AI Chips

NVIDIA’s H20 AI chip, designed for the Chinese market, faced export restrictions but received conditional clearance in mid-July, requiring a 15% revenue remit to the U.S. government. Despite this, H20’s future in China is uncertain. NVIDIA defends its H20 exports, emphasizing the U.S.’s need to compete with China in AI and avoid repeating the 5G situation where it lost its leadership position. They believe U.S. AI can become the global standard if the U.S. actively participates in the competition.

The U.S. maintains a strict stance on AI chip exports to China, particularly regarding advanced technology like Nvidia’s “Blackwell” chip. Former President Trump opposes its sale to China, citing national security concerns and the potential for accelerating Chinese AI development. He believes the U.S. holds a significant technological advantage in chip design. Nvidia’s CEO is expected to discuss export licenses, highlighting ongoing negotiations. While some argue restricting access could spur indigenous Chinese chip development, the U.S. aims to limit China’s access to cutting-edge AI capabilities.

Reports suggest NVIDIA has secured a deal with Samsung for its 12-layer stacked HBM3E memory, potentially for use in liquid-cooled server systems. The initial order is rumored to be 30,000-50,000 units. Samsung has neither confirmed nor denied the agreement. While previous validation claims proved premature, Wall Street analysts are optimistic that Samsung will complete certification by the end of August and begin shipments to NVIDIA in Q4. Samsung aims to reduce HBM3E production costs and expects increased HBM profitability.

Tesla is reportedly shutting down its Dojo project, a move indicating a shift towards external partnerships for its AI needs. Head of Dojo, Peter Bannon, is expected to leave. Tesla plans to enhance collaborations with Nvidia and AMD for computing, and deepen its partnership with Samsung in chip manufacturing, evidenced by a $16.5 billion deal. Samsung will reportedly produce Tesla’s AI6 chip, while TSMC will manufacture the AI5. Tesla also plans to diversify Dojo’s supply chain, engaging Samsung and Intel for chip manufacturing and packaging.

Tesla has reportedly signed a $16.5 billion deal with Samsung for AI chip manufacturing, intended for electric vehicles and robotics. The agreement focuses on the AI6 chip, potentially utilizing a 2nm process with production in Austin, Texas, starting in 2025. Despite this, Morgan Stanley analysts believe the deal will only minimally impact TSMC’s revenue, estimating a 1% reduction, as TSMC remains a key supplier for Tesla and xAI. TSMC is set to launch its 3nm A15 chip in 2026, while Samsung’s A16 isn’t expected until 2027.

In 2018, Nvidia’s Jensen Huang proposed a collaboration with Samsung encompassing HBM, advanced process node development, and CUDA advancement. Reportedly rejected due to Samsung’s internal challenges, this decision allowed Nvidia to partner with SK Hynix for HBM, a lucrative alliance that saw SK Hynix’s share price surge. Meanwhile, Samsung now lags in HBM production and the AI chip foundry market, dominated by TSMC for Nvidia’s chips. Samsung is now trying to catch up, aiming for HBM4 contracts.

Zhou Hongyi, chairman of 360 Group, views Nvidia’s H20 chip for China as a sign of ongoing US-China AI competition. Despite Nvidia’s 30-year lead, Zhou and Nvidia CEO Jensen Huang acknowledge Huawei’s remarkable progress in GPU development. Huang highlights Nvidia’s AI ecosystem investments, while Huawei’s Ren Zhengfei emphasizes using methods like stacking and clustering to compensate for single-chip limitations, focusing on mid-to-low-end chips and software advancements.

Nvidia CEO Jensen Huang visited China, praising its AI development and hinting at a shift in Nvidia’s strategy amid US chip restrictions. While downplaying GPU criticality, Nvidia aims to promote its new China-specific B30 chip and robotic technologies. Huang’s remarks suggest a willingness to tolerate CUDA compatibility from competitors, reflecting concerns about losing market share to domestic AI chip makers like Huawei and Cambricon, and a desire to maintain influence through its CUDA ecosystem, even with potential hardware revenue losses.

NVIDIA’s H20 series AI chips have U.S. approval for reintroduction to China alongside a new RTX Pro GPU for the Chinese market. Morgan Stanley sees this as a positive catalyst for BAT, anticipating increased capital expenditure for AI, cloud services, and e-commerce enhancements. Chinese firms like Tencent and ByteDance are placing orders, requiring U.S. government approval. The move aims to counter Huawei’s dominance in the Chinese market by providing an alternative, potentially limiting Huawei’s global competitiveness despite its technological advancements.

Malaysia is implementing a licensing regime for U.S.-made high-performance AI chip exports and transshipments to prevent diversion, especially to China. Effective immediately, entities suspecting misuse must notify authorities 30 days prior to export and obtain permits. This addresses regulatory loopholes, with plans to add AI chips to the Strategic Items List. This follows U.S. efforts to tighten export controls on advanced NVIDIA GPUs to Malaysia and Thailand, requiring licenses to prevent re-export to China.