On August 26, the 2026 AGIC Shenzhen International General Artificial Intelligence Industry Expo and the GERX Global Embodied Intelligent Robot Industry Expo opened in Shenzhen, drawing over 7,000 overseas buyers and thousands of enterprises to the 81,000-square-meter venue. From large language models and compute infrastructure to robots and intelligent agent applications, every segment of the AI industry chain was on display—where technology becomes a product that can be quoted, ordered, and delivered at scale.
A clear shift at the expo was the computing power narrative moving from “larger clusters” to “smarter scheduling.” Multiple cloud vendors presented end-to-end heterogeneous intelligent computing solutions, all pointing to the same conclusion: agent workloads are too mixed for a single GPU-centric architecture. The focus of compute supply is no longer just “how much computing power” but “how smoothly and cost-efficiently it can be delivered.”
Behind this focus on heterogeneous computing is a profound change in AI workload structure. A single agent operation often involves multiple stages—reasoning after receiving a command, calling external tools, and executing multiple instances concurrently. Each stage has very different compute demands, and a GPU-only architecture is rarely the most efficient match at every point. By coordinating GPUs, NPUs, and dedicated AI accelerators, computational resources can be dynamically matched to workload characteristics, ensuring stable agent response while avoiding idle compute waste.
The expo made the compute fluctuation caused by agent workloads highly visible. Office agents writing weekly reports and processing spreadsheets, robot control modules making real-time decisions, and agent clusters experiencing sudden compute spikes during multi-instance concurrency all highlighted why elastic resource reuse is critical. For enterprises deploying Agent projects, one of the biggest reasons compute costs remain high is the inability to reuse resources flexibly when loads fluctuate. Heterogeneous co-scheduling directly addresses this mismatch between fluctuating usage and inefficient allocation.
Unified scheduling across heterogeneous resources is becoming an unavoidable step for large-scale agent deployment. StarWar Cloud, focused on GPU compute platforms and compute scheduling, is advancing unified management and elastic orchestration of heterogeneous resources. This enables model training, inference, and hands-on training tasks to be flexibly scheduled across diverse computing resources, adapting to the volatility of multi-agent concurrency and enabling elastic resource reuse. The result is a practical path for enterprises to lower compute costs in Agent projects.
The expo also revealed another key trend in AI deployment: AI office tools, embodied intelligence, and agent applications are appearing at the same time, with technology suppliers and demand-side enterprises directly connected on the show floor. The maturity of the supply chain has become a decisive factor in how quickly AI can be deployed at scale. Shenzhen and the wider Greater Bay Area combine mature hardware manufacturing, open-source ecosystems, and compute infrastructure into one complete chain—and heterogeneous intelligent computing solutions are being delivered through exactly this kind of full-stack supply chain.