CNBC AI News May 18 — At the 2025 Sohu Tech Summit, Professor Zheng Weimin, a Chinese Academy of Engineering academician and Tsinghua University computer science professor, delivered a keynote titled ‘Infrastructure Development and Applied Frontiers of Large AI Models.’
Zheng outlined two pivotal trends shaping artificial intelligence in 2025: first, the explosion of multimodal capabilities where large models now process text, images and video simultaneously. Second, China’s strategic push to deploy AI in GDP-critical sectors — manufacturing, finance and healthcare — where he emphasized the nation’s implementation advantages over theoretical innovation.
The industry’s current reliance on large model training infrastructure creates a paradox, the expert noted. While NVIDIA’s GPU ecosystem remains unmatched for developer tools and mature architecture, export restrictions and unregulated secondary markets have transformed cutting-edge chips into commodities more valuable than gold in certain segments, with recurring supply chain bottlenecks disrupting projects nationwide.
In parallel, over 30,000 domestic hardware developers are aggressively refining local alternatives, achieving meaningful progress in core performance metrics. But this David and Goliath struggle faces reality checks, Zheng warned, as fragmented software compatibility issues continue creating roadblocks for mainstream adoption.
Turning to solutions, Zheng proposed a two-pronged approach: cultivating a ‘pseudo-CUDA’ development environment that minimizes retraining costs through familiar interfaces, while maintaining relentless focus on hardware benchmarks even with late-stage entry challenges.
“We shouldn’t wait until we reach 100% parity,” the academician argued, drawing applause from tech executives present. “Ironically, reaching 60% of international standards with superior localized service could actually trigger natural adoption depending on specific application scenarios. Reaching 70-80% would make migration almost inevitable.”
This pragmatic framework suggests China’s semiconductor industry may bypass traditional performance rat races through targeted interoperability enhancements, particularly in vision and speech processing where industry-specific optimizations deliver disproportionate commercial value. As global tech policy hardens against unrestricted cross-border data flows, Zheng’s calculus balances engineering aspirations with market realities on the cutting edge of AI-infrastructure development.
Original article, Author: Tobias. If you wish to reprint this article, please indicate the source:https://aicnbc.com/337.html
