AI Stocks: From Matrix Multiplication to Market Valuation

This blog examines the AI-related stock market, which has delivered exceptional returns over the past several years and whose long-term prospects remain promising. Despite this favorable outlook, AI-related stocks have recently experienced significant volatility. Much of this reflects the tension between exceptionally high market expectations and uncertainty surrounding AI monetization, compounded by elevated valuations, interest-rate sensitivity, geopolitical risks, and export controls. The objective of this blog is to develop a framework for understanding the long-term evolution of the AI investment landscape.

Before proceeding, I should note that I am not a professional financial analyst and have been actively investing for only about one year. My professional background lies instead in mathematics, artificial intelligence, engineering, medical imaging, and related quantitative disciplines. These perspectives, together with discussions with friends and extensive use of ChatGPT as a research and editing assistant, have helped me organize my thoughts on the rapidly evolving AI ecosystem.

AI Is Not a Single Industry. Although investors often speak of "the AI industry" as though it were a single sector, AI is more accurately described as a multi-layered industrial ecosystem. At a broad level, this ecosystem consists of three interconnected layers: Infrastructure, including semiconductors, memory, networking, data centers, electrical power, and cloud computing; Platforms, including operating systems, cloud services, enterprise software, and digital ecosystems; and Applications, including AI assistants, enterprise automation, robotics, healthcare, finance, manufacturing, and countless domain-specific solutions. Over the past several years, financial markets have largely focused on the first layer. The investment thesis was remarkably straightforward: More AI requires more computation. While seemingly simple, this statement captures one of the fundamental economic realities of modern artificial intelligence.

Why Computation Matters. Modern AI systems—and especially large language models based on the Transformer architecture—are fundamentally massive matrix-computation engines.   Matrix multiplications, such as  $Y = XW$, are performed repeatedly at an enormous scale during both training and inference. If \(X \in \mathbb{R}^{B \times d}\) represents a batch of token embeddings and \(W \in \mathbb{R}^{d \times m}\) denotes a trainable weight matrix, then the output feature representation is computed through matrix multiplication over billions or even trillions of parameters. In practice, modern Transformer models require not only a single matrix multiplication but a sequence of extremely large tensor operations associated with self-attention, feed-forward layers, and backpropagation updates. For example, the self-attention mechanism itself involves repeated computations of the form $Q = XW_Q,~ K = XW_K, ~ V = XW_V,$ followed by $\mathrm{Attention}(Q,K,V)=\mathrm{softmax}(\frac{QK^T}{\sqrt{d_k}})V.$. Since these operations must be executed across massive datasets and very deep neural architectures, the total computational complexity becomes enormous. During training, the optimization process repeatedly updates parameters through stochastic gradient-based iterations $\theta_{t+1}=\theta_t-\eta \nabla_{\theta}L(\theta_t),$ where \(L(\theta)\) denotes the loss function and \(\eta\) represents the learning rate. Consequently, modern AI training requires extraordinary computational throughput, memory bandwidth, inter-chip communication, electrical power, and cooling infrastructure. For a more detailed mathematical discussion, see my previous blog article on large language models.

Why Infrastructure Stocks Led the AI Boom. This computational reality explains why investors initially concentrated on companies supplying the hardware foundation of AI. The earliest beneficiaries included firms providing GPUs, high-bandwidth memory (HBM), advanced semiconductor manufacturing, chip packaging, networking equipment, and hyperscale cloud infrastructure. Companies such as NVIDIA, AMD, Broadcom, Micron Technology, TSMC, SK hynix, and Samsung Electronics became central to the AI investment narrative because they supply essential components required to build and operate modern AI systems. However, the infrastructure story extends well beyond semiconductors. 

Large AI data centers consume enormous amounts of electricity. Consequently, the rapid expansion of AI computing has also increased demand for transformers, substations, transmission networks, industrial automation systems, cooling technologies, and electrical grid upgrades. As a result, investors have increasingly broadened their attention to companies involved in power generation, electrical equipment, energy management, smart-grid technologies, and industrial infrastructure. The AI investment boom is therefore evolving into something much larger than a semiconductor story. Increasingly, it is becoming an infrastructure story encompassing semiconductors, networking, cloud computing, electricity, and industrial systems.

The First Phase of the AI Investment Cycle. Viewed from this perspective, the first phase of the AI investment cycle can be characterized as an infrastructure expansion phase. Investors initially focused on identifying the companies supplying the computational backbone of AI. As demand for increasingly powerful AI models accelerated, capital flowed aggressively into GPU manufacturers, memory producers, semiconductor foundries, networking companies, hyperscale cloud providers, and electrical infrastructure firms. The underlying logic was compelling. If AI adoption continued to expand rapidly, demand for computational infrastructure would also continue to grow. This investment thesis proved remarkably successful.

Yet financial markets rarely reward even the strongest narratives indefinitely. As expectations become increasingly optimistic, valuations begin to incorporate not only current growth but also substantial assumptions about future demand. At that point, even relatively minor disappointments—whether involving earnings, export restrictions, geopolitical developments, or concerns about overinvestment—can trigger significant price corrections. The question facing investors today is therefore no longer simply how much AI infrastructure will be built, but rather who will ultimately capture the long-term economic value created by AI. That question naturally leads to the next phase of the AI investment cycle.

From Building AI to Monetizing AI. As AI infrastructure becomes increasingly mature, investors are beginning to ask a different question. The issue is no longer simply who builds AI, but who ultimately captures its economic value.

The first phase of the AI investment cycle was driven largely by capital expenditure (CAPEX). Companies invested aggressively in GPUs, memory, networking equipment, data centers, and electrical infrastructure because larger AI models required ever greater computational resources. During this period, the market rewarded firms supplying the hardware necessary to build modern AI systems. Eventually, however, every infrastructure investment raises the same economic question: Will the investment generate sufficient returns? For AI, the answer depends not only on technological progress but also on whether businesses can translate AI capabilities into sustainable revenue growth, higher productivity, or lower operating costs. Consequently, investors are gradually shifting their attention from infrastructure construction toward AI monetization.

The second phase of the AI investment cycle is therefore less about building AI models and more about integrating AI into products, enterprise software, business processes, and digital platforms that generate recurring cash flows.

The Economics of AI Monetization. Although the technical capabilities of large language models have advanced rapidly, monetization remains considerably more challenging. Developing an advanced AI model is only the beginning. Long-term value depends on whether AI can solve practical problems that customers are willing to pay for repeatedly. This naturally favors companies possessing several structural advantages: large enterprise customer bases; established software ecosystems; cloud infrastructure; proprietary data; strong distribution channels; and recurring subscription-based business models. These characteristics enable firms not merely to develop AI, but to integrate it into existing workflows where customers already generate revenue. In the United States, companies such as Microsoft, Alphabet, Amazon Web Services, Meta, and Oracle are well positioned because they control extensive cloud infrastructure, enterprise software ecosystems, and global digital platforms. In Korea, firms including Samsung SDS, NAVER, Kakao, LG CNS, and KT Corporation possess important advantages in enterprise integration, cloud services, manufacturing systems, localization, and digital platforms. Although their competitive positions differ substantially from those of global hyperscalers, they may still benefit by embedding AI into domestic enterprise and industrial ecosystems.

AI as an Enterprise Productivity Tool. Perhaps the most important transition is conceptual. The first investment wave focused primarily on AI itself. The second increasingly focuses on what AI enables. For many businesses, the greatest economic value of AI may not come from selling chatbots, but from improving operational efficiency. AI has the potential to automate repetitive tasks, accelerate software development, optimize supply chains, improve customer service, assist knowledge workers, enhance industrial inspection, and reduce administrative costs. In many industries, even modest improvements in productivity can generate significant long-term economic value.

From an investment perspective, companies capable of embedding AI into existing enterprise workflows may ultimately generate more stable returns than firms competing solely to build increasingly capable foundation models.

The following framework represents my personal interpretation of how the AI investment cycle may evolve. It is intended as a conceptual model rather than an established financial theory, and serves as a way to organize the relationships among different groups of AI-related companies.This transition can be summarized using a simple conceptual framework.

Let X denote AI infrastructure suppliers, such as Micron, NVIDIA, Samsung Electronics, and TSMC, and let Y denote AI CAPEX spenders or IT service and platform companies, such as Microsoft, Alphabet, Meta, NAVER, Samsung SDS, and LG CNS. The relationship between X and Y is fundamentally state-dependent rather than characterized by a stable positive correlation. When AI capital expenditure is viewed as productive growth investment, both X and Y may appreciate simultaneously. Infrastructure suppliers benefit from rising demand for computational hardware, while platform companies are rewarded because AI investment is expected to generate future revenue and productivity gains. However, market sentiment can change. If investors become concerned that AI infrastructure spending is becoming excessively expensive relative to near-term returns, the cost burden of AI CAPEX becomes the dominant narrative. Under these conditions, platform companies may underperform because their profitability is pressured by rising investment costs, even while infrastructure suppliers continue to benefit from strong hardware demand. The opposite scenario is also possible. If investors begin to believe that AI infrastructure investment has become excessive and is likely to produce future oversupply, infrastructure suppliers may experience declining margins and weaker demand expectations. At the same time, companies purchasing AI infrastructure may benefit from falling hardware prices and lower computing costs.

Consequently, the correlation between X and Y is not constant. It changes according to how investors interpret AI infrastructure investment:

Growth driver: both X and Y tend to perform well. Cost burden: Y tends to underperform, while X tends to outperform. Future oversupply: X tends to underperform, while Y tends to outperform.  Rather than assuming a fixed relationship between the two groups, investors should recognize that their relative performance depends on the prevailing economic regime. This framework intentionally abstracts from interest rates, which remain arguably the single most important macroeconomic factor influencing valuations across both groups.

Beyond Software: The Next Layer of Monetization. Another area attracting growing attention is digital payments and financial infrastructure. As AI agents increasingly participate in commerce—making purchases, scheduling services, managing procurement, and executing cross-border transactions—the efficiency of digital payment systems may become strategically important. Stablecoins, digital payment networks, and programmable settlement systems could therefore emerge as complementary infrastructure supporting AI-driven commerce. Importantly, the greatest economic value may not accrue to the issuers of digital currencies themselves, but rather to companies controlling the platforms through which transactions occur. Firms with large user bases, extensive merchant ecosystems, and integrated payment networks may be particularly well positioned to benefit from this evolution. The long-term opportunity therefore extends beyond AI algorithms. It encompasses the broader digital ecosystem that enables AI systems to interact efficiently with businesses, consumers, and financial markets.

Looking Beyond Monetization. Even if AI becomes a profitable business, an even larger question remains. Can AI fundamentally improve economic productivity?

This question marks the beginning of the third phase of the AI investment cycle. Rather than focusing on who builds AI or who monetizes AI, investors may increasingly ask which companies can use AI to transform entire industries through higher productivity, lower costs, and better operational efficiency.

That transition is already beginning.