The AI and Semiconductor Subsector
The semiconductor industry has existed since the 1960s, but it has never mattered more to the stock market than it does now. Nvidia's market capitalization exceeded $3 trillion in 2024, making it one of the three most valuable companies in the world. The Philadelphia Semiconductor Index (SOX) outperformed the S&P 500 by a cumulative 200+ percentage points from 2019 to 2025. Artificial intelligence has transformed semiconductors from a cyclical, capital-intensive niche into the infrastructure layer of the most significant computing paradigm shift since the internet.
Understanding the AI and semiconductor subsector requires grasping three interconnected dynamics: the semiconductor industry's structural economics, the AI training and inference demand cycle, and the competitive landscape among chip designers, foundries, and equipment makers. Each layer of the semiconductor value chain has different margin structures, growth drivers, and competitive moats.
The Semiconductor Value Chain
The semiconductor industry is vertically segmented into four layers: design, fabrication (foundry), equipment, and materials. Each layer has different economics and competitive dynamics.
Chip designers create the intellectual property that defines what a chip does. Nvidia designs GPUs. AMD designs CPUs and GPUs. Qualcomm designs mobile processors. Broadcom designs networking and custom chips. These companies are "fabless," meaning they design chips but outsource manufacturing to foundries. The fabless model has low capital intensity (capex typically 3-8% of revenue) and high gross margins (55-75%) because the value is in the design, not the manufacturing.
Foundries manufacture chips designed by fabless companies. TSMC dominates this layer, producing over 60% of the world's semiconductors by revenue and approximately 90% of the most advanced chips (5nm and below). Samsung and Intel Foundry Services are the only other companies capable of manufacturing leading-edge chips. The foundry business requires extraordinary capital investment: TSMC spent $36 billion on capex in 2024, and its Arizona fab complex will cost approximately $65 billion. In exchange for this massive investment, TSMC earns gross margins of 53-57%, reflecting its technological monopoly at the leading edge.
Equipment makers supply the machines used to manufacture chips. ASML is the sole manufacturer of extreme ultraviolet (EUV) lithography systems, which are required for chips at 7nm and below. Each EUV machine costs approximately $380 million and takes 18 months to build. Applied Materials, Lam Research, KLA Corporation, and Tokyo Electron supply the deposition, etch, inspection, and other process equipment. The equipment layer has oligopolistic competition (2-3 major players per equipment type) and gross margins of 45-55%.
Materials companies supply the silicon wafers, specialty gases, photomasks, and chemicals used in chip fabrication. This layer is less concentrated and has lower margins but benefits from the same volume growth as the rest of the supply chain.
GPU Economics and Nvidia's Dominance
Nvidia's dominance of the AI chip market is the defining feature of the current semiconductor cycle. The company's data center revenue grew from $3 billion in fiscal year 2021 (ending January 2021) to approximately $115 billion in fiscal year 2025, a nearly 40x increase in four years. This growth was driven by the explosion in demand for GPU computing to train and run large language models, image generators, and other AI systems.
Nvidia's competitive moat has three components. First, the CUDA software platform, which Nvidia began developing in 2006, provides a programming framework for GPU computing that has become the industry standard. Over 4 million developers and virtually every AI research lab in the world uses CUDA. This software ecosystem creates switching costs that go far beyond the hardware itself.
Second, Nvidia's chip architecture has consistently led the competition in performance per watt for AI workloads. The H100 GPU, launched in 2023, delivered a 6x improvement in AI training performance over its predecessor. The B200, launched in 2024, delivered another 2.5x improvement. This performance leadership allows Nvidia to charge premium prices: H100 GPUs sold for $25,000-40,000 each, depending on configuration.
Third, Nvidia's networking products (acquired through the Mellanox acquisition in 2020) connect GPUs in the massive clusters required for AI training. A single AI training cluster can contain tens of thousands of GPUs, and the networking fabric that connects them is as important as the GPUs themselves. Nvidia's InfiniBand and Spectrum-X networking products generate additional revenue and deepen the competitive moat.
Nvidia's data center gross margin has been approximately 73-78%, extraordinary for a company with this revenue scale. The combination of high margins, explosive growth, and a competitive position that shows no signs of eroding has produced the largest single-stock wealth creation event in market history.
The AI Capital Expenditure Cycle
The demand for AI chips is driven by capital spending from a small number of hyperscale cloud providers and large enterprises. Amazon, Microsoft, Google, Meta, Oracle, and a handful of other companies account for the majority of AI chip purchases. These companies are building AI infrastructure for two purposes: training (developing AI models) and inference (running AI models in production).
Hyperscaler capital expenditure has accelerated dramatically. Amazon's capex reached approximately $75 billion in 2024. Microsoft's exceeded $55 billion. Google's exceeded $50 billion. Meta's exceeded $35 billion. A significant and growing share of this spending goes to AI-specific infrastructure, primarily GPUs, networking equipment, and the data centers that house them.
The sustainability of this capital expenditure cycle is the most debated question in the technology sector. Bulls argue that AI will generate trillions of dollars in economic value, justifying years of infrastructure investment. Bears argue that the current spending rate exceeds the near-term revenue opportunity from AI products and services, creating the conditions for an eventual correction when returns disappoint.
Historical parallels are instructive but imperfect. The build-out of fiber optic networks in the late 1990s produced a massive telecommunications capital spending cycle that ended in overcapacity and widespread bankruptcy. But the underlying demand for bandwidth was real and eventually exceeded even the most optimistic projections; it just took longer than the market expected. The railroads of the 1860s followed a similar pattern: overbuilding, financial crisis, and eventual validation of the investment thesis over a longer time horizon. This concept ties directly to Communication Services - Media and Platform Economics.
For semiconductor investors, the key question is not whether AI demand is real (it clearly is) but whether the current rate of capital spending is sustainable and whether the competitive landscape will allow current leaders to maintain their margins. If hyperscaler capex plateaus or declines, Nvidia's revenue growth will decelerate sharply, and the stock's high multiple will compress.
AMD, Intel, and Custom Silicon
AMD has emerged as Nvidia's primary competitor in the AI chip market. The company's MI300X GPU, launched in 2024, offered competitive performance at a lower price point. AMD's data center GPU revenue grew from near zero to approximately $5 billion in 2024. While this is a fraction of Nvidia's data center revenue, it represents a meaningful second source for hyperscalers seeking to reduce their dependence on a single supplier.
Intel's position in AI has been weaker. The company's Gaudi AI accelerator has gained limited market share, and Intel's broader financial challenges (declining PC market, manufacturing technology delays, massive restructuring costs) have constrained its ability to invest in AI chip development. Intel's turnaround under CEO Pat Gelsinger focused on rebuilding manufacturing capability through its IDM 2.0 strategy, but the company faces a multi-year path to competitiveness.
Custom silicon represents a growing competitive threat to merchant GPU vendors. Google's Tensor Processing Units (TPUs), Amazon's Trainium and Inferentia chips, and Microsoft's Maia chips are designed specifically for each company's AI workloads. These custom chips can be optimized for specific model architectures and offer lower total cost of ownership for their internal workloads. Broadcom and Marvell Technology are the primary designers of custom AI chips for hyperscalers, and both have seen their data center revenue surge as custom silicon adoption increases.
The long-term competitive landscape will likely include a mix of Nvidia's general-purpose GPUs, AMD's alternative GPUs, and custom chips designed for specific workloads. This diversification of the AI chip supply chain is healthy for the industry but could pressure Nvidia's pricing power and margins over time.
Semiconductor Cyclicality
Despite the AI growth overlay, semiconductors remain a cyclical industry. The semiconductor cycle is driven by inventory dynamics: when end-market demand is strong, chip companies and their customers build inventory, amplifying the upcycle. When demand softens, excess inventory must be worked down, amplifying the downcycle. The cycle typically lasts 3-5 years from peak to peak.
The most recent downcycle hit memory chips (DRAM and NAND) and analog semiconductors in 2023, while AI-driven demand kept GPUs and data center chips in an upcycle. This bifurcation was unusual. Typically, the semiconductor cycle affects most subsectors simultaneously. The AI demand wave has partially decoupled data center chips from the broader cycle, but this decoupling may not persist if macroeconomic conditions weaken enough to reduce overall IT spending.
Memory companies (Samsung, SK Hynix, Micron) have benefited from AI demand because training clusters require enormous amounts of high-bandwidth memory (HBM). SK Hynix, as the leading HBM producer, saw its stock price more than double in 2024. Micron's HBM revenue grew rapidly as well. HBM pricing has been strong due to limited supply and intense demand, but memory remains fundamentally cyclical, and the inevitable capacity additions will eventually create a supply glut.
Valuation Approaches
Semiconductor companies are valued on P/E, EV/EBITDA, and price-to-sales, with the appropriate metric depending on the company's growth rate and cyclical position.
For Nvidia, the P/E ratio must be evaluated against the growth rate. Nvidia trading at 35-45 times forward earnings with 50%+ revenue growth has a PEG ratio below 1.0, which by historical standards is not expensive. But the growth rate is decelerating, and the terminal growth rate will eventually settle at a much lower level. The debate is over when that deceleration occurs and what the steady-state margin and growth profile look like.
For cyclical semiconductor companies (Micron, Texas Instruments, NXP), the P/E ratio should be evaluated against normalized or mid-cycle earnings, not current earnings. Buying a memory company at 8x peak earnings is expensive. Buying it at 30x trough earnings may be cheap.
For TSMC and the equipment makers (ASML, Applied Materials), the combination of secular growth trends and oligopolistic market positions justifies premium multiples. TSMC at 20-25x forward earnings, given its technological monopoly and AI-driven growth, has historically been well-valued. ASML, with its literal monopoly on EUV lithography, commands a similar premium.
Equipment and Materials: The Picks and Shovels
The semiconductor equipment and materials layer offers a different risk-return profile than chip designers and foundries. Equipment companies sell into every major chip manufacturer, providing diversified exposure to the semiconductor industry without the concentration risk of betting on a single chip architecture or end market.
ASML's position is unique in the history of industrial equipment. The company is the sole manufacturer of EUV lithography systems, which are required for manufacturing chips at 7nm and below. Each EUV system contains over 100,000 parts, 457 individual lasers firing 50,000 pulses per second to create a plasma that emits ultraviolet light at 13.5nm wavelength. The physics and engineering challenges of EUV are so extreme that after 30 years of development and over $10 billion in cumulative R&D investment, no other company has been able to replicate the technology.
ASML's monopoly translates into extraordinary pricing power. EUV systems sell for $380 million each, with the newest high-NA EUV systems expected to sell for over $400 million. The company's gross margin exceeds 50%, and its backlog provides multi-year revenue visibility. For investors, ASML represents a unique asset: a literal monopoly on the enabling technology for the most important industry of the 21st century.
Applied Materials and Lam Research hold oligopoly positions in deposition and etch equipment, respectively. These companies benefit from the same industry growth trends as ASML but face more competitive pressure because their technologies, while complex, are not monopolistic. Both companies earn gross margins of 45-48% and have grown revenue at approximately 10-15% annually over the past decade.
Geopolitical Risk in Semiconductors
The semiconductor industry is at the center of geopolitical tension between the United States and China. The U.S. government's export controls, first imposed in October 2022 and subsequently expanded, restrict the sale of advanced semiconductor equipment and chips to Chinese entities. ASML cannot sell its EUV systems to Chinese customers. Nvidia cannot sell its most advanced AI GPUs to China. Applied Materials, Lam Research, and KLA face restrictions on selling certain equipment to Chinese fabs.
These restrictions have created revenue headwinds for affected companies. China represented approximately 25% of semiconductor equipment revenue before the restrictions. The affected companies have partially offset lost Chinese revenue with growth in other markets, but the net impact has been a reduction in their addressable market.
For investors, geopolitical risk in semiconductors is not a temporary concern. The strategic importance of semiconductors means that export controls and industrial policy will remain significant factors for the foreseeable future. Companies with high China revenue exposure (semiconductor equipment makers, memory companies) face the most direct risk. Companies with minimal China exposure (most fabless design firms, which sell chips to end customers outside China) face less direct impact.
The CHIPS Act and similar legislation in Europe, Japan, and South Korea represent the other side of geopolitical semiconductor policy: governments subsidizing domestic chip manufacturing to reduce dependence on concentrated supply chains, particularly in Taiwan. TSMC's Arizona fab, Intel's Ohio and Arizona expansions, and Samsung's Texas facility are all recipients of government incentives that will reshape the geographic distribution of semiconductor manufacturing over the next decade.
The semiconductor subsector is the most dynamic and rewarding area of the stock market for investors who understand the technology, the competitive dynamics, and the cyclical patterns. The AI demand cycle has created extraordinary investment opportunities, but it has also raised the stakes: the difference between correctly timing the cycle and getting caught at the peak has never been larger.
Put these principles into practice. Track fundamentals, build portfolios, and analyze stocks with AI-powered insights.
Start Free on GridOasis →