What Happened
- Chinese AI startup DeepSeek denied Nvidia and AMD early access to its forthcoming V4 AI model, giving Chinese chipmakers (including Huawei) a head start of several weeks in optimising their chips for the new model.
- This is a departure from standard industry practice, in which AI developers provide pre-release model versions to major chipmakers so their hardware and software run smoothly on widely adopted chips.
- Separately, a senior Trump administration official alleged that DeepSeek's latest model was trained on Nvidia's Blackwell chip (its most advanced AI chip) inside mainland China — likely in violation of US export controls, with the chips believed to be clustered at a data centre in Inner Mongolia.
- The US believes DeepSeek will remove technical indicators that might reveal the use of banned American chips.
- Reports also indicate the model may have used "distillation" from leading US AI companies' models (Anthropic, Google, OpenAI, xAI) — a technique where a newer model learns from an older, established model's outputs.
- These developments arrive amid the intensifying US-China AI chip competition, as export controls tighten and China accelerates domestic chip development.
Static Topic Bridges
US AI Chip Export Controls: Regulatory Architecture
The US Bureau of Industry and Security (BIS) under the Department of Commerce administers export controls through the Export Administration Regulations (EAR). Since October 2022, the US has imposed sweeping restrictions on exporting advanced AI chips and semiconductor manufacturing equipment to China. Chips above certain performance thresholds (measured in total processing performance, or TPP) require export licences for China. Nvidia's Blackwell chips (B100, B200 series) — the current leading-edge AI training chips — are restricted from export to China. The BIS also maintains an Entity List of companies subject to enhanced licensing requirements. China's Huawei, YMTC, and several AI firms are on the Entity List. The Foreign Direct Product Rule (FDPR) extends US jurisdiction to chips made with US technology anywhere in the world.
- BIS Entity List: companies subject to export licence requirements for items subject to EAR
- October 2022 rules: restricted advanced AI chips (Nvidia H100, A100, then A800/H800) and chip-making equipment
- Blackwell chips (Nvidia B100/B200): the latest generation; restricted from China; allegedly used by DeepSeek
- Foreign Direct Product Rule (FDPR): extends US export control jurisdiction to chips made abroad using US technology
- Distillation technique: using outputs of a powerful model to train a newer model — a workaround for training without the most powerful chips
- US CHIPS Act (2022): $52.7 billion to build domestic semiconductor manufacturing, reducing dependence on Taiwan
Connection to this news: DeepSeek's alleged use of banned Blackwell chips illustrates the enforcement challenges of export controls — sophisticated actors find workarounds, and tracing chip provenance is technically difficult.
Artificial Intelligence: Strategic Technology and Governance
Artificial intelligence has become a core dimension of great power competition. The US AI advantage has historically rested on Nvidia's dominant GPU ecosystem, advanced cloud infrastructure (AWS, Azure, Google Cloud), and leading frontier model developers (OpenAI, Anthropic, Google DeepMind). China's DeepSeek emerged as a significant challenger in early 2025, achieving performance comparable to frontier US models at substantially lower training cost, partly through algorithmic innovation. India has its own AI Mission — IndiaAI, announced in 2024 — with ₹10,372 crore allocated for compute infrastructure, a foundational model, datasets, and application development. The National AI Strategy emphasises "AI for all" with focus on agriculture, healthcare, education, and governance applications.
- DeepSeek R1 (January 2025): matched GPT-4-level performance at dramatically lower training cost — ~$6 million vs hundreds of millions for comparable US models
- IndiaAI Mission (2024): ₹10,372 crore; compute infrastructure, foundational model, data governance
- Nvidia: controls ~80% of AI training chip market; H100 GPU is the standard AI training chip
- AI distillation: technique transferring "knowledge" from a large teacher model to a smaller student model
- Model weights: mathematical parameters of a trained AI model; their export is a key control challenge
- US-China AI competition: key battleground for technological and military advantage
Connection to this news: DeepSeek's withholding of chip access from US companies and alleged use of banned hardware represent China's strategy to compete in AI despite export controls — raising questions about the long-term effectiveness of technology denial strategies.
Technology Transfer, IP, and AI Governance Challenges
AI development raises novel intellectual property and technology transfer challenges. The "distillation" technique — using outputs from proprietary AI models (like ChatGPT) to train new models — sits in a legal grey zone. OpenAI and other US companies have alleged that DeepSeek used their model outputs without permission. At the governance level, there is no comprehensive international framework for AI governance, though efforts include the UN Advisory Body on AI, the Bletchley Declaration (November 2023, signed by 28 countries including India and China), and the Council of Europe's AI Convention. India has not yet enacted a comprehensive AI regulation, though the Digital India Act (in development) and the AI framework under the Ministry of Electronics and IT address aspects of AI governance.
- Distillation legality: no clear international law on using AI model outputs for training competing models
- Bletchley Declaration (November 2023): 28 countries acknowledged frontier AI safety risks; India and China signed
- EU AI Act (2024): first comprehensive AI regulation globally; risk-based classification
- India's AI governance: MeitY AI framework; Digital India Act under development; no standalone AI law yet
- US Executive Order on AI (October 2023): required safety reporting for frontier AI models; rescinded by Trump in 2025
- Key governance challenge: jurisdictional reach of AI regulations across borders
Connection to this news: DeepSeek's alleged model distillation from US AI companies' outputs highlights the governance gap — current laws inadequately address knowledge transfer through AI outputs, a challenge that will grow as AI capabilities advance.
Critical Minerals and the Semiconductor Supply Chain
The semiconductor supply chain depends on a complex web of critical minerals — rare earth elements, gallium, germanium, tungsten, silicon — many of which China controls at the processing stage. China mines approximately 60% of the world's rare earths and controls ~85% of the processing capacity. In December 2024, China imposed export controls on gallium and germanium (used in compound semiconductors for defence, satellite, and high-speed electronics) as retaliation for US chip controls. This creates a supply chain interdependence that complicates the US strategy of denying China advanced chips — China can retaliate by restricting minerals needed to make those same chips. India has significant rare earth deposits (6.9 million tonnes, 5th largest globally) but currently under-exploits them.
- China's rare earth dominance: ~60% of global mining, ~85% of global processing
- Gallium and germanium export controls (China, December 2024): used in compound semiconductors, satellites, defence
- US CHIPS Act response: $52.7 billion for domestic semiconductor manufacturing
- India's rare earth deposits: 6.9 million tonnes (5th largest globally); Indian Rare Earths Ltd (IREL) is primary state entity
- Critical Minerals Mission (India, 2023): identified 30 critical minerals; mission for exploration and processing
- Significance: whoever controls critical mineral processing controls the semiconductor supply chain bottleneck
Connection to this news: DeepSeek's alleged use of banned Nvidia chips highlights how export controls and critical mineral dominance are two sides of the same US-China technology competition — India's rare earth processing capacity positions it strategically in this contest.
Key Facts & Data
- DeepSeek V4: access denied to Nvidia and AMD for chip optimisation; Chinese chipmakers (Huawei) got weeks-long head start
- Alleged violation: DeepSeek trained on Nvidia Blackwell chips in mainland China — subject to US export controls
- Blackwell chip location: alleged to be clustered at a DeepSeek data centre in Inner Mongolia, China
- Distillation technique: used by DeepSeek — learning from outputs of US frontier models (OpenAI, Anthropic, Google, xAI)
- US export control regime: October 2022 restrictions on advanced AI chips; subsequent tightening rounds
- China's critical mineral export controls: gallium and germanium (December 2024); rare earths, tungsten expanded in 2025
- Nvidia's market share in AI training chips: ~80%
- DeepSeek R1 (January 2025): comparable performance to GPT-4 at ~$6 million training cost
- India's rare earth deposits: 6.9 million tonnes (5th largest globally); strategic diversification opportunity