Skip to main content

2026 CSDI: AI's Next Leap—From Models to Intelligent Systems

The Shift from Models to Systems

At WAIC 2026, the large model and generative AI track featured 130 exhibitors, but only 18 were pure base model companies. The rest—83 of them—were showcasing AI agents or intelligent applications. The conversation has moved from "how smart is the model?" to "what real value can it create?" The fundamental unit of AI competition is no longer a single model but an entire system.

This shift reflects a broader industry trend: AGI is no longer just a distant dream. It's becoming the strategic anchor for research organizations worldwide. DeepSeek's founder, Liang Wenfeng, recently outlined a six-step roadmap to AGI, emphasizing that each step—from language models to embodied intelligence—builds on the previous one. "Agents should use CoT, and CoT should use previous steps," he said, highlighting the importance of a progressive approach.

The Building Blocks: Compute, Data, Algorithms, Talent, and Organization

If AGI defines the destination, then compute, data, algorithms, talent, and organizational structure determine how far we can go. These are the core development elements that every serious AI player must master.

Compute: From Single Chips to Systems

China's intelligent computing power reached 2185 EFLOPS by June 2026, according to the Ministry of Industry and Information Technology. But the real challenge isn't just about raw chip performance. Interconnection, memory sharing, and task scheduling are what turn a cluster of chips into effective computing power. The focus is shifting from training to inference, and the competition is now about building entire systems, not just faster processors.

Data: From Managing to Using

In 2025, China used 199.48 exabytes of data for AI training and inference—a 42.86% increase from the previous year. For the first time, inference data surpassed training data. But as Liu Liehong, director of the National Data Administration, points out, the real value lies in using data, not just managing it. Breaking down data silos and enabling shared governance is essential for AI to take root in real-world industries.

Algorithms: Beyond Micro-Innovations

Algorithmic progress has reached international standards, but the next breakthroughs won't come from tweaking existing frameworks. The industry is moving toward multimodal world models that understand physical laws. The ability to handle long-range tasks—turning grand goals into thousands of executable subtasks—is the new frontier.

Talent: The Scarcest Resource

The demand for high-performance computing engineers is intense, with a supply-demand ratio of just 0.15. AI scientists command average monthly salaries of 132,000 yuan. But attracting talent is only half the battle. Retaining and inspiring creativity requires a strong organizational culture. DeepSeek, for instance, unites its researchers around the long-term goal of AGI.

Organization: From Racing to Unifying

A notable trend in 2026 is the consolidation of AI research efforts. ByteDance merged its Feishu product team into Douyin, Alibaba integrated its Agent product lines, and Tencent pooled its R&D resources. Even Amazon shut down its AGI Lab to merge departments, while Google DeepMind restructured to focus on AGI development. These moves signal a shift from internal competition to unified, system-level innovation.

Why Agents Still Depend on Base Models

Agents have made huge strides, moving from simple chatbots to autonomous workers that can execute tasks for hours. But their capabilities are ultimately limited by the pre-trained models underneath. As Professor Tang Jie from Tsinghua University puts it, "Pre-training determines the ceiling; fine-tuning decides whether you reach it."

Tencent Cloud's Li Qiang uses a car analogy: "The large model is the engine, and the engineering chain is the process of turning the engine into a car. The engine determines the upper limit." No matter how sophisticated the agent framework, if the base model isn't strong, the agent will hit a wall.

This is why some researchers are returning to pre-training itself. Shanghai AI Lab's Agents-A1 model, for example, was trained on high-quality long-range trajectory data to improve its agent capabilities from the ground up. The lesson is clear: the competition in agents is really a competition in base models.

The Rise of Agentic Programming

Software development is undergoing a paradigm shift. Gartner predicts that by 2028, AI coding agents will boost software engineering productivity by 30% to 50%. Companies like Stripe, Ramp, and Coinbase have already deployed internal coding agents that work asynchronously, triggered by team communications rather than individual commands.

But the real value comes from organizational-level deployment, not just individual use. As IDC notes, the difference lies in whether companies have a comprehensive plan for platform engineering, governance, and developer role transformation. Those that treat Agentic AI as an enterprise capability will gain long-term advantages in speed, quality, and innovation.

Key Points

  • AI competition is shifting from single models to integrated systems
  • AGI is becoming a strategic goal for major research organizations
  • Compute, data, algorithms, talent, and organization are the core development elements
  • Agent capabilities are fundamentally limited by pre-trained base models
  • Agentic programming is transforming software development at the enterprise level