Author: BIT US Stock Channel; Source: X, @BITstocks_CN

The chips that power artificial intelligence are reshaping geopolitics, restructuring supply chains, and driving the fastest growth in the history of the semiconductor industry.

Key data: The global semiconductor market size (in 2025) is approximately US$792 billion · Q1 sales in 2026 are US$298.5 billion · Forecast in 2026 is approximately US$975 billion · Nvidia’s fiscal year 2026 revenue is US$215.9 billion · TSMC’s Q1 net profit in 2026 increased by 58% year-on-year

1. Why semiconductors are more important than ever

Semiconductors are the material basis for artificial intelligence, cloud computing, smartphones, electric vehicles and defense systems. Every time the AI ​​model generates a response, the chip completes billions of operations in milliseconds. All of this, running on silicon.

Unlike previous cycles driven by a single device, such as a mobile phone or PC, the current surge is underpinned by spending on AI infrastructure. In 2026, the five largest hyperscalers have committed to investing more than US$600 billion in AI infrastructure, a year-on-year increase of 36%.

This fundamental shift in demand structure is reflected in the fact that high-value AI chips contribute about half of industry revenue, but account for less than 0.2% of total shipments. Semiconductors have evolved from consumer electronic components to strategic assets for giants with a market value of over US$10 trillion. Educational note: A modern AI chip contains billions of transistors etched onto a chip of silicon the size of a fingernail. The "nanometer" value of a chip represents the size of these features. The smaller the nanometer number, the more transistors integrated on each chip and the stronger the computing power. The more advanced the node, the more difficult the manufacturing process required.

2. Four core tracks: Who controls the silicon chip blueprint?

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

Designer (Architect) : This type of company designs chips but does not manufacture them themselves. They own the intellectual property and hand over the blueprints to manufacturers. Because there are no factories to operate, its gross profit margins are among the highest in the technology sector, often exceeding 70%. Nvidia, AMD, Qualcomm, Apple and Broadcom are all fabless companies.

Foundry (Manufacturer): Foundries carry out large-scale chip manufacturing in large facilities called fabs. The construction cost of a single factory is as high as 20 billion US dollars or more. TSMC accounts for about 70% to 72% of the revenue share of the overall global foundry market and produces about 90% of the world's most advanced chips at 3 nanometers and below. Every Nvidia Blackwell GPU, every Apple A-series processor, and every advanced AI accelerator from hyperscale cloud manufacturers are produced in TSMC’s wafer fab in Taiwan. This concentration means that the world's most critical technology supply chains operate within a geographical area approximately the same size as Belgium and only 180 kilometers away from mainland China.

Equipment maker (tool maker): Without the machines to make chips, you cannot make chips. ASML is the only company in the world that can manufacture extreme ultraviolet lithography machines, which are necessary to pattern chip features at the 7-nanometer node and below. Without ASML, the entire semiconductor technology roadmap would have stalled. Applied Materials, Lam Semiconductor and Kelei provide other key tools required for deposition, etch and inspection processes.

Memory vendor (storage layer): High-bandwidth memory (HBM) is placed next to the GPU in the data center server, delivering data to the chip at a speed unmatched by any traditional memory. Without enough HBM, even the fastest GPUs in the world can only idle and wait. SK Hynix, Samsung and Micron are the three major manufacturers. HBM sales will exceed US$30 billion in 2025, and total memory revenue is expected to reach approximately US$200 billion in 2026.

3. Regional dynamics: Game and reconstruction of global supply chain

The semiconductor industry has become the core of global economic security. In the current complex international environment, investors need to focus on the in-depth adjustment of the supply chain structure and policy spillover effects:

Industrial reshoring and localization: With the implementation of semiconductor incentive bills in many countries, the geographical concentration of advanced processes has begun to be moderately dispersed. The progress of TSMC's Arizona factory has become a yardstick for measuring "supply chain resilience." Early purchase agreements by giants such as Apple mark the transformation of global advanced production capacity from a single region to a multi-polar distribution.

Technology access and market adaptation: Strict export controls are forcing multinational chip giants to re-evaluate their revenue structures. Under the compliance framework, companies such as NVIDIA and ASML are developing customized products to maintain global market share. This kind of "compliance-driven innovation" is not only a survival strategy for enterprises, but also reflects the rigid demand for high-performance computing power in the global market.

Redistribution of computing resources: In areas where access to computing power is limited, industrial logic is shifting from "pursuing extreme computing power" to "optimizing computing power efficiency." Domestic leading manufacturers and model developers are trying to alleviate the structural contradiction between supply and demand of computing power through software optimization, architectural innovation (such as storage and computing integration), and the deployment of local alternatives in specific scenarios.

New forms of cross-border flows: Under the inertia of globalization, the cross-border flow of computing resources has taken on a more hidden and diverse form. Policymakers are strengthening supervision by improving supply chain transparency and establishing chip traceability mechanisms. For investors, this means compliance risk has become a key dimension in assessing premiums on semiconductor assets.

4. Key companies worth studying

NVDA

Nvidia is the most iconic company in the current semiconductor cycle. Its GPU has become the default hardware for training AI models, and the CUDA software platform has built a software ecological moat that is more durable than any hardware advantage.

Key Financial Data:

• Total revenue for fiscal year 2026: US$215.9 billion, a year-on-year increase of 65% (SEC Form 8-K, February 2026)

• Data center revenue: approximately US$193.7 billion to US$194 billion, a year-on-year increase of 68%

• Q4 revenue in fiscal year 2026: US$68.1 billion, a year-on-year increase of 73%

• NVIDIA accounts for approximately 15.8% of the global semiconductor market revenue share

• Forward P/E ratio: approximately 32 times

Core issues that investors are concerned about:

• The Vera Rubin platform is based on TSMC's 3nm process and is equipped with 336 billion transistors. The inference cost is up to 10 times lower than that of Blackwell. AWS, Google Cloud, Microsoft Azure and Oracle Cloud have all committed to deployment. Nvidia has locked in most of its HBM4 supply from SK Hynix and Samsung.

• CUDA’s moat is deeper than most investors realize. Millions of developers have written AI software based on CUDA. Switching to competing chips means rewriting years of code accumulation, creating huge migration friction.

• Google, Amazon and Microsoft each build in-house self-developed chips to reduce their dependence on Nvidia, which is the most important long-term structural risk.

• Export controls to China are currently one of the most significant hidden pressures on revenue among technology companies.

TSMC (TSM)

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

Key Financial Data:

• Revenue in 2025: approximately US$122.5 billion to US$122.9 billion, a year-on-year increase of approximately 31% to 36%

• 2026 Q1 net profit: a year-on-year increase of 58%, a record high for the fourth consecutive quarter

• 2026 Q2 revenue guidance: US$39 billion to US$40.2 billion

• Capital expenditures for fiscal year 2026: US$52 billion to US$56 billion

• 74% of Q1 wafer revenue in 2026 will come from advanced processes of 7 nanometers and below

• Forward P/E ratio: approximately 24 times

Core issues that investors are concerned about:

• TSMC is the most direct beneficiary that will benefit no matter who spends money on AI chips. It is a quantitative infrastructure target that is betting on the entire AI theme, rather than a directional bet on a specific winner.

• The geopolitical risk premium explains TSMC's valuation discount relative to Nvidia and Broadcom, even though its revenue growth is comparable to or stronger than both. Investors must proactively determine whether a forward P/E ratio of 24 times reasonably reflects the risks involved in a scenario that has never occurred.

• Arizona decentralization is real, but currently limited in scale. The second factory is expected to start 3nm production by the end of 2026, and Apple's chip purchase agreement provides early commercial verification.

ASML (ASML)

ASML is the only company in the world that can manufacture EUV lithography machines. Without these machines, it would be impossible to manufacture chips below 7 nanometers; without these chips, there would be no advanced AI.

Core issues that investors are concerned about:

• ASML’s EUV monopoly is the culmination of decades of expertise accumulated in the fields of physics, optics and precision mechanical engineering. No other company has come close to developing a similar device, and this moat cannot be replicated anytime soon.

• Every new wafer fab in the world, whether it is a project supported by the "Chip Act", Japan's semiconductor investment plan or TSMC's expansion plan, represents the demand for ASML equipment.

• Restrictions on exports to China have compressed the addressable market and will continue to exist as long as the current geopolitical environment remains unchanged.

• The long order backlog provides ASML with rare revenue visibility, as customers place orders years in advance, a rarity among most technology companies.

AMD (AMD)

AMD is Nvidia's most substantial AI accelerator competitor, benefiting from the same TSMC foundry relationship as Nvidia and is attracting hyperscale cloud players looking to spread supplier dependence.

Key Financial Data:

• MI308 downgraded version (can be exported to China upon approval) sales reached US$390 million in a single quarter

• Data center GPU revenue guidance: compound annual growth rate of 60% in the next five years

Core issues that investors are concerned about:

• The bullish logic lies in the supplier decentralization needs of ultra-large-scale cloud vendors. No large technology company is willing to rely entirely on a single chip supplier, and Nvidia's market dominance creates structural incentives for introducing AMD as a second supplier.

• AMD's ROCm software platform is its most critical challenge. While it has made great strides, it still lags behind CUDA in terms of developer adoption. Closing the software gap is more important than bridging the hardware gap.

Broadcom (AVGO)

Broadcom specializes in designing custom AI accelerators (ASICs) for ultra-large-scale cloud vendors, which are chips optimized for specific workloads rather than general-purpose GPUs. The TPU used by Google throughout its entire AI product system is a chip designed by Broadcom.

Key Financial Data:

• AI semiconductor revenue is expected to exceed US$30 billion in fiscal year 2026

• Forward price-to-earnings ratio: approximately 41 times, the highest among major semiconductor companies

Core issues that investors are concerned about:

• As hyperscale cloud vendors expand the scale of AI deployments, custom chips optimized for specific workloads will become increasingly attractive. Broadcom has a deep and solid cooperative relationship with Google and Meta, and it occupies a leading position in the field of custom chips.

• The forward price-to-earnings ratio of 41 times requires Broadcom to maintain strong execution. Any slowdown in custom chip orders from hyperscale cloud vendors will have a significant impact on this valuation level.

SK Hynix

SK hynix leads the HBM market with a market share of approximately 53% to 62%. Its HBM3e is the memory standard for NVIDIA Blackwell GPUs. HBM4 will be integrated into the NVIDIA Rubin platform. NVIDIA has locked in most of the HBM4 supply.

Core issues that investors are concerned about:

• HBM is the real bottleneck in AI chip deployment. Even if Nvidia delivers every GPU on time, these GPUs cannot operate at full capacity without enough HBM, which gives SK Hynix extraordinary pricing power in the current wave of AI infrastructure construction.

• SK hynix is listed on the Korean Exchange and can gain exposure through Korean brokerage accounts, some international brokerages, or indirectly through semiconductor ETFs.

• Memory has historically been highly cyclical. Although HBM has certain natural barriers to oversupply due to special manufacturing process requirements, investors still need to understand the cyclical risks carried by the memory sector.

5. Semiconductor ETF

SMH — Invesco Semiconductor ETF

The most widely used semiconductor ETF has a management scale of approximately US$46 billion to US$47 billion and holds 26 companies, covering chip designers, foundries, equipment manufacturers and memory manufacturers. Main holdings: Nvidia accounts for approximately 19.4%, TSMC accounts for approximately 11.6%, and Broadcom accounts for approximately 7.7%. Management fee: 0.35%. It is widely regarded as the most efficient single tool covering the entire supply chain of AI semiconductor topics.

SOXX — iShares Semiconductor ETF

SMH’s closest competitor holds 30 companies, and its historical long-term return rate is basically the same as SMH. Management fee: 0.35%. The five-year return to 2025 is about 140%.

SOXQ — Invesco PHLX Semiconductor ETF

The coverage is roughly the same as that of SMH and SOXX sectors, and the management fee is significantly lower. Management fee: 0.19%, the lowest among major semiconductor ETFs, and the best choice for cost-conscious investors to gain exposure to similar sectors.

Educational Note: When comparing ETFs, pay attention to how the weights are constructed. SMH uses capped market capitalization weighting to ensure that Nvidia does not form excessive concentration. Understanding how ETFs are constructed can help you understand what you are actually holding and how it behaves differently when sectors rotate.

VI. 2026 Key Risk Warning

AI concentration risk. The entire industry puts all its eggs in the AI ​​basket. If AI infrastructure spending slows due to lower than expected realization, geopolitical shocks or efficiency breakthroughs, the impact on semiconductor revenue will be direct and immediate. Deloitte clearly lists this as a core risk despite the industry's record revenue.

Geopolitics and supply chain risks. TSMC produces about 90% of the world's most advanced chips in Taiwan. The impact on the entire global technology industry of any kind of disruption to Taiwan's manufacturing operations is too real to overstate. Arizona's decentralized layout is moving forward, but it will still take several years to truly shift the manufacturing focus away from Taiwan.

Export control policy uncertainty. U.S. semiconductor export controls are affected by political factors and there is a risk of policy changes. The current administration has maintained some controls and relaxed other restrictions, including revoking Biden-era AI proliferation rules. Future policy decisions could open up new markets for U.S. chip companies or close existing channels.

Memory Cyclic Risk. Affected by AI-driven demand, consumer-grade memory prices have increased approximately 4 times between September and November 2025, and are expected to further increase by as much as 50% in early 2026. Deloitte warned that memory capacity expansion could trigger oversupply and price collapse by the end of 2026 or 2027. Markets that go too far on the upswing tend to go too far on the downswing.

Valuation Risk. Nvidia’s forward price-to-earnings ratios of about 32 times and Broadcom’s about 41 times have embedded extremely high growth expectations. A lower-than-expected single-quarter revenue, a guidance cut, or a change in market sentiment could trigger a sharp decline in stock prices, even if the underlying business remains solid.

7. Key catalysts worthy of attention

Trillion dollar milestone. Semiconductor sales in Q1 of 2026 will reach US$298.5 billion, making the full-year target of US$975 billion to US$1 trillion within reach. Whether the momentum can be maintained in the second half of the year, or whether AI spending will slow down and weaken at the end of the year, is the core issue that attracts the most attention for the entire sector.

TSMC’s Arizona plant capacity ramps up. The second Arizona factory will start 3-nanometer chip production at the end of 2026. Yield and output will determine how quickly the United States reduces its reliance on Taiwanese manufacturing; Apple's chip purchase agreement provides the first meaningful commercial verification.

NVIDIA Vera Rubin platform deployment. The promise of a 10x reduction in inference costs is Nvidia’s most important product milestone. The successful deployment of hyperscale cloud vendors will significantly extend Nvidia's data center revenue growth curve; any delays or substandard performance are significant negative catalysts.

AMD market share progress. AMD’s MI350 and MI400 products, expected to launch in 2026, will test whether its ROCm software improvements are enough to attract large-scale deployment by hyperscale cloud vendors, rather than just remaining in the current pilot project stage.

Memory pricing and HBM4 availability. The integration of HBM4 and NVIDIA’s Rubin platform creates new demand pull. Tracking SK Hynix's HBM4 production yield, as well as Samsung and Micron's progress in HBM4 product certification, will be key signals for judging the dynamics of memory tier pricing in 2027.

Thinking framework for studying this section:

• Investors seeking the highest degree of confidence in AI chip exposure will focus on NVIDIA and accept the risks inherent in export control revenue constraints and current valuation levels

• Investors looking to gain exposure to AI infrastructure while reducing concentration risk in individual stocks will look into SMH or SOXX, which cover the complete supply chain

• Investors who believe that TSMC's geopolitical discount is too significant relative to its ongoing diversification may find its lower valuation multiple relative to growth worthy of further investigation

• Investors looking for exposure to the most defensive segments of the supply chain will focus on ASML, as every new fab built anywhere in the world creates demand for it

The demand is real and the growth is extraordinary. Risks, including geopolitical concentration, AI demand dependence, memory cyclicality and valuation, are also real. Only investors who understand these four dimensions at the same time can examine this sector with the sobriety and thoroughness required.