Xi Jinping's keynote and China's AI governance vision

President Xi Jinping delivered a speech titled "Joining Hands to Build a Just and Reasonable Global AI Governance System," in which he laid out four observations: adhere to openness and win-win cooperation; strengthen risk-awareness and ensure safety; uphold the principle of developing AI for good; and jointly build global AI governance frameworks . Xi emphasized a "people-centered" approach and positioned China as a champion of inclusive, multilateral AI governance—a direct counterpoint to U.S.-led models. He pledged to share technology with developing countries and called for collective rather than unilateral AI rule-setting . The address framed AI as a force for shared prosperity and common security, not merely national advantage .

Kimi K3: The largest open-source AI model ever

Chinese AI startup Moonshot AI launched Kimi K3, a 2.8-trillion-parameter open-weight model, making it the largest open-source AI model in the world at the time . Key details:

• Performance: Benchmarked against Anthropic's Claude Fable 5 and OpenAI's GPT 5.6 Sol, Moonshot acknowledged that while K3 still trails these models on overall performance, it outperforms them on several specific benchmarks, particularly long-context reasoning and coding tasks . Independent reports note it consistently outperformed other tested models, including Claude Opus 4.8 and GPT 5.5 .
• Specs: Features a 1-million-token context window and native vision capabilities, built on Moonshot's proprietary Delta Attention architecture . It activates just 16 of its 896 experts per token .
• Availability: Released as open-weight on July 16, with full weights promised by July 27, signaling China's push for open-source leadership as U.S. frontier models remain largely closed . API pricing was set at $3 per million input tokens and $15 per million output tokens .

AI for social good: The framing shift

A recurring theme across Xi's speech and the conference overall was that AI development should prioritize social welfare and human benefit over national security or corporate profit alone . The conference chair statement emphasized that AI should be used to address global challenges such as poverty, healthcare access, and education inequality . This positioning—"AI for good" as a core principle—served both as a diplomatic value proposition and as a subtle critique of what China frames as the U.S.'s more security-driven, competitive approach .

Hardware showcase: Huawei's Atlas 950 SuperPoD

Huawei staged the most dramatic hardware reveal of the conference: the Atlas 950 SuperPoD, an ultra-large AI supernode system built entirely without U.S. components .

• Scale: The show-floor unit connected 1,024 Ascend NPUs across 16 cabinets; the full configuration scales to 8,192 Ascend 950DT processors . The demo unit was rated at 1 EFLOPS at FP8 precision and 2 EFLOPS at FP4 .
• Performance claim: Huawei says the full system delivers 6.7 times the computing power and 15 times the memory capacity of Nvidia's NVL144 (built on Nvidia's Rubin architecture) . The system uses Huawei's proprietary Lingqu protocol for high-speed interconnection .
• Status: A working physical unit appeared on the show floor for the first time—previously the system had existed only as roadmap slides—and is slated for commercial availability in Q4 2026 .
• Significance: The Atlas 950 is designed to directly compete with Nvidia's rack-scale offerings while proving China's ability to build advanced AI infrastructure under U.S. export controls .

Exhibition scale: 1,100+ companies

The conference floor featured over 1,100 exhibitors showcasing a wide range of AI applications :

• Humanoid robots from dozens of Chinese and international firms demonstrating advances in dexterity, locomotion, and human-robot interaction.
• Autonomous vehicles—including robotaxis, delivery bots, and autonomous shuttles—displayed live or in simulated environments.
• AI chips from Huawei (Ascend 950 series), Cambricon, and a host of domestic startups competing in the post-sanctions semiconductor ecosystem.
• AI glasses from multiple manufacturers, reflecting the rapid rise of wearable AI assistants in China's consumer market.
• Medical AI, smart manufacturing, and education AI were also major exhibition themes, consistent with the "AI for good" framing .

The conference drew high-level participation from dozens of countries and multiple international organizations, and concluded with a chair's statement calling for continued multilateral dialogue on AI safety standards and equitable access .

Xi Jinping's keynote and China's AI governance vision

President Xi Jinping delivered a speech titled "Joining Hands to Build a Just and Reasonable Global AI Governance System," in which he laid out four observations: adhere to openness and win-win cooperation; strengthen risk-awareness and ensure safety; uphold the principle of developing AI for good; and jointly build global AI governance frameworks . Xi emphasized a "people-centered" approach and positioned China as a champion of inclusive, multilateral AI governance—a direct counterpoint to U.S.-led models. He pledged to share technology with developing countries and called for collective rather than unilateral AI rule-setting . The address framed AI as a force for shared prosperity and common security, not merely national advantage .

Kimi K3: The largest open-source AI model ever

Chinese AI startup Moonshot AI launched Kimi K3, a 2.8-trillion-parameter open-weight model, making it the largest open-source AI model in the world at the time . Key details:

• Performance: Benchmarked against Anthropic's Claude Fable 5 and OpenAI's GPT 5.6 Sol, Moonshot acknowledged that while K3 still trails these models on overall performance, it outperforms them on several specific benchmarks, particularly long-context reasoning and coding tasks . Independent reports note it consistently outperformed other tested models, including Claude Opus 4.8 and GPT 5.5 .
• Specs: Features a 1-million-token context window and native vision capabilities, built on Moonshot's proprietary Delta Attention architecture . It activates just 16 of its 896 experts per token .
• Availability: Released as open-weight on July 16, with full weights promised by July 27, signaling China's push for open-source leadership as U.S. frontier models remain largely closed . API pricing was set at $3 per million input tokens and $15 per million output tokens .

AI for social good: The framing shift

A recurring theme across Xi's speech and the conference overall was that AI development should prioritize social welfare and human benefit over national security or corporate profit alone . The conference chair statement emphasized that AI should be used to address global challenges such as poverty, healthcare access, and education inequality . This positioning—"AI for good" as a core principle—served both as a diplomatic value proposition and as a subtle critique of what China frames as the U.S.'s more security-driven, competitive approach .

Hardware showcase: Huawei's Atlas 950 SuperPoD

Huawei staged the most dramatic hardware reveal of the conference: the Atlas 950 SuperPoD, an ultra-large AI supernode system built entirely without U.S. components .

• Scale: The show-floor unit connected 1,024 Ascend NPUs across 16 cabinets; the full configuration scales to 8,192 Ascend 950DT processors . The demo unit was rated at 1 EFLOPS at FP8 precision and 2 EFLOPS at FP4 .
• Performance claim: Huawei says the full system delivers 6.7 times the computing power and 15 times the memory capacity of Nvidia's NVL144 (built on Nvidia's Rubin architecture) . The system uses Huawei's proprietary Lingqu protocol for high-speed interconnection .
• Status: A working physical unit appeared on the show floor for the first time—previously the system had existed only as roadmap slides—and is slated for commercial availability in Q4 2026 .
• Significance: The Atlas 950 is designed to directly compete with Nvidia's rack-scale offerings while proving China's ability to build advanced AI infrastructure under U.S. export controls .

Exhibition scale: 1,100+ companies

The conference floor featured over 1,100 exhibitors showcasing a wide range of AI applications :

• Humanoid robots from dozens of Chinese and international firms demonstrating advances in dexterity, locomotion, and human-robot interaction.
• Autonomous vehicles—including robotaxis, delivery bots, and autonomous shuttles—displayed live or in simulated environments.
• AI chips from Huawei (Ascend 950 series), Cambricon, and a host of domestic startups competing in the post-sanctions semiconductor ecosystem.
• AI glasses from multiple manufacturers, reflecting the rapid rise of wearable AI assistants in China's consumer market.
• Medical AI, smart manufacturing, and education AI were also major exhibition themes, consistent with the "AI for good" framing .

The conference drew high-level participation from dozens of countries and multiple international organizations, and concluded with a chair's statement calling for continued multilateral dialogue on AI safety standards and equitable access .