US delays blacklisting DeepSeek and other Chinese firms

🔑 Enhanced Key Takeaways

•The delay in blacklisting DeepSeek and CXMT is part of a broader pause by the US government, which has not added new entities to the Commerce Department's Entity List since October 2025, marking the longest such gap in over a decade.
•DeepSeek gained global attention in January 2025 after launching a low-cost AI model that challenged existing assumptions about the financial investment required to build advanced artificial intelligence systems.
•US officials have accused DeepSeek of supporting China's military and intelligence operations and attempting to illicitly acquire advanced US chips through shell companies operating in Southeast Asia.
•AI companies Anthropic and OpenAI have both alleged that DeepSeek engaged in campaigns to extract capabilities from their respective AI models, Claude and OpenAI's proprietary systems, to enhance DeepSeek's own offerings.
•CXMT, recognized as China's leading memory chipmaker, was previously designated as a Chinese military company by the US Defense Department under the Biden administration, a separate designation from the Commerce Department's Entity List.
•DeepSeek recently completed its first external funding round, raising over $7.4 billion (50 billion yuan) and achieving a valuation exceeding $50 billion, making it China's most valuable AI startup.

📊 Competitor Analysis▸ Show

Feature/Model	DeepSeek V4 Pro (DeepSeek)	GPT-4o (OpenAI)	Claude 3.5 Sonnet (Anthropic)	Gemini 2.5 (Google)	Llama 3 (Meta)
Parameters	1.6 Trillion (49B active MoE)	Not publicly disclosed (estimated ~175B for GPT-4)	Not publicly disclosed	Not publicly disclosed	8B, 70B, 400B (open-weights)
Context Window	1 Million tokens	Not publicly disclosed	Large (unspecified)	Multimodal, large context (unspecified)	Customizable
Multimodal	Native multimodal (text, images, video, audio)	Advanced image and video generation	Not specified	Multimodal (text, images, audio, video, code)	Not specified
API Pricing	10-50x cheaper than GPT-5.4 and Claude Opus 4.8 (for V4)	Industry standard, more expensive than DeepSeek	More expensive than DeepSeek	Not specified	Open-source, self-hosted control
Architecture	MoE with Hybrid Attention (CSA + HCA), Transformer-based	Transformer-based	LLM	Multimodal, Generative AI	Large language model
Open-source	Open-source weights expected (Apache 2.0 for V4)	Proprietary	Proprietary	Proprietary	Open-weights
Performance (General)	Outperforms Gemini, GPT, Llama in some classification tasks, underperforms Claude	Top choice for general-purpose tasks	Superior reasoning, ethical, conversational	Excels in logical reasoning and factual accuracy	Popular for R&D
Performance (Speed)	Comparably slower than Gemini, GPT	Much faster than DeepSeek	Not specified	Much faster than DeepSeek	Not specified

🛠️ Technical Deep Dive

DeepSeek AI Models:
- DeepSeek-R1 (Released January 2025): Utilizes a Mixture-of-Experts (MoE) framework and an advanced transformer-based design. It has 671 billion parameters, but only approximately 37 billion are activated during a single forward pass, contributing to its cost-effectiveness and efficiency. It also incorporates Multi-Head Latent Attention (MLA) to optimize the attention mechanism, reducing memory overhead.
- DeepSeek V4 (Released April 2026): This flagship model is multimodal, trained on text, images, video, and audio simultaneously. It features approximately 1 trillion total parameters (with 32-37 billion active per token via MoE) and a 1 million token context window. The V4 Pro variant has 1.6 trillion total parameters and 49 billion activated parameters, employing a hybrid attention architecture combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA).
- Training: DeepSeek models leverage reinforcement learning and distillation techniques to transfer knowledge from larger models into more efficient ones. They also use FP8 precision for training to overcome compute limitations.
CXMT Memory Chips:
- Products: CXMT specializes in Dynamic Random-Access Memory (DRAM) and produces DDR4, LPDDR4X, LPDDR5, and DDR5 memory chips for mobile devices, PCs, tablets, and servers.
- DDR5 Technology: CXMT's DDR5 chips achieve speeds up to 8000Mbps with die capacities of 16Gb/24Gb. They incorporate on-die Error Correction Code (ECC) and enhanced anti-interference designs for stability and data integrity.
- Process Nodes: As of 2020, CXMT manufactured LPDDR4 and DDR4 memory on a 19 nm process. The company entered mass production of 16nm DRAM in early 2025 and is targeting its G5 (15nm-class) process for completion by late 2025, with samples expected in 2026.
- Advanced Memory: CXMT has also advanced into mass production of High-Bandwidth Memory (HBM3) and is reportedly collaborating with Huawei on the next-gen Ascend 950 AI accelerator.

🔮 Future ImplicationsAI analysis grounded in cited sources

The delay in blacklisting could signal a strategic shift in US policy towards China, prioritizing de-escalation over immediate punitive measures.

The Trump administration's decision to hold off on these listings, despite interagency approval, suggests a political calculation to prevent further deterioration of US-China relations, as reported by Reuters.

Chinese AI and semiconductor firms, including DeepSeek and CXMT, may use this reprieve to further accelerate domestic technological self-sufficiency.

DeepSeek's success, despite previous US sanctions, has already been attributed to catalyzing Chinese innovation and forcing improvements in efficiency and local hardware development, making future sanctions potentially less impactful.

The continued operation of DeepSeek and CXMT without Entity List restrictions could intensify competition for US AI and semiconductor companies in global markets.

DeepSeek's low-cost, high-performance AI models already challenge Western giants like OpenAI and Anthropic, and CXMT is rapidly catching up to leading DRAM manufacturers, potentially impacting market share and technological leadership.

⏳ Timeline

2016-05

ChangXin Memory Technologies (CXMT) founded as a commercial startup.

2019-09

CXMT announced 8Gb DDR4 for mass production.

2023-07

DeepSeek founded by Liang Wenfeng.

2025-01

DeepSeek launched its DeepSeek-R1 model and eponymous chatbot, gaining global attention for its low cost and performance.

2026-04

DeepSeek V4, a 1 trillion parameter multimodal AI model, was launched.

2026-06

DeepSeek completed a funding round of over $7.4 billion, valuing the company at more than $50 billion.

US delays blacklisting DeepSeek and other Chinese firms

⚡ 30-Second TL;DR

🧠 Deep Insight

🔑 Enhanced Key Takeaways

🛠️ Technical Deep Dive

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline

📎 Sources (31)

👉Related Updates

Taiwan debates NT$240bn investment in drone defense systems