Zhejiang University Holding white paper: AI investment logic shifts from parameters to deployment, model adaptability and low cost become core demands

The fifth Global Digital Trade Expo, themed “Meet the AI Future at the Digital Trade Expo,” was held in Hangzhou from September 23 to 27.

Digital Trade Venture Capital Day Expands in Scope

Building on the success of its inaugural edition, this year’s “Digital Trade Venture Capital Day” has been comprehensively upgraded and expanded. Enterprise screening, capital matching, and main activities were further integrated and advanced. From September 22 to 23, the event focused on sectors including artificial intelligence, hard technology, healthcare, and biotechnology, conducting selection of innovative tech companies, institutional matching, and intent tracking. The initiative attracted 100 well-known venture capital institutions, 100 professional investors, and 100 high-quality science and technology innovation projects, facilitating precise exchanges and matchmaking.

AI Competition Shifts from “Parameter Wars” to “Real-World Deployment”

During the Venture Capital Day, Zheda Holding Group released the White Paper on China’s AI Technology, New Consumption, and Investment Trends (2026) (hereinafter referred to as the “White Paper”). The White Paper notes that China’s AI industry is accelerating from technological breakthroughs to large-scale commercial deployment. Open-source ecosystems are rising rapidly, and “tech-driven new consumption” has become a key lever for industrial application. “Investing early, investing small, and investing in hard tech” has become a market consensus.

“The logic of AI investment is systematically moving upstream,” the White Paper states. Key observations include:

• AI competition is shifting from “parameter wars” to “real-world deployment,” with model adaptability and low cost becoming core enterprise demands.
• Tech-driven new consumption is a key scenario for large-scale AI deployment, with agentic AI and physical AI jointly forming a closed loop of “vague need identification to precise value delivery.”
• Embodied intelligence has become the primary investment theme, with capital moving from “investing in complete machines” to deeper areas such as “investing in brains and components.”
• Going global has entered a “system competition” phase, requiring enterprises to simultaneously possess capabilities in product definition, model iteration, compliance operations, and localized services.

The AI industry is now entering an accelerated phase of commercial deployment. Yu Feipeng, Deputy Secretary of the Party Committee and General Manager of Zhejiang University Holding Group, suggested observing this from four dimensions:

1. Policy-driven: Both national and local governments have introduced top-level plans to promote AI empowerment across industries. “AI empowers everything” has become an industry consensus.
2. Capital-backed: Investment scale and the number of participating institutions in the AI sector continue to rise, with new projects constantly emerging.
3. Commercial deployment: The industry is generally concerned that a complete closed loop among AI investment, output, and cash flow has not yet formed, leading to some debate over potential bubbles.

“However, infrastructure investment inherently has an advanced nature. High investment does not mean the business model is unviable; it simply means that AI penetration across industries remains uneven,” Yu said. In terms of technological iteration, large model parameters and comprehensive capabilities are rapidly evolving, and vertical domains are likely to generate many new AI agents.

Yu noted that the industry often compares AI to the Fourth Industrial Revolution, though no unified definition has yet been established. The market remains divided on whether AI is a bubble—some liken it to the 2000 internet wave, while others compare it to the steam engine or the electric revolution. Different reference points lead to vastly different views on the industry’s current stage and future potential.

“China was largely absent from the first three industrial revolutions. In this transformation, however, China has the opportunity to run alongside—or even lead—and achieve leapfrog development,” Yu said.

AI Agents to Restructure Manufacturing Processes

Physical AI refers to intelligent systems capable of understanding the laws of the physical world and perceiving, deciding, and interacting within real-world environments. It is seen as a critical phase in AI’s evolution from the digital to the physical world. Currently, Physical AI remains in a validation stage. Its technical feasibility has been preliminarily demonstrated at the proof-of-concept level, and the focus is shifting to real-world scenario validation to drive the formation of application closed loops.

How should we view the evolution of Physical AI from real-world scenarios? Cao Guoxiong, Chairman of Puhua Group and Co-Founder of Toutoushidao, said the core of robot development lies in redefining the relationship between humans and machines. The first applications should involve replacing humans in hazardous environments, making humanoid form less important. In specific scenarios, hotel food delivery has already been implemented, while cleaning robots are still being advanced. Due to persistent labor shortages in service roles, these scenarios have become essential needs. Hazardous environment inspection is the next direction for development.

Cao believes that, in the long term, the true maturity of embodied intelligence will restructure production models. Current factory processes are designed around human operational habits. In the future, intelligent agents will re-architect production workflows, with humans only responsible for defining and supervising tasks. Many startups are already moving in this direction.

“The internet has accumulated vast amounts of text data, supporting the training of large models like ChatGPT and Kimi. But usable data from the physical world is almost nonexistent,” said Wang Xiang, Founding Partner of Gaoshan Xinyu and former Partner at Xiaomi. Real-world data is the core prerequisite for training Physical AI foundation models. Without high-quality real-world data, embodied intelligence cannot truly be deployed in social and household settings. Drawing from the experience of autonomous driving, the ideal data collection method is to acquire real-scene data directly during operations. Existing collection methods have limitations: video capture, even from real scenes, has flaws; wearable devices also face numerous challenges. An efficient approach is to collect real-world data naturally during daily work and life.

Analyzing electromechanical signals represents a transformative technological path. Wang explained that when a person performs actions like picking up a cup or folding clothes, from the brain sending instructions to the hand executing them, corresponding muscle electrical signals are generated. Collecting, standardizing, and annotating these signals could significantly advance the Physical AI industry ecosystem, with data being the most critical element.

AI-Assisted Diagnosis and Medical Insurance Reimbursement to Become Routine

AI is reshaping medical innovation, opening a new pathway from scientific breakthroughs to clinical applications and from laboratory results to industrial transformation.

Liang Xiao, Vice Dean of Zhejiang University School of Medicine, stated that Sir Run Run Shaw Hospital began building a “Future Hospital” in 2014, continuously exploring AI-empowered medical scenarios. For example, the team independently developed a multimodal medical AI large model for liver cancer diagnosis, which was selected as a key project for the National Pilot Base. The system can receive imaging and text data and automatically output a structured report within three minutes, requiring no manual operation. Verified across multiple centers nationwide with over 10,000 cases, the detection rate for liver lesions larger than 1 cm exceeds 95%, and the accuracy for distinguishing benign from malignant lesions reaches 93.46%.

“The global AI medical market is expected to maintain rapid growth of 30% to 40% over the next five to ten years. However, clinical applications still face many practical challenges,” Liang said. These include insufficient evidence for AI-empowered healthcare, fragmented medical data forming isolated silos, and large models performing well in labs but showing significantly reduced generalization ability when equipment or hospitals change. Issues such as data privacy protection and payment mechanisms are also prominent.

Liang believes that, in the future, AI-assisted diagnosis and medical insurance reimbursement will become increasingly routine. Multimodal medical large models and agents will cover full life-cycle health management. At the same time, medical AI access should emphasize traceability and explainability. As the closed loop of AI preliminary screening and doctor review is fully implemented, human-machine collaboration will become the standard clinical workflow. On the talent front, interdisciplinary training should be systematically integrated into medical education, with medical schools cultivating students equipped with AI literacy.