On July 28, the MoneyFrontier 2026 New Infrastructure Intelligent Computing Summit was held at the Hopewell Hotel in Hong Kong, bringing together speakers from Moore Threads, Tencent Cloud, Yanji Microelectronics and Xinke Liquid Cooling to discuss domestic GPU ecosystems, enterprise AI agents, the migration of mining know-how into AI compute, and the upgrade path from conventional facilities to AIDC centers.
The central message from the event was straightforward: the AI compute story is shifting away from who controls the largest number of cards and toward who can turn compute into actual productivity. Across the morning sessions, speakers pointed to the same set of changes — token consumption is rising quickly, agents are moving beyond chat interfaces into task execution environments, capabilities built in the mining industry are beginning to transfer into AI infrastructure, and the move from PoW machine rooms to AIDC is, at its core, a rebuild of the engineering stack.
Moore Threads says domestic GPUs are entering the token era
Zhao Zhenyang, general manager of the solutions division at Moore Threads, delivered a presentation titled “The Token Era, Intelligence Everywhere,” focusing on token demand growth, the industrial path for domestic GPUs and the compute infrastructure required in the age of intelligent agents.
Zhao said AI is moving from visual AI and generative AI toward Agentic AI and Physical AI. As agents repeatedly call models, tools and other agents while carrying out tasks, token consumption is expanding from a pattern of humans using AI to one in which AI increasingly uses AI. In that setting, compute demand is no longer centered only on single-chip performance. It now depends on system-level support for training, inference, simulation, agent operation and terminal applications.
Against that backdrop, Moore Threads introduced three frameworks: a training factory, a token production factory and an agent factory, corresponding to model training, token production and agent operation. The presentation argued that if domestic GPUs are to enter real industrial settings, the challenge is no longer simple hardware substitution. What is needed is a closed loop that covers chips, software stacks, training platforms, inference frameworks, simulation capability and application adaptation.
Tencent Cloud sees agents moving beyond chat into execution
Cai Peng, chief architect for Tencent Cloud’s Yunnan region, spoke on “The Agent Wave: From Technical Breakout to a Full Reshaping of a New Industrial Era,” with a focus on AI agent evolution and enterprise deployment.
Cai said AI agents are moving past the large-model application stage defined by question-and-answer interactions. The next phase, he said, is one in which agents can understand goals, break down tasks, formulate plans, call tools and keep responding through feedback loops. Once agents enter industrial settings, the decisive factors extend beyond model capability to include memory mechanisms, knowledge management, tool invocation, skill accumulation, multi-agent collaboration and secure execution environments.
On the enterprise side, Tencent Cloud presented WorkBuddy, CodeBuddy, the ADP agent development platform, a training and inference platform, and AI Sandbox. The set of products and platforms presented at the summit suggested that enterprise agents are moving from personal productivity assistants toward task execution, organizational collaboration and changes to production workflows.
Yang Zuoxing says the larger AI compute opening is on the inference side
Yang Zuoxing, founder of Yanji Microelectronics, delivered a presentation titled “From Bitcoin Compute to AI Compute,” covering cooling routes, energy choices, AI compute opportunities and the future of Bitcoin mining.
Yang said that for high-density compute, cooling and energy are no longer supporting considerations. In his view, they are core variables that shape a project’s long-term competitiveness. After comparing air cooling, oil cooling and water cooling, he said water cooling is better suited to future new-build and retrofit projects. On the energy side, he said natural gas combines stability with dispatch flexibility and is an incremental option worth watching in the near term.
On AI compute, Yang said the training market has already become concentrated, while real incremental opportunities are more likely to appear in inference. As AI applications, agents and robots spread, inference demand is expected to move outward toward regional, edge and industry nodes. He also said open-source models, distributed deployment and domestic chips will create room for new participants.
When speaking about Bitcoin mining, Yang said the industry’s golden era has passed. Even so, the mining sector’s experience in low-power chips, energy-efficiency optimization and energy operations still retains value. In his view, mining machines, natural gas, solar-plus-storage and AI data centers may eventually form new synergies.
Xinke Liquid Cooling says AIDC conversion starts with engineering upgrades
Wang Chong, CEO of Xinke Liquid Cooling, gave a presentation titled “The Iteration of the Compute Wave: Transformation Opportunities and Real-World Challenges in the Data Center Industry,” focusing on the practical issues involved in upgrading traditional data centers and blockchain machine rooms into AIDC intelligent computing centers.
Wang said turning a traditional data center into an AIDC facility is not a matter of simply swapping in GPU servers. It requires a broader rebuild across power distribution, liquid cooling, networking, fire protection, load-bearing capacity, operations platforms and delivery systems. High-density GPU racks are quickly moving into the range of tens of kilowatts and even 100-kilowatt-class deployments, raising the bar for power redundancy, liquid-cooling control, fire-safety compliance and continuous operations.
He added that existing data centers still have reuse value. Land, factory buildings, power access and primary-side heat dissipation facilities can all be retained. But whether a site can complete the transition depends on filling gaps in secondary-side liquid cooling, dual-path power supply, backup power, gas-based fire protection, network latency control, BCIM operations platforms and standardized quality verification capability. For the traditional compute industry, Wang said, the first step in an AIDC shift is an upgrade in engineering capability.
Four presentations, one conclusion
The four sessions approached the market from different angles, but they converged on the same judgment: AI compute is entering a stage where system engineering will decide outcomes.
Moore Threads focused on how domestic GPUs move from merely being able to run workloads to becoming practical across the full stack. Tencent Cloud focused on how agents move from experimental tools to enterprise production systems. Yang Zuoxing focused on transferring a decade of mining-sector energy and engineering experience into inference compute. Xinke Liquid Cooling focused on shifting data centers from a simple “power on and run” model to enterprise-grade standards built around zero interruption and verifiability.
Taken together, the talks suggested that competition in AI compute is no longer confined to chips or hardware on their own. The contest is now taking shape across the combined system of training, inference, energy, cooling, operations and delivery.