What Happened
- A senior Trump administration official stated that China's DeepSeek used Nvidia's most advanced AI chip — the Blackwell series — to train a new AI model, despite US export controls prohibiting the shipment of such chips to China
- The US believes DeepSeek will remove technical indicators from the model that might reveal the use of American-origin chips; the Blackwell chips are reportedly clustered at DeepSeek's data centre in Inner Mongolia, China
- The DeepSeek model in question reportedly also relied on "distillation" — a technique where outputs from leading US AI models (including those by Anthropic, Google, OpenAI, and xAI) are used to improve a newer model's performance
- The revelation has intensified the debate within the US administration between two factions: those advocating tighter chip controls (China hawks) and those arguing that strict controls push China toward developing its own competitive chip ecosystem (Nvidia's position, supported by AI Czar David Sacks)
- The administration is also evaluating whether to allow DeepSeek to purchase Nvidia's H200 chips — a decision not yet taken as of the time of reporting
Static Topic Bridges
US Semiconductor Export Control Framework — Export Administration Regulations and the Entity List
The US regulates the export of dual-use technologies — items with both commercial and military applications — through the Export Administration Regulations (EAR), administered by the Bureau of Industry and Security (BIS) within the Department of Commerce. The Export Control Reform Act of 2018 (ECRA) strengthened BIS's authority, particularly over "emerging and foundational technologies."
The Entity List is BIS's primary enforcement tool: companies or individuals placed on it require a special licence from BIS before US firms can supply them with controlled items. Advanced AI chips (Nvidia A100, H100, H200, Blackwell series) are controlled items under EAR — their export to China requires a BIS licence, which is typically denied.
- Legal basis for semiconductor export controls: Export Administration Regulations (EAR); Export Control Reform Act, 2018 (ECRA)
- Administering body: Bureau of Industry and Security (BIS), US Department of Commerce
- Entity List: Maintained by BIS; listed entities require special licences (generally denied policy applies for Chinese advanced tech firms)
- Key Chinese entities on Entity List related to semiconductors: Huawei (2019), YMTC, Naura Technology Group, SMIC (partial restrictions), and numerous others
- October 2022 controls: BIS imposed comprehensive controls on advanced computing chips and semiconductor manufacturing equipment for China — covered chips above certain performance thresholds (e.g., A100-class)
- December 2024 update: BIS added 24 types of semiconductor manufacturing equipment and 3 software tools; 140 additional Chinese entities added to Entity List
- January 2026 rule: BIS revised export review policy for advanced AI chips to China — case-by-case evaluation with rigorous supply chain and security conditions
- Nvidia Blackwell chips: Next-generation AI accelerators (GB200, GB100 series) — subject to export controls to China
Connection to this news: DeepSeek's alleged use of Blackwell chips — despite their being controlled items — represents a potential violation of US export controls, highlighting enforcement gaps and the difficulty of preventing diversion once chips are physically in China.
AI Model Training — Large Language Models and Distillation
Large Language Models (LLMs) are AI systems trained on vast text datasets using transformer architecture. Training LLMs requires massive computational resources — thousands of high-performance GPUs (Graphics Processing Units) running for weeks or months. Nvidia's AI accelerators (A100, H100, H200, Blackwell series) are specifically designed for this parallel computing workload.
"Distillation" in AI refers to knowledge distillation — a technique where a smaller or newer "student" model is trained by learning from the outputs of a larger, more powerful "teacher" model. The teacher model's predictions (not its weights or architecture) guide the student's training, effectively transferring knowledge. This allows a newer model to achieve high performance even with limited training data.
- Transformer architecture: Foundation of modern LLMs; introduced in the 2017 "Attention is All You Need" paper (Google Research)
- GPU vs CPU for AI training: GPUs can perform many operations in parallel (thousands of cores) vs CPUs (few cores, sequential); this parallelism is essential for matrix multiplication in neural network training
- Nvidia's dominance: Nvidia controls approximately 70–80% of the AI accelerator market [Unverified — approximate market estimate]
- Knowledge distillation: First formalised by Geoffrey Hinton et al. (2015); allows transfer of a "teacher" model's generalised knowledge to a "student" model
- DeepSeek's distillation sources: Reportedly included outputs from Anthropic (Claude), Google (Gemini), OpenAI (GPT series), and xAI models
- Computational advantage of distillation: Reduces training compute requirements significantly — key to DeepSeek's claimed cost efficiency
Connection to this news: DeepSeek's approach — combining high-end Blackwell chips with distillation from US AI models — represents a sophisticated strategy to overcome both hardware and data limitations imposed by US controls.
Geopolitics of AI Chip Supremacy — Technology as Strategic Competition
Advanced semiconductor chips have become a key dimension of US-China strategic competition. The US has pursued a strategy of "chokepoint" control — maintaining dominance over the most advanced chip design and manufacturing capabilities, using export controls to deny China access to leading-edge AI accelerators and semiconductor manufacturing equipment (especially extreme ultraviolet lithography machines from ASML of the Netherlands).
China's response has been to accelerate domestic semiconductor development, with Huawei's Ascend AI chips and SMIC's manufacturing advances representing its primary counter-measures. However, China's chip capabilities remain approximately 2–3 generations behind the global frontier, making access to Nvidia chips highly valuable.
- ASML (Netherlands): Only manufacturer of Extreme Ultraviolet (EUV) lithography machines — essential for chips below 7nm; Netherlands has restricted ASML exports to China under US pressure
- EUV vs DUV lithography: EUV uses 13.5nm wavelength light for sub-7nm chips; DUV (Deep Ultraviolet) enables only older nodes
- Huawei Ascend chips: China's primary domestic AI accelerator; Ascend 910B is China's best currently; roughly equivalent to Nvidia A100 performance range
- SMIC (Semiconductor Manufacturing International Corporation): China's largest chip foundry; placed on US Entity List in 2020; reportedly achieving 7nm nodes using multi-patterning DUV
- "Small yard, high fence" doctrine: US strategy of tightly controlling a narrow set of the most critical technologies rather than broad restrictions
- India's semiconductor policy: India Semiconductor Mission (2021); ₹76,000 crore incentive scheme for semiconductor fabs and ATMP facilities; Tata Electronics-Powerchip JV fab in Gujarat approved (2024)
Connection to this news: The DeepSeek-Blackwell episode demonstrates the limits of hardware-only controls: even with export restrictions, chip diversion and distillation techniques allow Chinese AI firms to partially circumvent the technology barriers.
Key Facts & Data
- Chip allegedly used: Nvidia Blackwell series (GB200/GB100 — most advanced AI accelerators as of 2026)
- Alleged location of chips: DeepSeek's data centre in Inner Mongolia, China
- Training technique: Knowledge distillation from US AI models (Anthropic, Google, OpenAI, xAI)
- Administering body for US export controls: Bureau of Industry and Security (BIS), Department of Commerce
- Statutory basis: Export Administration Regulations (EAR); Export Control Reform Act, 2018
- US Supreme Court ruling (Feb 2026): Separate matter — struck down IEEPA-based tariffs (does not affect BIS export controls, which operate under different statutory authority)
- Huawei Ascend 910B: China's best domestic AI chip; roughly A100-equivalent range
- India Semiconductor Mission: ₹76,000 crore incentive scheme; launched 2021; first approved fab: Tata-Powerchip JV in Dholera, Gujarat
- Nvidia's market share in AI accelerators: approximately 70–80% [Unverified]