1. The World AI Conference (WAIC) in Shanghai opened with the buzz of a trade fair at full tilt and the unease of an industry under pressure to show it can turn technical advances into durable businesses. Over four days in mid-July, it recorded over 400,000 visits, 1,117 companies displaying 4,486 exhibits, and 351 product debuts—the highest metrics since the conference began in 2018. [para. 1][para. 2]
2. Beneath the crowds gathered around dancing robots and generative AI demos, a more pragmatic shift was taking shape. The industry has moved quickly from asking what the technology can do to confronting the harder questions of how to deploy it at scale and achieve profitability across computing infrastructure, foundation models, and enterprise software. These concerns over demand and market volatility were amplified by a sharp pullback in global tech stocks just days before the conference, which revived talk of an AI bubble. [para. 3][para. 6]
3. Chinese startup Moonshot AI set the early tone of the conference by unveiling Kimi K3, an open-source foundation model with 2.8 trillion parameters. According to a testing platform, the model ranked third globally in overall performance, behind Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol, pushing a Chinese company into the global top tier despite tight domestic computing constraints. [para. 4][para. 5]
4. The computing demands required to support such models were visible on the exhibition floor in the form of "supernodes." Huawei debuted its commercial Atlas 950 SuperPoD, connecting as many as 1,024 chips across 20 cabinets, while rivals Alibaba, Moore Threads, and Enflame presented competing systems. A single system often costs more than 100 million yuan, and a single hardware failure can disrupt the whole architecture. [para. 9][para. 10][para. 11][para. 12][para. 13]
5. The financial incentives behind this massive computing buildout are vast. Venture investor Zhou Zhifeng reported that U.S. technology giants have revised their 2026 AI computing budgets to more than $800 billion, while China’s comparable market is about $100 billion. Crucially, inference demand has caught up with training demand, with ByteDance already at a one-to-one balance. [para. 14][para. 15]
6. Beyond the computing displays, embodied AI was the conference’s most conspicuous visual. The robotics section expanded sharply to over 200 exhibitors from companies like Unitree and Deep Robotics showing robots boxing, dancing, and sorting medicine. The emphasis shifted from agility to proving real-world utility in manufacturing and logistics, though developers conceded that learning and generalization abilities remain constrained. [para. 16][para. 17][para. 18][para. 19]
7. In enterprise software, the foundation-model market narrowed rapidly at WAIC. Of the more than 100 large-model exhibitors, only about a dozen still described foundation-model development as their core business; most pivoted to AI agents for enterprise clients. Tencent, Alibaba, and Baidu showcased their platforms. Enterprise demand has at least doubled, but integration with legacy corporate databases and workflows remains the bottleneck. Kai-Fu Lee warned that corporate AI adoption must be CEO-driven or companies may not survive three years. [para. 22][para. 23][para. 24][para. 25][para. 26]
8. Consumer AI-native hardware, including smartphones and smart glasses, showed a clear gap between promise and performance. Executing a multi-step command, a smartphone responded slowly and was disrupted by an ad. Smart glasses focusing on narrower functions appeared more grounded, but Li Yannan of HSG concluded they must still solve a genuine user need rather than exist for AI’s sake. [para. 27][para. 28][para. 29][para. 30][para. 31]
9. On the governance front, President Xi Jinping called for a human-centered approach to AI. The World AI Cooperation Organization was formally established with 29 founding countries, including China, Russia, Brazil, and Indonesia. The diplomatic signal was clear: China is seeking a larger role in shaping the global rules governing the technology. [para. 33][para. 34][para. 35]
AI generated, for reference only
Sign in with Google
Sign in with Facebook
Sign in with 财新