Huang Renxun's "Five-Layer Cake" Industrial Logic and China's Opportunities

Recently, NVIDIA CEO Jensen Huang proposed an inspiring industry model—the “Five Layers of AI Cake.” In the view of this one of the most powerful tech leaders today, artificial intelligence is not just a single software iteration but a complex, highly coupled system built from the ground up. These five layers are: Energy, Chips, Infrastructure, AI Models, and Applications.

While public and capital markets are captivated by dazzling application-layer products like ChatGPT and “Lobster” (OpenClaw), Huang’s “Five Layers” theory acts as a wake-up call. It not only sketches a panoramic view of the generative AI industry chain but also reveals a harsh reality often overlooked: the competition in AI is not just about code but a brutal race for atomic-level resources in the physical world.

1. Energy Layer: The “New Oil” and the Ultimate Bottleneck of the AI Era

At the bottom of Huang’s “Five Layers,” it’s not data but energy—more precisely, electricity. Over the past 20 years, the prosperity of the internet was based on the assumption that the marginal cost of “bits” approaches zero. However, large models have shattered this myth. From training to inference, generative AI consumes vast amounts of electricity with every breath.

Huang placing energy at the first layer hits the industry’s most hidden pain point. In fact, not only NVIDIA but also Silicon Valley leaders like Sam Altman and Elon Musk have keenly sensed that the limit of computing power is electricity. If the previous generation of tech giants built moats through data traffic monopolies, the next AI giants’ lifeline will depend on grid capacity and clean energy supply.

This explains why many Silicon Valley capital flows are now into frontier energy projects like nuclear fusion, geothermal, and solar power. In the “Five Layers,” the scarcity of energy at the bottom directly determines the expansion boundary of the upper layers. Without sufficient, stable, and low-cost electricity, even the grandest AI visions are castles in the air.

2. Chip Layer: The “Silicon Heart” of Computing Power Hegemony

Built upon energy is the second layer: chips. This is NVIDIA’s absolute domain and the heartbeat of the entire AI industry. As Moore’s Law wanes, “Huang’s Law” is taking over the pace of computing growth. From A100 and H100 to the latest Blackwell architecture, AI chips are no longer just silicon stacks but integrated with cutting-edge packaging and high-bandwidth memory in super-engineering feats.

The industry logic here has fundamentally changed. Barriers in the AI chip market are no longer just about hardware design finesse but are built through software moats like the CUDA ecosystem. This integrated hardware-software monopoly gives the bottom layer of computing power the strongest value capture ability in the current AI supply chain. NVIDIA’s soaring market value is a direct confirmation of its dominant position in this layer. However, this highly centralized computing power also pushes other tech giants (like Google’s TPU, Microsoft’s Maia, Meta’s MTIA) to accelerate in-house chip development, trying to carve out a slice of the “cake.”

3. Infrastructure Layer: The Underestimated “Invisible Artery”

The third layer is infrastructure, including data centers, compute networks, advanced cooling systems, and supporting power grid facilities. This is the most asset-heavy layer in the “Five Layers” and currently the most underestimated field in commercial opportunities. Training large AI models requires thousands of GPUs working in ultra-low latency collaboration, testing not only chips but also network topology and data exchange capabilities. Meanwhile, the intense heat generated by dense compute demands is forcing data centers to shift from traditional air cooling to liquid cooling and phase-change cooling technologies.

In this layer, we see a deep integration of the tech industry with traditional manufacturing. Coolant suppliers, optical module manufacturers, server rack assemblers—these formerly peripheral “water sellers” are experiencing a historic valuation reset. The prosperity of AI is reconstructing global heavy-asset infrastructure at an unprecedented intensity—a “rust belt revival” with no visible gunfire but costing trillions.

4. AI Model Layer: The Arena for Intelligence Equality and Organizational Restructuring

Reaching the fourth layer, we enter the realm of cognition—AI models. This is the main battlefield for OpenAI, Google, Meta, and numerous startups competing in large models. Notably, the business models and organizational forms here are undergoing dramatic evolution. On one side are the “closed-source” camps attempting to build all-encompassing AGI (Artificial General Intelligence) through massive compute barriers and data flywheels; on the other are the “open-source” factions trying to democratize technology to break the monopolies of the former.

This process also triggers profound organizational changes in Silicon Valley. top AI researchers and scientists are breaking away from traditional paths tied to large corporate research institutes, sparking a new wave of entrepreneurship. They are either acquired at sky-high prices or forming new AI R&D alliances through restructuring. Competition at the model layer is essentially a contest for top human intellectual capital and a reorganization of it. However, from a business perspective, the model layer faces “involution”: as training costs rise exponentially, how to sustain a profitable business loop remains a tough problem.

5. Application Layer: The Final Anchor of AI Value

At the top of the “Five Layers” is the application layer, directly interacting with the physical world and end-users. Huang specifically highlights robotics and autonomous driving. The second half of generative AI will inevitably shift from virtual text and image generation to embodied intelligence and complex system decision-making (like autonomous vehicles and industrial control). The application layer is crucial for transforming AI from a capital bubble into a real industry.

This explains why concepts like “Lobster” (OpenClaw) and AI agents are now highly sought after. Pure chatbots cannot sustain trillion-dollar business visions. Only when AI becomes capable of understanding environments, calling tools, and autonomously executing tasks—possibly embedded in humanoid robots—penetrating manufacturing, services, healthcare, and transportation—can the enormous sunk costs of the first four layers realize true value.

6. The “Barrel Effect” of the Industry Chain and Insights for China

Huang’s “Five Layers” model provides a coordinate system for examining the AI industry. Its greatest insight is revealing the strong coupling and “barrel effect” of the industry chain. Any weakness in one layer limits the overall industry’s ceiling. For Chinese enterprises, this presents both severe challenges and structural opportunities.

From the challenge side, Chinese companies face significant “middle-layer constraints.” In the second layer (chips) and parts of the third (advanced networks and infrastructure), due to geopolitical export controls and physical barriers of advanced processes, China faces an objective gap in computing hardware. The high costs of underlying compute power directly raise the training barriers of the fourth layer (AI models), forcing us to bear higher capital and trial-and-error costs in catching up with top global general models.

On the opportunity side, China holds relative advantages at both ends of the “cake.”

In the energy layer, China leads in photovoltaics, wind power, ultra-high-voltage transmission, and energy storage infrastructure. Of course, large energy capacity does not mean AI companies can immediately access low-cost, stable, green, and load-appropriate electricity. Market mechanisms, cross-regional dispatch, and energy absorption efficiency still determine how much this advantage can be realized.

In the model layer, recent years have seen rapid progress in open-source models in China, with notable advances in low-cost inference, industry-specific fine-tuning, and vertical scenario adaptation. Huang’s mention of DeepSeek indicates Chinese model companies are now entering the global forefront.

In the application layer, China uniquely possesses all industrial categories in the UN’s industry classification, with abundant testing scenarios and strong engineering capabilities in robotics, autonomous driving, and intelligent manufacturing.

Facing these circumstances, Chinese enterprises and policymakers can adopt strategies of “cross-layer coordination and asymmetric competition”:

First, build foundational advantages with “green electricity + compute power.” Further deepen macro strategies like “East Data, West Computing,” guiding heavy-asset compute infrastructure to transfer to clean energy-rich central and western regions. Through institutional innovation, enable AI companies to access low-cost green power directly, offsetting hardware premiums at the chip layer.

Second, focus on data element circulation and vertical large model ecosystems. While keeping pace in the general large model track, prioritize breakthroughs in vertical models. Accelerate clarifying data ownership, pricing, and supply-demand mechanisms, activating vast dormant data assets. Use industry-specific proprietary data with high barriers to develop large models that truly solve business pain points in finance, healthcare, and industrial manufacturing.

Third, leverage application-layer profits to fund R&D at the chip layer. Chinese companies should fully utilize the advantages of “super factories” and a vast domestic market, accelerating the commercialization of embodied intelligence and autonomous driving in the physical world. Only when the application layer generates real cash flow can it sustain the domestic replacement of chips, high-end optical modules, and liquid cooling systems, ultimately enabling a downstream-to-upstream industry chain breakthrough.

(Author Wang Xiang is a researcher at Fudan University’s Digital and Mobile Governance Laboratory)