The new architecture of intelligence

Editor's note: Michael Wang is an anchor for CGTN's "Global Business" program. The article reflects the author's views and not necessarily those of CGTN.

When Shanghai's merchants gathered in the city's Chamber of Commerce more than a century ago, they were grappling with the technologies that would shape modern commerce: steamships, banking systems and global shipping routes. At The Asian Banker's Shanghai International AI Finance Summit, held in the same historic building, the discussion turned to another kind of infrastructure: the architecture of artificial intelligence itself.

For much of the past half-decade, the AI race has been framed as a contest to build ever more powerful models. More data, more computing and larger neural networks were widely assumed to be the primary drivers of progress. But the assumption that scaling alone drives AI progress is being actively and broadly challenged. The catalyst is a new generation of AI agent systems.

Unlike traditional chatbots that answer a single question and stop, agent systems such as OpenClaw can execute sequences of actions: Retrieving information, interacting with software, running code and carrying out multi-step tasks autonomously. At the summit, Lin Yonghua, vice president and chief engineer at the Beijing Academy of Artificial Intelligence (BAAI), offered a framework for understanding why this matters. In Lin's analysis, an AI system's capability is no longer determined primarily by its model. Instead, performance emerges from the interaction of three components: The model, the agent system that orchestrates tasks, and the specialized "skills" that allow it to operate in professional domains.

In BAAI's internal testing, changing the underlying model could alter the cost of completing a complex task by as much as 10 times. But altering the agent system could change the cost by as much as 100 times. The surrounding architecture may now matter as much as the intelligence of the model itself.

For businesses, the implication is pointed. Large models will consolidate around a handful of global companies. But the skills layer, the certified, domain-specific capabilities that allow AI to function in professional contexts, is where industries retain leverage. For example, without high-quality financial skills, Lin argues, no agent system will function meaningfully in banking, risk management, or compliance. The institutions that build and certify these skills may define what AI can actually do in their sectors.

This shift in architecture is driving a parallel shift in hardware. Agent systems may access large models hundreds of times to complete a single task, running continuously around the clock. The result is surging demand for inference, running AI models, rather than training them. Deloitte estimates inference will account for two-thirds of all AI computing in 2026, while Lin estimates it will reach over 70-percent. Nvidia is preparing a dedicated inference chip. OpenAI is building one too. In China, Alibaba, Baidu, Cambricon and Huawei are all prioritizing inference hardware under different competitive conditions. The era of one general-purpose chip doing everything is ending.

The proliferation of chip architectures has created its own problem: What Lin calls "one chip, one stack," where each chip requires its own software ecosystem, making it costly to move models between platforms. BAAI's response is FlagOS, an open-source universal software stack developed with Peking University, Tsinghua University, the Chinese Academy of Sciences and more than a dozen chip manufacturers. It already supports over 20 chips. The goal: Just write a model that can run on any chip.

But even if the hardware and software problems are solved, a deeper question remains: Whether can we trust the data on which these systems are based. As agents move from advising humans to autonomously placing orders, transferring funds and managing supply chains, the consequences of flawed data are amplified exponentially. In supply chain finance, cases of fabricated trade activities, goods pledged multiple times, and containers that turned out to hold little value have caused losses of hundreds of millions of dollars. These are not failures of AI. They are failures of the data environment AI is now being asked to operate within. As Lin puts it: The records must match reality. Without that foundation, more powerful AI does not reduce risk. It magnifies it.

BAAI and its partners are deploying multimodal large models to interpret complex physical environments while specialized smaller models perform high-precision verification. Agents connect these outputs to enterprise systems, closing the loop between the physical world and the digital record. If this architecture matures, it could unlock activities that financial institutions currently avoid simply because they cannot verify what is happening on the ground.

The race to build the most powerful model has not ended, but it is no longer the only race that matters. The deeper contest over who builds the architecture that makes intelligence trustworthy, affordable and universal is just beginning.