TSMC: The Semiconductor Foundry Monopoly Powering the AI Revolution

The Foundry Monopoly Nobody Talks About

Nvidia gets the headlines. TSMC makes the chips.

That sentence contains one of the most underappreciated dynamics in global technology. Every single Nvidia H100 and B200 GPU is manufactured by TSMC. Every Apple M4 chip. Every AMD EPYC server processor. Every Qualcomm Snapdragon mobile SoC. Every Broadcom networking ASIC, every MediaTek 5G modem, every Amazon Graviton CPU. All of them — fabricated in TSMC's fabs in Hsinchu, Tainan, and Kaohsiung, Taiwan.

TSMC's share of advanced node manufacturing (defined as sub-7nm) sits above 90%. We want to be precise about what that means. It means there is no meaningful alternative. Samsung Foundry, TSMC's closest competitor, holds roughly 8% of advanced node capacity and has been losing share, not gaining it. Intel Foundry Services remains years away from competing at the leading edge (their 18A node is targeting 2025–2026 production, and yield ramp timelines remain uncertain). GlobalFoundries exited the advanced node race entirely in 2018.

So when investors debate whether Nvidia or AMD or Broadcom will “win” the AI chip race, they are debating which TSMC customer will capture the most value. TSMC wins regardless.

N3 and N2: The Technology Gap Widens

N3 Process Family

TSMC's N3 (3nm) process entered volume production in late 2023, initially for Apple's A17 Pro and M3 chips. By mid-2025, the N3E and N3P variants had ramped across Nvidia, AMD, Qualcomm, and MediaTek. N3 delivers approximately 30–35% power reduction and 15–20% speed improvement versus N5, with a transistor density of roughly 291 million transistors per square millimeter.

Here is what matters for investors: Samsung's competing 3nm GAA (gate-all-around) process has suffered yield problems that multiple industry sources peg at 20–30% below TSMC's N3 yields. Yield is everything in semiconductor manufacturing. A 20% yield disadvantage means Samsung must use 20% more silicon wafers to produce the same number of functional chips, which destroys economics. This is why Qualcomm, once a meaningful Samsung Foundry customer, shifted its flagship Snapdragon 8 Gen 3 production to TSMC. It is why Nvidia has never used Samsung for a flagship GPU.

N2 and A16: The Next Frontier

TSMC's N2 (2nm) process, scheduled for volume production in the second half of 2025, will be the company's first node to use gate-all-around (GAA) transistors — replacing FinFET architecture that has been the industry standard since 14nm. N2 targets 10–15% speed improvement and 25–30% power reduction versus N3E, with further density gains.

Beyond N2, TSMC's A16 node (expected 2026) will incorporate backside power delivery — routing power connections through the back of the wafer rather than the front, freeing up space for signal routing and enabling another step-function in density. This is a technology that Intel has been developing under the name “PowerVia” for its 18A node, and the race to deliver backside power in volume will be one of the defining competitive battles in semiconductor manufacturing through 2027.

We think the market underestimates the compounding effect of TSMC's node-over-node execution. Each generation where TSMC delivers superior yields and on-time production reinforces customer loyalty and makes the switching costs higher. Apple has been exclusive to TSMC since the A6 chip in 2012. Nvidia has been exclusive to TSMC for its entire data center GPU lineup. AMD moved its entire product stack to TSMC after GlobalFoundries exited advanced nodes. Once a customer builds their design team, IP libraries, and validation workflows around TSMC's process, switching foundries requires 18–24 months of redesign work and hundreds of millions in engineering costs.

Customer Dependency: Who Needs TSMC Most

Apple is TSMC's largest customer, accounting for an estimated 23–25% of total revenue. Every iPhone, iPad, Mac, Apple Watch, and Vision Pro ships with a TSMC-fabricated chip. Apple typically gets first access to TSMC's newest process nodes — a relationship built over more than a decade of exclusive partnership.

Nvidia is the fastest-growing customer relationship. The explosive demand for H100 and B200 GPUs has made Nvidia one of TSMC's top three customers by revenue (estimated 12–15% of revenue in 2025, up from roughly 7% in 2023). Nvidia's next-generation Rubin GPU architecture will continue on TSMC's most advanced nodes. The AI boom has fundamentally altered the customer mix — high-performance computing and AI now represent the majority of TSMC's advanced node demand, surpassing smartphones for the first time.

AMD relies entirely on TSMC for its EPYC server CPUs, Ryzen desktop and laptop CPUs, Radeon GPUs, and MI300X AI accelerators. Broadcom, the second-largest fabless chip company by revenue, depends on TSMC for its custom AI ASICs (built for Google, Meta, and others) and its networking silicon. Qualcomm, MediaTek, Marvell — the list goes on.

The concentration cuts both ways. TSMC depends on Apple for nearly a quarter of revenue — a single design loss would be painful. But Apple has even fewer alternatives than TSMC has customers. That asymmetry matters. TSMC can lose a mid-tier customer and fill the capacity with one of dozens of other fabless companies waiting for allocation. Apple cannot switch foundries without a multi-year, multi-billion-dollar redesign effort with no guarantee Samsung or Intel could match TSMC's yields.

CoWoS: The AI Bottleneck Nobody Expected

Advanced packaging has become the surprise constraint of the AI era. Not transistor density. Not lithography. Packaging.

TSMC's CoWoS (Chip-on-Wafer-on-Substrate) technology enables the integration of GPU dies with HBM (High Bandwidth Memory) stacks on a single silicon interposer. Every Nvidia H100, every B200, every GB200 NVL72 rack requires CoWoS packaging. The interposer must be large enough to accommodate the GPU die plus 6–8 HBM stacks — resulting in package sizes exceeding 100mm x 100mm, which are enormous by semiconductor standards.

Through 2023 and most of 2024, CoWoS supply was the binding constraint on Nvidia GPU shipments. Not wafer fabrication. Not HBM supply from SK Hynix. CoWoS. TSMC has been scrambling to expand capacity — roughly doubling CoWoS output in 2024 and guiding for another 60–80% increase in 2025. But demand continues to outstrip supply, with lead times at 6–9 months for new orders.

Why does this matter for TSMC's stock? Three reasons. First, CoWoS carries higher margins than standard wafer fabrication because the technology is proprietary and customers have limited alternatives (Samsung's competing I-Cube technology has much less volume). Second, CoWoS revenue is almost entirely driven by AI demand, making it the purest AI exposure within TSMC's business. Third, as AI chip architectures evolve toward larger interposers (Nvidia's next-gen designs are rumored to require even bigger CoWoS packages), the revenue per package increases with each generation.

The $65 Billion Arizona Question

TSMC's Arizona expansion is the most consequential fab buildout outside of Asia in a generation. Three fabs. $65 billion-plus in total investment. The largest foreign direct investment in American history.

Fab 1 (N4 process, roughly equivalent to the chips in current iPhones) is expected to begin production in 2025, after multiple delays from the original 2024 target. Fab 2 will run N3 or N2 processes, targeting 2028 production. Fab 3, announced in late 2024, will push to N2 or beyond and is expected by 2030. The CHIPS Act provides approximately $6.6 billion in direct grants and $5 billion in government loans — significant but covering only a fraction of total costs.

The bear case on Arizona is straightforward: it is expensive. Semiconductor manufacturing in the US costs an estimated 20–30% more than in Taiwan. Labor costs are higher. Construction costs are higher. The supply chain for chemicals, gases, and photomask materials is less developed. TSMC has publicly acknowledged these cost challenges and has experienced workforce issues — the initial plan to use primarily American construction workers was supplemented by flying in thousands of experienced Taiwanese workers, creating political friction.

Our view? Arizona is margin-dilutive in the medium term (2025–2028) but strategically essential. It secures TSMC's position with the US government and defense industry (which cannot rely on Taiwan-based production for classified chip programs). It provides a hedge against Taiwan geopolitical risk for TSMC's largest customers. And it may command premium pricing — customers willing to pay 10–15% more for “made in USA” chips, particularly for defense, automotive safety, and critical infrastructure applications.

Geopolitical Risk: The Taiwan Question

We cannot write about TSMC without addressing it directly. Taiwan is 100 miles from mainland China. Beijing considers it a breakaway province. A military confrontation — whether invasion, blockade, or coercive actions short of war — would disrupt 90% of the world's advanced chip supply.

The economic consequences would be staggering. A RAND Corporation study estimated a Taiwan conflict could reduce global GDP by $2.5 trillion in the first year. Every smartphone, every server, every car, every military system that depends on advanced chips would face immediate supply disruption. There is no substitute capacity. Not Samsung. Not Intel. Not the Arizona fabs (which won't have meaningful volume until 2026–2027 at the earliest).

But here is the contrarian take we hold: the market has priced Taiwan risk into TSMC's stock for decades, and it consistently overprices it relative to the actual probability-weighted impact. TSMC trades at 20–22x forward earnings. Nvidia trades at 28–35x. Broadcom trades at 25–30x. The P/E gap between TSMC and its customers is roughly 10–15 turns — an implied “Taiwan discount” of 30–40%.

Is a 30–40% discount appropriate? Only if you believe a Taiwan conflict is both likely and imminent. We do not. The “silicon shield” theory has merit: China itself depends on TSMC (and by extension, chips designed by Apple, Qualcomm, and others) for its domestic technology ecosystem. A military action against Taiwan would destroy China's own chip supply. That is a meaningful deterrent, even if imperfect. US military deterrence, TSMC's geographic diversification to Arizona and Japan, and the sheer economic cost to all parties make a near-term conflict a low-probability (though high-impact) tail risk.

We size the position accordingly. TSMC is not a bet-the-farm conviction position. But the valuation discount for Taiwan risk gives investors a margin of safety that does not exist in Nvidia or Broadcom at current multiples.

Financial Profile: The Numbers Behind the Moat

Several things jump out from these numbers. First, TSMC's gross margins have been steadily expanding — from 53% in 2022 to 56%+ in 2024 — despite massive capex spending. This reflects TSMC's pricing power. When you are the only manufacturer capable of producing the world's most important chips, you set the price. TSMC has implemented multiple rounds of price increases over the past three years (estimated 5–10% average ASP increases at leading-edge nodes), and customers have accepted them because they have no alternative.

Second, the $30B+ annual capex is not a burden — it is the moat. No other foundry can match this level of investment. Samsung Foundry's semiconductor capex was approximately $10–12 billion in 2024 (across memory and logic). Intel's foundry-related capex, while significant in absolute terms, has been spread across process development and existing fab upgrades rather than greenfield advanced node expansion. TSMC's capex advantage compounds: more investment leads to better yields, which leads to more customers, which generates more revenue to fund even more capex.

Third, the forward P/E. We keep coming back to this. At 20–22x forward earnings for a company growing revenue 20%+ with expanding margins and a monopoly position — that is cheap. It is cheap relative to the company's own quality. It is cheap relative to its customers. It is cheap relative to the broader semiconductor sector (SOX index trades at 25–28x forward). The Taiwan discount is real, but we believe it creates opportunity rather than reflecting fair value.

The $30B+ Capex Moat: Why Nobody Can Catch TSMC

Semiconductor manufacturing has the highest barriers to entry of any industry on Earth. A single leading-edge fab costs $15–20 billion. It takes 3–5 years to build. It requires ASML EUV lithography machines that cost $150–380 million each (and have a two-year delivery backlog). It requires thousands of engineers with highly specialized expertise in semiconductor physics, chemical engineering, and process integration.

TSMC spent $30.5 billion on capex in 2023 and $29.8 billion in 2024. The 2025 capex budget is guided at $32–36 billion. Over the 2021–2025 period, TSMC will have invested approximately $150 billion in manufacturing capacity and process development. To put that number in context: it exceeds the entire GDP of 130 countries.

This is the moat. Not patents (which expire). Not software ecosystems (which can be replicated given enough time). Physical manufacturing infrastructure that took decades to build and costs tens of billions annually to maintain and expand. Intel learned this lesson the hard way: despite spending $25–30 billion annually on capex, Intel Foundry remains at least two years behind TSMC on process technology. Money alone is not enough — you also need the accumulated process know-how, the yield engineering expertise, and the customer relationships that TSMC has compounded over 30+ years.

Bull Case vs. Bear Case

The Bull Case

AI is a structural demand driver that extends TSMC's growth runway well beyond the current cycle. HPC (high-performance computing, which includes AI accelerators) has overtaken smartphones as TSMC's largest revenue segment, growing from 33% of revenue in 2022 to over 50% in 2025. Unlike smartphones (which face cyclical replacement cycles and market saturation), AI infrastructure demand is in its infancy. Hyperscaler capex continues to accelerate — Microsoft, Google, Amazon, and Meta collectively guided over $250 billion in 2025 capital expenditure, with the majority directed at AI infrastructure. Every dollar of that spending translates to chip orders at TSMC.

Margin expansion has room to run. TSMC's price increases are sticking, and the mix shift toward advanced nodes (which carry higher ASPs and margins) is a multi-year trend. If gross margins reach 60% by 2027 (from 56% in 2024), the earnings leverage is substantial on a revenue base growing 20%+ annually. That combination — revenue growth plus margin expansion — could drive EPS growth of 25–30% annually through 2028.

Multiple expansion is the cherry on top. If the Arizona fabs execute well and geopolitical tensions do not escalate, the Taiwan discount could narrow from 30–40% to 15–20%. On a 2027 EPS of approximately $9–10 per ADR share, multiple expansion to 25x would imply a stock price of $225–250 — roughly 40–60% above current levels, depending on where you enter.

The Bear Case

Taiwan risk is the obvious one, and we have covered it. But the more probable bear case is a cyclical demand correction. Semiconductors are cyclical. Full stop. The AI capex boom has been extraordinary, but history shows that periods of aggressive infrastructure buildout are followed by digestion phases. If hyperscalers pause or reduce chip orders — even temporarily — TSMC's revenue growth could stall, and the stock would de-rate quickly. The 2022–2023 downcycle, when TSMC's revenue declined 4.5% and the stock fell over 30%, is a recent reminder.

Customer concentration is a real risk. Apple and Nvidia together represent roughly 35–40% of revenue. A design change at Apple (unlikely but possible if Intel or Samsung ever close the technology gap) or a demand slowdown at Nvidia would have outsized impact.

Arizona margin dilution is quantifiable. If Arizona fabs operate at 20–30% higher cost and represent 10–15% of total capacity by 2030, the blended gross margin drag could be 100–200 basis points. Not catastrophic, but meaningful for a stock that the bull case depends partly on margin expansion.

Finally, there is the long-tail risk of foundry competition. Intel under new CEO Lip-Bu Tan is refocusing the company, and Intel 18A could (if yields cooperate) attract a few design wins in 2026–2027. Samsung continues to invest heavily. Neither is likely to dent TSMC's 90%+ share in the next three years, but the five-to-ten-year outlook is less certain.

How We Would Size the Position

We think TSMC belongs in most growth-oriented portfolios. The question is sizing.

For investors who already own Nvidia, AMD, or Broadcom: TSMC is the common denominator. Owning the foundry alongside the fabless customers provides a natural hedge — if Nvidia loses share to AMD, TSMC fabricates both. If custom ASICs from Google and Amazon displace merchant GPUs, TSMC fabricates those too. A 3–5% portfolio allocation to TSMC complements a broader semiconductor position without excessive overlap.

For investors specifically concerned about Taiwan risk: limit the position to 2–3% and pair it with semiconductor equipment stocks (ASML, Lam Research, Applied Materials), which benefit from TSMC's capex but are not exposed to Taiwan geopolitical risk.

For deep-value investors: the 20–22x forward P/E on a company with TSMC's quality characteristics is rare. If you can stomach the geopolitical risk (and we believe the probability-weighted expected return favors the long side), a 5–7% position is justified. TSMC's earnings growth trajectory means the effective P/E on a two-year forward basis is closer to 16–18x — bargain territory for a monopoly.

Frequently Asked Questions