半导体世纪：2026 AI狂飙下的投资路线图

Key Data: Global semiconductor market size (2025) approximately $792 billion · Q1 2026 sales $298.5 billion · 2026 forecast approximately $975 billion · NVIDIA FY2026 revenue $215.9 billion · TSMC Q1 2026 net profit up 58% year-over-year

1. Why Semiconductors Are More Important Than Ever

Semiconductors are the physical foundation of artificial intelligence, cloud computing, smartphones, electric vehicles, and defense systems. Every time an AI model generates a response, chips perform billions of operations in milliseconds. All of this runs on silicon.

Unlike previous cycles driven by a single device (such as smartphones or PCs), the current surge is supported by AI infrastructure spending. In 2026, the five major hyperscalers have committed to investing over $600 billion in AI infrastructure, a 36% increase year-over-year.

This fundamental shift in demand structure manifests as follows: high-value AI chips contribute approximately half of the industry's revenue but account for less than 0.2% of total shipments. Semiconductors have evolved from consumer electronics components into strategic assets for giants with market capitalizations exceeding $10 trillion.

Educational note: A modern AI chip contains billions of transistors etched onto a silicon die the size of a fingernail. The "nanometer" value of a chip represents the size of these features; a smaller nanometer number means more transistors integrated on each chip, resulting in greater computational power. The more advanced the node, the more difficult the manufacturing process required.

2. Four Core Tracks: Who Controls the Silicon Blueprint?

Investors need to discern four key roles in the supply chain, rather than lumping them together:

Designers (Architects): These companies design chips but do not manufacture them. They own the intellectual property and hand over the design blueprint to manufacturers. Since they don't operate factories, their gross margins are among the highest in the tech industry, typically exceeding 70%. NVIDIA, AMD, Qualcomm, Apple, and Broadcom are fabless companies.

Foundries (Manufacturers): Foundries perform large-scale chip manufacturing in facilities called fabs, which can cost $20 billion or more to build. TSMC holds approximately 70% to 72% of the global foundry market revenue share and produces around 90% of the world's most advanced chips at 3nm and below. Every NVIDIA Blackwell GPU, every Apple A-series processor, and every advanced AI accelerator from hyperscalers comes out of TSMC's fabs in Taiwan. This concentration means the world's most critical technology supply chain operates within a geographic area roughly the size of Belgium, located just 180 kilometers from mainland China.

Equipment Suppliers (Toolmakers): Without the machines to make chips, no chips can be made. ASML is the only company in the world capable of manufacturing extreme ultraviolet (EUV) lithography machines, which are necessary for patterning chip features at 7nm and below. Without ASML, the entire semiconductor technology roadmap would stall. Applied Materials, Lam Research, and KLA provide other critical tools for deposition, etching, and inspection processes.

Memory Makers (Storage Layer): High Bandwidth Memory (HBM) sits right next to the GPU in data center servers, feeding data to the chip at speeds no traditional memory can match. Without sufficient HBM, even the world's fastest GPU would sit idle waiting. SK hynix, Samsung, and Micron are the three main producers. HBM sales exceeded $30 billion in 2025, and total memory revenue is expected to reach approximately $200 billion in 2026.

3. Regional Dynamics: The Game and Restructuring of the Global Supply Chain

The semiconductor industry has become central to global economic security. In the current complex international environment, investors need to focus on the deep restructuring of the supply chain and policy-driven spillover effects:

Reshoring and Localization: As multiple countries implement semiconductor incentive acts, the geographic concentration of advanced manufacturing nodes is beginning to disperse moderately. The progress of TSMC's Arizona fab has become a yardstick for measuring "supply chain resilience." Early procurement agreements by giants like Apple signal that global advanced capacity is transitioning from a single region towards a multipolar distribution.

Technology Access and Market Adaptation: Strict export controls are forcing multinational chip giants to reassess their revenue structures. Companies like NVIDIA and ASML, operating within compliance frameworks, are maintaining global market share by developing customized products. This "compliance-driven innovation" is both a survival strategy for companies and a reflection of the global market's inelastic demand for high-performance computing power.

Reallocation of Computing Resources: In regions where access to computing power is restricted, the industry logic is shifting from "pursuing ultimate computing power" to "optimizing computing efficiency." Domestic leading companies and model developers are attempting to alleviate the structural contradiction in computing power supply and demand through software optimization, architectural innovation (such as in-memory computing), and deploying local alternatives in specific scenarios.

New Forms of Cross-Border Flow: Driven by the inertia of globalization, cross-border flows of computing resources take on more covert and diverse forms. Policymakers are strengthening regulation by enhancing supply chain transparency and establishing chip traceability mechanisms. For investors, this means compliance risk has become a key dimension for assessing the premium on semiconductor assets.

4. Key Companies Worth Researching

NVIDIA (NVDA)

NVIDIA is the most iconic company in the current semiconductor cycle. Its GPUs have become the default hardware for training AI models, and the CUDA software platform builds a software ecosystem moat more durable than any hardware advantage.

Key Financial Data:

• FY2026 Total Revenue: $215.9 billion, up 65% year-over-year (SEC Form 8-K, February 2026)
• Data Center Revenue: Approximately $193.7 to $194.0 billion, up 68% year-over-year
• FY2026 Q4 Revenue: $68.1 billion, up 73% year-over-year
• NVIDIA accounts for approximately 15.8% of global semiconductor market revenue
• Forward P/E Ratio: Approximately 32x

Core Issues for Investor Focus:

• The Vera Rubin platform, based on TSMC's 3nm process, packs 336 billion transistors and reduces inference costs by up to 10x compared to Blackwell. AWS, Google Cloud, Microsoft Azure, and Oracle Cloud have all committed to deployment. NVIDIA has secured the majority of HBM4 supply from SK hynix and Samsung.
• The depth of the CUDA moat is beyond most investors' perception. Millions of developers have written AI software based on CUDA. Switching to a competitor's chip means rewriting years of accumulated code, creating massive switching friction.
• Google, Amazon, and Microsoft are building their own internal chips to reduce dependence on NVIDIA, representing the most significant long-term structural risk.
• Export controls on China represent one of the most significant implicit revenue pressures among current tech companies.

TSMC (TSM)

TSMC is both the world's most critical and geographically concentrated technology supply chain node.

Key Financial Data:

• 2025 Revenue: Approximately $122.5 to $122.9 billion, up approximately 31% to 36% year-over-year
• Q1 2026 Net Profit: Up 58% year-over-year, a record high for the fourth consecutive quarter
• Q2 2026 Revenue Guidance: $39.0 to $40.2 billion
• FY2026 Capital Expenditure: $52.0 to $56.0 billion
• Q1 2026 Wafer Revenue: 74% came from advanced process nodes of 7nm and below
• Forward P/E Ratio: Approximately 24x

Core Issues for Investor Focus:

• TSMC is the most direct beneficiary regardless of which AI chip company wins spending. It represents a capacity-oriented infrastructure bet on the entire AI theme, rather than a directional bet on a specific winner.
• Geopolitical risk premium explains TSMC's valuation discount relative to NVIDIA and Broadcom, despite having comparable or even stronger revenue growth. Investors must actively judge: whether a 24x forward P/E ratio reasonably reflects the risk of a scenario that has never occurred.
• Diversification into Arizona is real but currently limited in scale. The second fab is expected to begin 3nm production by the end of 2026, with Apple's chip procurement agreement providing early commercial validation.

ASML (ASML)

ASML is the world's only company capable of manufacturing EUV lithography machines. Without these machines, chips below 7nm cannot be made; without these chips, there is no advanced AI.

Core Issues for Investor Focus:

• ASML's EUV monopoly is the culmination of decades of accumulated expertise in physics, optics, and precision mechanical engineering. No other company is close to developing a comparable machine; this moat cannot be replicated in the short term.
• Every new fab built globally, whether supported by the CHIPS Act, Japan's semiconductor investment plans, or TSMC's expansion, represents demand for ASML's equipment.
• Export restrictions on China have somewhat compressed its addressable market, and as long as the current geopolitical environment persists, these restrictions will continue.
• The long-term backlog of orders provides ASML with rare revenue visibility, with customers needing to place orders years in advance – a rarity among most tech companies.

AMD (AMD)

AMD is NVIDIA's most substantial competitor in AI accelerators, benefiting from the same TSMC foundry relationship as NVIDIA and attracting hyperscalers looking to diversify supplier dependence.

Key Financial Data:

• Quarterly sales of the downgraded MI308 (approved for export to China) reached $390 million
• Data Center GPU Revenue Guidance: 60% compound annual growth rate over the next five years

Core Issues for Investor Focus:

• The bull case rests on hyperscalers' need for supplier diversification. No major tech company wants to rely entirely on a single chip supplier; NVIDIA's market dominance itself creates a structural incentive to adopt AMD as a second source.
• AMD's ROCm software platform is its most critical challenge. While it has made significant progress, it still lags behind CUDA in developer adoption. Bridging the software gap is even more important than bridging the hardware gap.

Broadcom (AVGO)

Broadcom specializes in designing custom AI accelerators (ASICs) for hyperscalers – chips optimized for specific workloads rather than general-purpose GPUs. The TPUs that Google uses throughout its AI product ecosystem are chips designed by Broadcom.

Key Financial Data:

• FY2026 AI Semiconductor Revenue Expected to Exceed $30 billion
• Forward P/E Ratio: Approximately 41x, the highest among major semiconductor companies

Core Issues for Investor Focus:

• As hyperscalers scale up their AI deployments, custom chips optimized for specific workloads will become increasingly attractive. Broadcom's partnerships with Google and Meta are deep and durable, giving it a leading position in the custom chip space.
• A 41x forward P/E ratio requires Broadcom to maintain strong execution. Any slowdown in hyperscaler custom chip orders will have a significant impact at this valuation level.

SK hynix

SK hynix leads the HBM market with a share of approximately 53% to 62%. Its HBM3e is the memory standard for NVIDIA's Blackwell GPU, and HBM4 will be integrated into NVIDIA's Rubin platform, with NVIDIA securing most of the HBM4 supply.

Core Issues for Investor Focus:

• HBM is the true bottleneck for AI chip deployment. Even if NVIDIA delivers every GPU on time, without sufficient HBM, these GPUs cannot run at full capacity. This gives SK hynix extraordinary pricing power in the current AI infrastructure buildout.
• SK hynix is listed on the Korea Exchange. Exposure can be gained through Korean brokerage accounts, some international brokerages, or indirectly via semiconductor ETFs.
• Memory has historically been highly cyclical. Although HBM has some natural barrier to oversupply due to special manufacturing process requirements, investors still need to understand the cyclical risk inherent in the memory sector.

5. Semiconductor ETFs

SMH – VanEck Semiconductor ETF

The most widely used semiconductor ETF, with assets under management (AUM) of approximately $46 to $47 billion, holding 26 companies covering chip designers, foundries, equipment makers, and memory producers. Major holdings: NVIDIA approximately 19.4%, TSMC approximately 11.6%, Broadcom approximately 7.7%. Expense ratio: 0.35%. Widely considered the most efficient single tool for covering the full AI semiconductor supply chain.

SOXX – iShares Semiconductor ETF

SMH's closest competitor, holding 30 companies with historical long-term returns roughly on par with SMH. Expense ratio: 0.35%. Five-year return as of 2025 was approximately 140%.

SOXQ – Invesco PHLX Semiconductor ETF

Offers roughly equivalent sector coverage to SMH and SOXX but with a significantly lower expense ratio. Expense ratio: 0.19%, the lowest among major semiconductor ETFs, making it the best option for cost-conscious investors seeking similar sector exposure.

Educational Note: When comparing ETFs, pay attention to the weighting construction methodology. SMH uses market-cap weighting with caps, ensuring NVIDIA does not become overly concentrated. Understanding how an ETF is constructed helps you know what you actually hold and how its performance might differ during sector rotations.

6. Key Risk Warnings for 2026

AI Concentration Risk. The entire industry has put all its eggs in the AI basket. If AI infrastructure spending slows due to monetization falling short of expectations, geopolitical shocks, or efficiency breakthroughs, the impact on semiconductor revenue would be direct and immediate. Deloitte explicitly lists this as a core risk, even against the backdrop of record industry revenue.

Geopolitical and Supply Chain Risk. TSMC produces approximately 90% of the world's most advanced chips in Taiwan. Any disruption to manufacturing operations in Taiwan would have an impact on the entire global tech industry that is real and difficult to overstate. Diversification into Arizona is underway, but shifting the center of gravity away from Taiwan will take years.

Export Control Policy Uncertainty. US semiconductor export controls are subject to political influences and policy changes. The current administration has maintained some controls while relaxing others, including revoking Biden-era AI diffusion rules. Future policy decisions could open new markets for US chip companies or close existing ones.

Memory Cyclicality Risk. Driven by AI-related demand, consumer memory prices rose approximately 4x between September and November 2025 and are expected to rise further by up to 50% in early 2026. Deloitte warns that memory capacity expansion could trigger oversupply and a price crash by late 2026 or 2027. Markets that overextend on the upside often do so on the downside as well.

Valuation Risk. Forward P/E ratios of approximately 32x for NVIDIA and 41x for Broadcom embed very high growth expectations. A single quarter of revenue falling short of expectations, guidance cuts, or a shift in market sentiment could trigger a sharp stock price decline, even if the underlying business remains sound.

7. Key Catalysts to Watch

The One Trillion Dollar Milestone. Q1 2026 semiconductor sales reached $298.5 billion, making the annual target of $975 billion to $1 trillion a tangible possibility. Whether momentum can be sustained in the second half of the year, or whether a slowdown in AI spending leads to a weaker end of the year, is the core issue garnering the most attention across the sector.

TSMC Arizona Fab Ramp-Up. The second Arizona fab is scheduled to begin production of 3nm chips by the end of 2026. Yield and output volume will determine the speed at which the US reduces its reliance on Taiwanese manufacturing; Apple's chip procurement agreement provides the first meaningful commercial validation.

NVIDIA Vera Rubin Platform Deployment. The promise of a 10x reduction in inference costs is NVIDIA's most important product milestone. Successful deployment by hyperscalers will significantly extend NVIDIA's data center revenue growth curve; any delays or performance shortfalls would be a major negative catalyst.

AMD Market Share Progress. AMD's MI350 and MI400 products are expected to launch in 2026, testing whether its ROCm software improvements are sufficient to attract large-scale deployments from hyperscalers, rather than remaining at the current pilot project stage.

Memory Pricing and HBM4 Supply. The integration of HBM4 with NVIDIA's Rubin platform creates new demand pull. Tracking SK hynix's HBM4 production yield, along with Samsung and Micron's progress in HBM4 product qualification, will be key signals for gauging memory pricing dynamics in 2027.

Framework for Researching this Sector:

• Investors seeking the highest conviction AI chip exposure will focus their research on NVIDIA, accepting the risks inherent in export control revenue constraints and current valuation levels.
• Investors seeking AI infrastructure exposure while reducing single-stock concentration risk will research SMH or SOXX, covering the full supply chain.
• Investors who believe TSMC's geopolitical discount has become too pronounced relative to its ongoing diversification efforts may find its lower valuation multiple relative to growth worth deeper investigation.
• Investors seeking exposure to the most defensible link in the supply chain will focus their research on ASML, as every new fab built anywhere in the world creates demand for its equipment.

The demand is real. The growth is extraordinary. The risks – including geopolitical