At the 2026 World Robot Expo (WRC), held from August 19-23 at the Beijing Yichuang International Convention and Exhibition Center, Yin Jilei, co-founder and CEO of WeFond AI, participated in a roundtable discussion on "Scenario-Driven Application Innovation and Industrial Chain Collaboration" during the "AI Large Models Empowering Robots and Embodied Intelligence: A New Industry Paradigm" special session. During the dialogue, Yin shared insights on the company's positioning in the robotics chip sector, the challenges of domestic substitution, and his projections for the humanoid robot market's growth trajectory.

Introducing himself, Yin explained that he brings two decades of semiconductor industry experience to his current role. His first decade focused on non-AI chip development, covering areas such as cloud computing chips, PC northbridge and southbridge chips, mobile phone chips, and digital TV chips. Over the past decade, he has concentrated on AI inference chips for wearable edge devices, including those used in earbuds, smartwatches, and fitness bands. More recently, WeFond AI has pivoted toward the robotics sector, developing "big brain and small brain" fusion chips designed for embodied intelligence applications.

Addressing the current state of the computing market, Yin noted that China's high-end edge computing market remains heavily reliant on imported solutions. Nvidia's Orin and Thor platforms currently dominate the landscape, and WeFond AI's mission is to create cost-effective alternatives to these chips. "We're approaching this with an ecosystem-building strategy," he explained. "On the software side, we're embracing CUDA compatibility to support mainstream development ecosystems, while on the hardware side, we're combining general-purpose computing with neuromorphic computing principles to create brain-inspired, low-power GPUs."

When asked about the feasibility of creating "replacement" chips for Nvidia's offerings, Yin acknowledged that the market currently shows limited demand for alternatives. The robotics industry is still nascent, with few companies achieving mass production, and the entire supply chain is still grappling with fundamental questions about reliability—specifically, how to ensure robots can operate continuously for hundreds or even thousands of hours without failure. Nevertheless, he emphasized that chips serve as the essential infrastructure for the industry, acting as the master controller that processes multimodal sensory data, executes planning and decision-making, and coordinates joint movements to complete end-to-end task loops.

Yin highlighted three critical challenges facing the current chip landscape. First, power consumption remains a significant hurdle, as embodied intelligence models—typically distilled from cloud-based large language models with parameters ranging from several billion to tens of billions—demand substantial computing power while generating excessive heat. Second, cost presents a major barrier to scaling. The Nvidia Thor chip alone commands a price tag of at least 50,000 RMB, making it prohibitive for widespread adoption. Third, existing chips were never natively designed for robotics; they were repurposed from autonomous driving applications, representing a "make-do" solution rather than an optimized approach.

For startups, this gap represents a genuine opportunity. "Creating domestic alternatives is absolutely necessary," Yin stated. "Chips are essentially the 'industrial grain' of the sector. Whether we're talking about autonomous driving or robotics, this is new quality productivity, and no country would allow such a critical industry to remain without self-sufficiency." He acknowledged the difficulty of the path forward but described it as "difficult yet correct," addressing the dimensions of cost, efficiency, power consumption, and autonomous control simultaneously.

Discussing neuromorphic computing, Yin highlighted its fundamental advantage in power efficiency. "The human brain operates on roughly 20-30 watts of power, yet it can perform arithmetic, write code, do laundry, and cook—it's true AGI," he observed. By contrast, current cloud-based large models require hundreds of watts to run. He cited a recent example: during a sports event project with partners, a Thor-powered outdoor scenario application crashed after just 30 minutes due to excessive power consumption. Neuromorphic computing addresses this through sparsity, event-triggered processing, and binary spike-based computation, which can be integrated into general-purpose architectures to dramatically reduce energy requirements.

Yin also expressed optimism about neuromorphic intelligence as a computing paradigm. Noting that some industry peers are exploring similar brain-inspired approaches, he suggested that the conventional "Scaling Law" model—which relies on brute-force computation—may prove unsustainable. "The human brain's approach appears to be the most economical and efficient," he said, adding that neuromorphic computing and brain-inspired algorithms represent promising frontiers worth deeper exploration.

Looking ahead, Yin offered a contrarian perspective on industry momentum. While 2026 is widely considered the "year one" for robotics commercialization, he cautioned that this designation doesn't necessarily translate into explosive growth. However, he did project meaningful breakthroughs. Citing industry consensus figures, he noted that humanoid robot shipments reached approximately 18,000 units in 2025, with estimates of around 60,000 units in 2026. By 2027, he forecasts shipments could reach 100,000 to 200,000 units—a substantial leap driven not by any single company or scenario, but by incremental progress across multiple verticals.

"Breakthroughs will come from specific specialized scenarios," Yin explained, "and when a single point achieves critical mass, whether in electronic skin, dexterous hands, or computing power, it will catalyze the maturation of the entire industrial chain." He emphasized that these developments signal genuine industrial chain collaboration and real-world deployment rather than mere laboratory demonstrations.

Offering a candid assessment of the domestic chip market, Yin acknowledged that near-term demand for domestic alternatives in robotics will remain limited. Robotics applications fall into the industrial-grade category, where products follow 3-5-10 year replacement cycles rather than annual upgrades. "Nvidia's solutions will remain the mainstream for the next two years," he conceded. However, he advised startups to leverage this window to prepare thoroughly for the coming production surge. "2027 will be an exciting year for many companies," he predicted.

Reflecting on the evolving technology landscape, Yin noted that last year's buzzword was VLA, while this year it's World Model—and next year's hot topic remains unknown. The rapid turnover of trends is itself a positive signal, he argued, as it indicates growing industry participation and exploration. "Robots and AI are no longer unfamiliar to people; they're part of everyday life," he observed. "The novelty of watching a robot dance has worn off, but the imagination around robots working in real scenarios is incredibly compelling." He concluded by expressing enthusiasm for the cognitive shifts in how people perceive AI and the maturation of the robotics industrial chain, both of which he believes will drive meaningful innovation in the years ahead.