Beyond the Model: Reflections on the AWS Summit China
This summer, I attended the AWS Summit China at the Shanghai World Expo Center. The scale of the event was striking: activities, exhibitions, and talks occupied six floors, and “agents” or “agentic AI” appeared almost everywhere. Yet what interested me most was not any particular product demonstration. It was the contrast between how academia and industry currently imagine the future of AI.
In academic discussions, attention often gravitates toward foundation models themselves: their capabilities, architectures, benchmarks, and limitations. At the summit, however, the foundation model was rarely the main character. It was treated more like infrastructure—a powerful but increasingly standardized layer beneath the actual business problem. Most conversations focused instead on how agents could be integrated into existing management systems and operational workflows to improve efficiency, reduce costs, and increase productivity.
The audience also shaped the atmosphere. Many attendees appeared to represent companies evaluating technologies, platforms, or potential business partnerships. Consequently, nearly every presentation was grounded in questions of deployment and commercialization: Can this system be connected to existing enterprise software? Is it reliable enough for daily operations? Can its value be measured? Who is accountable when it fails?
This perspective was quite different from what I usually encounter in academic research. In a paper, a modest improvement on a benchmark can be meaningful. In an enterprise, however, technical performance is only one part of a much larger equation. A system must also fit organizational structures, comply with policies, work with legacy infrastructure, and justify the cost of changing established processes. The most advanced model is not necessarily the most useful one; often, the more important innovation lies in designing the surrounding system well.
The summit made me reconsider what “progress” in AI actually means. Scientific breakthroughs remain essential, but their social and economic influence depends on a less visible layer of engineering, product design, and organizational adaptation. Agentic AI will not transform companies simply because models become more capable. It will do so only when autonomy can be made dependable, governable, and compatible with human institutions.
For me, the summit offered a valuable view of AI from the demand side. It showed that the distance between a convincing demonstration and a genuinely useful system is still substantial—and that bridging this distance may be one of the defining challenges of the agent era.