At the recent AI Investment Summit hosted in Beijing on September 16, industry discourse centered on the theme of "Defining Opportunities in the AI Infrastructure Era." Among the key voices was He Yongzhan, General Manager of Basic Hardware R&D at Baidu Intelligent Cloud, who offered a grounded perspective on how artificial intelligence interaction models are evolving and what this means for the backbone of computing power infrastructure.
He observed that over the past year, user interaction with AI has undergone four pivotal shifts. First, the search box has transformed into a command execution interface, moving away from mere link retrieval toward handling tasks involving long text, documents, and multimodal audio-visual inputs. Second, AI has stepped out of standalone applications to become deeply embedded across a wider array of everyday digital touchpoints. Third, user demand has transitioned from straightforward Q&A to complex multimodal task flows, with a stronger emphasis on end-to-end operational capability. Fourth, the consumer mindset has grown more pragmatic and discerning, as creative outputs are often met with tolerance, while factual queries now demand credible, traceable sources.
According to He, the single most significant industry development is the scaling of AI "digital employees" in real-world applications. Enterprises are increasingly using AI to automate linked, sequential tasks. Within his own team, internally developed digital employees now handle roughly 60% to 80% of repetitive, question-and-answer work. This changes the core business logic: the focus is no longer on merely selling computing power or models, but on delivering tangible business outcomes. Correspondingly, the key metric for gauging AI’s commercial value is shifting from token consumption to daily active agents (DAA).
As inference demand climbs sharply, intelligent agent tasks originating from consumer-facing channels are increasingly loading onto enterprise cloud platforms. A balanced focus on both training and inference workloads has become the standard expectation in the field. He highlighted the continuous evolution of supernode architecture as a critical piece of infrastructure. The industry, he argued, must tackle the challenge of achieving thousand-card-scale Scale-Up fully connected networking, while simplifying maintenance standards to enable point-to-point connectivity across 1,024 accelerators.
Acknowledging the technical gap between domestic and imported chips, He stressed the need to address two major bottlenecks: memory walls and storage walls. Advancing memory pooling alongside persistent storage solutions is essential. Test data indicates that such storage approaches can boost prefill-stage input throughput sevenfold. The sector also needs to adapt to a growing variety of emerging specialized chips, establish robust GPU interconnect systems, and accelerate the full-stack localization of supernode technology. This covers compute, storage, high-speed interconnect, mechanical structure, liquid cooling, and power supply chains, thereby creating expansive deployment scenarios for domestic chips and building a positive industry flywheel.
He outlined four industry signals worth tracking closely. The first is the surge in high-density storage demand, driven by agent workloads leading to exponential growth in data storage volumes. The second is the critical incubation phase for dedicated AI inference chips, where specialized inference areas are showing increasing commercial value. The third points to domestic interconnect chips nearing volume-production readiness, with switch and retimer chips emerging as core bottlenecks where local alternatives are gaining maturity. Fourth, the liquid cooling market is approaching its prosperity window, as advanced overseas hardware commonly adopts liquid cooling, while high power consumption in domestic chips further entrenches the technology’s necessity, with widespread deployment expected next year.
Comments