Capital stratification from chips, cloud, models to applications

Over the past two years, the most common question when discussing AI in the US stock market has been: Is the P/E ratio of a certain stock too high? This question is not meaningless, but using a single layer of P/E to judge the entire AI industry chain often leads to overly simplistic conclusions. AI is not a single industry but a multi-layered capital structure composed of chips, servers, cloud, models, data, software, and applications. Each layer differs in revenue quality, capital expenditure, depreciation cycles, customer concentration, and pricing power; thus, valuation methods should also differ.

The bottom layer consists of chips and semiconductor equipment. This layer plays the role closest to 'shovels and railroads,' with companies like Nvidia $NVIDIA (NVDA.US)$ , AMD, Broadcom $Broadcom (AVGO.US)$ , Taiwan Semiconductor $Taiwan Semiconductor (TSM.US)$ , ASML $ASML Holding (ASML.US)$ , Applied Materials $Applied Materials (AMAT.US)$ Companies such as these are at the forefront of the AI boom, addressing the earliest and most tangible demand: computing power. When cloud giants, model companies, and enterprise clients all need training and inference capabilities, GPUs, advanced packaging, HBM memory, optical communication, and networking chips naturally become bottlenecks. Where there is a bottleneck, there is often profit. Therefore, the market is willing to assign higher valuations to chip leaders not just because of the 'AI story,' but because they hold real pricing power in the supply chain.

However, the valuation of chip stocks cannot be based solely on high growth. The semiconductor industry is inherently cyclical; once customers overstock, technology shifts, or cloud giants accelerate their in-house chip development, the growth curve could suddenly slow down. A chip company's high P/E ratio actually contains two assumptions: first, that AI computing power demand will remain higher than supply in the long term; second, that leading companies can maintain their technological and ecosystem advantages. If either assumption falters, valuation compression can happen quickly.

The second layer is cloud and data center infrastructure. Companies like Microsoft, Amazon, Google, Oracle, and Meta are heavily investing in GPUs, data centers, networks, and power capacity. They are not just buying chips but building the next-generation computing platform. The difficulty in valuing this layer lies in the fact that capital expenditures come first, while returns lag behind. Investors cannot just look at cloud revenue growth; they also need to ask: how much incremental revenue can each dollar of AI capex ultimately generate? Will services like Copilot, Gemini, AWS AI services, and Oracle Cloud GPU rentals deliver sufficiently high returns?

Cloud giants have the advantage of strong capital, a vast customer base, and mature distribution channels, but the risk is the sharp rise in capital intensity. In the past, the market liked cloud businesses because they exhibited economies of scale and high operating leverage. However, AI clouds may change this model: servers are more expensive, depreciation is faster, and electricity and cooling costs are higher. If AI revenue cannot quickly cover depreciation, cloud profit margins will be under pressure. Therefore, the valuation of cloud stocks should not only consider traditional software multiples but also incorporate analysis of return on capital and free cash flow.

The third layer consists of large models and foundational model platforms. This layer captures the market's imagination the most but is also the hardest to value. Model companies require massive computing power, top talent, and data, with extremely high training costs, but product prices might not fully reflect these costs. As model capabilities increasingly converge, competition may shift from technological leadership to factors like distribution, branding, security, enterprise integration, and ecosystem lock-in. In other words, the model itself may not always be the most profitable layer; what truly holds value is whether the model can become part of enterprise workflows, development tools, and consumer entry points.

This also explains why platform companies like Microsoft, Google, and Meta are more readily accepted by the market in the AI race compared to pure model companies. These companies don't necessarily rank first in every model benchmark, but they can embed models into search, advertising, office software, social platforms, and cloud services. Without distribution capabilities, a model can easily turn into a high-cost R&D project; when connected to existing users and workflows, it has the chance to become a recurring revenue product.

The fourth layer is enterprise software and applications. This includes companies like Palantir, ServiceNow, Salesforce, Adobe, Snowflake, Datadog, and CrowdStrike. Their AI narrative isn’t about who owns the most GPUs but who can turn AI into features that customers are willing to pay for. Enterprises won’t endlessly increase budgets for 'AI' alone—they want to reduce labor, boost sales, accelerate development, lower risks, or improve decision-making. Therefore, the key to valuing the application layer isn’t how impressive the demo is but whether AI can improve retention rates, drive price increases, expand usage, and ultimately reflect in revenue growth and gross margins.

The advantage of the application layer is lower capital expenditure, and if the product is truly adopted, operating leverage can be very strong; the downside is that competitive barriers may be lowered. Generative AI makes many functions easier to replicate, and the moat originally built through interfaces and processes could potentially be broken by AI agents. In the future, investors will need to distinguish not which companies 'have AI capabilities,' but which companies make it harder for customers to leave because of AI.

(Chip and Computing Power Series, Part 57)

110K Views