerates the development of China's domestic AI supply chain. Huawei's Ascend ecosystem is already seeing increased adoption. The Chinese government is likely to increase subsidies for domestic chip purchases and cloud services. However, the performance gap between domestic chips and NVIDIA's latest offerings remains significant. The latest NVIDIA B200 GPU delivers approximately 4x the training performance of Huawei's Ascend 910B for large language models. This gap will constrain the scale of models that Chinese companies can train.
Long-Term Impact: The US strategy may backfire if it pushes China to develop completely independent AI infrastructure. The blacklist incentivizes investment in alternative computing architectures (e.g., analog computing, optical computing, neuromorphic chips) and software optimizations (e.g., quantization, pruning, knowledge distillation). DeepSeek's efficiency-focused approach could become the dominant paradigm in China, potentially leading to models that achieve comparable performance with significantly less computing power.
| Metric | US Ecosystem (NVIDIA) | Chinese Ecosystem (Huawei) | Gap |
|---|---|---|---|
| Peak GPU FLOPs (FP16) | 1,979 TFLOPS (H100) | 320 TFLOPS (Ascend 910B) | 6.2x |
| Memory Bandwidth | 3.35 TB/s (H100) | 1.2 TB/s (Ascend 910B) | 2.8x |
| Interconnect Bandwidth | 900 GB/s (NVLink) | 200 GB/s (HCCS) | 4.5x |
| Training Time (GPT-3 175B) | ~34 days (1,000 H100s) | ~120 days (1,000 Ascend 910Bs) | 3.5x |
| Cost per 1M tokens (inference) | $0.10 (GPT-4o) | $0.08 (DeepSeek-V2) | -20% (advantage China) |
Data Takeaway: While China's hardware gap is substantial, its efficiency advantage in inference (due to model optimization and lower labor costs) creates a competitive niche. The blacklist may accelerate a strategic divergence where China prioritizes efficiency over raw scale.
Risks, Limitations & Open Questions
Risk of Over-Targeting: The blacklist's broad scope risks creating a self-fulfilling prophecy. By designating over 100 companies as security risks, the US may push these companies into closer collaboration with the Chinese military and state security apparatus, precisely the outcome the policy aims to prevent.
Enforcement Challenges: The blacklist is only as effective as its enforcement. Chinese companies have demonstrated creativity in circumventing controls through shell companies, third-country transshipment, and technology licensing. The US lacks the resources to monitor every transaction, and allied countries (e.g., Netherlands, Japan, South Korea) may not fully cooperate.
Unintended Consequences for US Companies: The blacklist disrupts supply chains for US companies that sell to Chinese firms. NVIDIA, for example, has already lost billions in potential revenue from China. The blacklist expansion will further reduce the market for US semiconductor equipment makers and design software companies.
DeepSeek's Strategic Ambiguity: The biggest open question is whether DeepSeek will remain exempt. The company's open-source strategy makes it difficult to control, but its growing influence could trigger inclusion. DeepSeek's recent hiring of hardware engineers and its exploration of custom chip design suggest it may be moving toward infrastructure control, which would make it a target.
Data Sovereignty Concerns: The blacklist targets data pipelines, but the US has limited ability to control data flows within China. Chinese companies are already building massive domestic data centers and training datasets. The question is whether these datasets are sufficient for cutting-edge AI development or whether access to global data (e.g., multilingual corpora, scientific literature) remains essential.
AINews Verdict & Predictions
The selective blacklist of DeepSeek while targeting over 100 other Chinese tech firms is a masterclass in strategic differentiation. The US is correctly identifying that the most dangerous Chinese AI companies are not those building the best models, but those building the infrastructure that could make the US AI supply chain irrelevant.
Prediction 1: DeepSeek will be added to the blacklist within 12 months. The company's open-source ecosystem is growing too fast, and its influence on China's AI hardware roadmap is becoming too significant. The US will find a pretext—likely related to data security or military applications of its models.
Prediction 2: China will accelerate investment in alternative computing architectures. The blacklist makes it clear that relying on von Neumann architectures (GPUs) is a strategic vulnerability. Expect increased funding for optical computing (Lightmatter, Lightelligence), analog AI chips (Mythic), and neuromorphic processors (Intel Loihi, but Chinese equivalents).
Prediction 3: The US-China AI competition will bifurcate into two distinct ecosystems. The US ecosystem will prioritize scale and raw performance (larger models, more GPUs), while the Chinese ecosystem will prioritize efficiency and integration (smaller models, specialized hardware, tight software-hardware co-optimization). This divergence will make direct model comparisons increasingly meaningless.
Prediction 4: The blacklist will fail to prevent China from achieving AI parity in specific domains. China will likely match or exceed US capabilities in AI applications for manufacturing, logistics, and government services, where data access and integration matter more than raw model performance. The US will maintain leadership in foundational research and cutting-edge applications.
What to Watch: Monitor DeepSeek's hiring patterns (especially hardware engineers), Huawei's Ascend ecosystem adoption rates, and Chinese government subsidies for domestic AI chips. The next 6-12 months will determine whether the US strategy succeeds in constraining China's AI ambitions or merely accelerates the creation of a parallel, independent AI ecosystem.