Whalesbook
Chip Bottleneck Stymies China's AI Ambitions
Elite Chinese AI researchers are expressing a more pessimistic outlook on the nation's ability to catch up with the U.S. Artificial Intelligence capabilities. The primary hurdle, they state, is a critical bottleneck in advanced semiconductor chips, exacerbated by U.S. export controls.
Tang Jie, founder of Chinese AI startup Zhipu, stated at a recent Beijing conference that the gap with the U.S. is likely widening. "While we're doing well in certain areas, we must still acknowledge the challenges and the disparities we face," he remarked. This sentiment is amplified by the fact that when AI chip leader Nvidia unveiled its next-generation Rubin hardware, no Chinese AI developers were named as customers due to U.S. restrictions.
Workarounds and Resource Constraints
Chinese companies are exploring circuitous routes to access these vital chips, including discussions about renting computing power in data centers across Southeast Asia and the Middle East. These arrangements, following similar efforts for Nvidia's Blackwell series chips last year, are legally permissible but introduce significant inconvenience and yield fewer chips compared to their well-funded American counterparts.
Justin Lin, who leads AI model development for Alibaba's Qwen, projected a mere 20% or less chance for any Chinese company to surpass giants like OpenAI and Anthropic within the next three to five years. Washington's stringent export controls deter many firms from pursuing state-of-the-art AI, which is highly compute-intensive. Instead, Chinese companies focus on applying AI for everyday uses.
Divergent Investment and Innovation Paths
American firms are investing heavily in the latest chips to push technological boundaries, while Chinese companies are "stretched thin," with meeting current delivery demands consuming most resources, according to Lin. UBS analysts estimate the combined capital spending of China's internet leaders on AI was around $57 billion last year, a stark contrast to roughly ten times that amount for their U.S. peers.
Despite these challenges, innovation persists. Developers like DeepSeek are demonstrating skill in adapting to resource limitations. Zhipu, formally known as Knowledge Atlas Technology, and MiniMax recently raised over $1 billion combined in Hong Kong IPOs, signaling investor confidence in the potential for technological catch-up. MiniMax's shares more than doubled in just two trading days post-IPO.
DeepSeek, in particular, has published techniques to improve AI development efficiency, including papers on new training architectures for larger models using fewer chips and more efficient memory designs. While models from DeepSeek and Alibaba are closing the gap, the reliance on less-advanced domestic or imported chips, manufactured using limited technology, remains a fundamental constraint for China's AI sector.