The Three Layers of Business and Profit Pools in the AI Industry Chain

Viewed through the lens of profit pools, the AI industry actually consists of three layers: the first layer earns revenue from chips, the second from computing power and cloud rentals, and only the third layer generates profits from applications and services.

Currently, the deepest profit pool undoubtedly remains in chips and hardware infrastructure. GPUs, specialized AI accelerators, high-speed interconnects, HBM memory, advanced packaging, liquid cooling systems, and server platforms—required for AI training and inference—form the foundation of the entire industry chain. The business model at this layer is the most straightforward: customers must purchase hardware before they can deploy AI; the larger the models, the more parameters they have, and the higher the inference demands, the greater their reliance on computing power. Chip companies thus enjoy a position akin to collecting an 'entry toll'—as long as AI-related capital expenditures continue, hardware suppliers will be the first to benefit from incremental demand.

However, while chip-related profits are the most attractive, they also come with the highest barriers to entry. This is not a market where ordinary companies can gain traction simply by telling compelling stories; success requires integrated capabilities in chip architecture, software ecosystems, process technology supply, packaging expertise, customer certifications, and substantial capital investment. This explains why the AI hardware market tends to reinforce a 'winner-takes-most' dynamic. The real winners aren’t all semiconductor firms, but specifically those that control critical bottlenecks. GPUs are a bottleneck, HBM is a bottleneck, advanced packaging is a bottleneck, and data center power and cooling can sometimes become bottlenecks as well. Investors who see only the phrase 'AI chips' without discerning who controls these bottlenecks, who is merely a contract manufacturer, and who is just riding on conceptual hype are likely to misjudge where the true profit pools lie.

The second profit pool lies in cloud platforms and computing power leasing. Most enterprises lack the capability to build their own AI data centers and cannot afford to purchase, maintain, or manage large quantities of high-end chips themselves. As a result, cloud platforms have become the new landlords of the AI era. They procure chips, build data centers, and monetize through cloud services, API calls, model hosting, inference services, and other offerings. This layer doesn’t generate one-time hardware sales but instead earns recurring rental income. As businesses continuously invoke models, process data, deploy AI-powered customer service, generate content, or run internal AI tools daily, cloud platforms charge usage-based fees—much like utilities billing for water and electricity.

This rental-style business may appear stable, but it is not without pressure. Cloud platforms must first bear massive capital expenditures—buying chips, building data centers, signing power contracts, and enhancing networking and cooling capabilities. Second, they face hardware depreciation, as AI chips evolve rapidly; today’s cutting-edge equipment may lose its cost advantage within just a few years. More importantly, competition among cloud platforms is fierce, and as computing power supply gradually increases, the price per unit of computing power could decline. Therefore, whether cloud platforms can continue earning high 'rent' depends on two factors: first, their ability to leverage economies of scale to reduce costs; and second, their ability to tightly integrate computing power with data, models, development tools, and enterprise customer stickiness. Merely renting out GPUs may, in the long run, become a capital-intensive business; only by transforming computing power into an enterprise AI operating system can they build a deeper moat.

The third layer is application monetization—the layer easiest for the market to conceptualize, yet the most uncertain in terms of profitability. AI applications can emerge across scenarios such as office software, software development, customer service, financial analysis, medical imaging, education, legal services, design, gaming, advertising, e-commerce, and industrial manufacturing. In theory, this layer is closest to end users and has the largest addressable market; in reality, however, many AI application companies are still at a stage where 'users exist, but few are willing to pay full price.' The reason is simple: foundational large model capabilities are diffusing rapidly, making application features easy to replicate. Without embedding deeply into workflows, owning proprietary data, or increasing switching costs, these innovations risk becoming mere plugins.

This is also where valuing AI applications becomes most challenging. Some companies may appear to be growing rapidly, but their revenue may simultaneously come with high model inference costs, substantial customer education expenses, and pressure to retain users. If a company merely wraps a large model into a chat interface without truly transforming industry workflows, its gross margins and pricing power are unlikely to be strong. Conversely, truly valuable AI applications don’t just 'generate a piece of text' or 'answer a question'; they shorten delivery cycles, reduce labor requirements, improve conversion rates, lower error rates, and even reshape an industry’s cost structure. Only then do application companies move beyond selling tools—they begin sharing in the economic gains generated by their customers’ improved efficiency.

$NVIDIA (NVDA.US)$$Oracle (ORCL.US)$$Meta Platforms (META.US)$$Alphabet-C (GOOG.US)$$MetaX Integrated Circuits (688802.SH)$

338K Views