The AI Chokepoint Nobody Talks About: How TSMC Became the World's Most Systemically Important Company

Every AI chip powering the world's largest data centers — NVIDIA Blackwell, AMD MI300X, Apple M4 Pro, Google's TPU — has something in common that doesn't appear in any earnings call slide deck. They all run through one company's fabs on a 35-kilometer strip of land in Hsinchu and Tainan, Taiwan. That company is TSMC. And almost nobody in retail investing thinks about it.

The narrative around AI investing focuses on the fabless companies: NVIDIA for accelerators, AMD for alternatives, Broadcom for custom silicon. The underlying manufacturer is treated as infrastructure — invisible, assumed, reliable. That assumption deserves a closer look.

This analysis is not financial advice. Figures are estimates based on publicly available data. Always verify against the latest SEC filings before making investment decisions.

---

The Concentration: One Foundry, Every Major AI Chip

The clearest way to understand TSMC's position is to read what its customers say about it — in their own SEC filings, where legal standards demand precision.

According to NVIDIA's FY2026 10-K, filed February 25, 2026:

"Our supply chain is mainly concentrated in the Asia-Pacific region. We utilize foundries, such as Taiwan Semiconductor Manufacturing Company Limited, or TSMC, and Samsung Electronics Co., Ltd., or Samsung, to produce our semiconductor wafers."

That sentence buries the lead. "Mainly concentrated" and "such as TSMC" is corporate understatement. NVIDIA's Blackwell platform — the B200 and GB200 — is manufactured on TSMC's N4P node. There is no Samsung alternative at that process geometry for those chips.

Qualcomm's FY2025 10-K, filed November 5, 2025, is equally candid about the fabless model's structural dependency:

"QCT utilizes a fabless production model, which means that we do not own or operate foundries for the production of silicon wafers from which our integrated circuits are made. Therefore, we primarily rely on third parties to perform the manufacturing and assembly, and most of the testing, of our integrated circuits."

AMD follows the same model. Its MI300X AI accelerator, EPYC server processors, and Ryzen client chips are all manufactured on TSMC N5 or N4 nodes. Apple designs its own silicon but contracts 100% of its manufacturing to TSMC — the M4, M4 Pro, M4 Max, and iPhone 16's A18 chip are all TSMC N3E. Google's TPU v5, the accelerator powering its internal AI workloads, runs on TSMC N5/N4.

The node breakdown makes the concentration concrete:

Intel's own fabs produce some products; AI accelerators are outsourced to TSMC.

The one major chip company with its own foundry operation is Intel — and Intel's foundry business is struggling. The company recorded $18.8 billion in impairment charges in FY2024, with full-year EPS of -$4.38. Intel Foundry is not a credible near-term alternative to TSMC at advanced nodes.

The CoWoS Bottleneck

The concentration runs deeper than wafer fabrication. NVIDIA's Blackwell platform requires TSMC's CoWoS (Chip-on-Wafer-on-Substrate) advanced packaging — a process that stacks HBM3e memory directly on the compute die using TSMC's proprietary substrate technology. No other manufacturer can currently execute CoWoS at the volumes and yields Blackwell requires.

In 2024 and into 2025, CoWoS capacity was the single biggest constraint on NVIDIA's ability to ship Blackwell systems. Lead times stretched to 52 weeks. Hyperscaler orders went unfulfilled not because NVIDIA couldn't design the chips or procure the HBM memory, but because TSMC's packaging line was the bottleneck. TSMC doesn't just make the chips — it's the only place that can assemble them into the form factor the world's largest data centers need.

---

TSMC's Own Metrics: Beneficiary, Not Victim

TSMC is not a passive participant in this concentration dynamic. It is the primary beneficiary. The company's FY2025 financials reflect what happens when every major AI infrastructure build routes through a single supplier with 60% gross margins.

Key FY2025 figures (estimated at NT$/USD 32:1):

Those operating margins are extraordinary for a capital-intensive manufacturer. For context: Intel's operating margin in FY2024 was deeply negative. Samsung Foundry runs margins well below TSMC's. The combination of technological leadership, customer lock-in, and the network effect of the advanced node ecosystem gives TSMC pricing power that pure manufacturing companies almost never sustain.

TSMC is actively working to reduce geographic concentration — though the timeline matters:

• Arizona (Fab 21): Phase 1 on N4 node began production in 2024; Phase 2 on N2 is targeted for 2028. The US government provided significant CHIPS Act support.
• Japan (JASM, Kumamoto): N16/N12 production began 2024, with Sony as anchor customer and Japanese government subsidy.
• Europe (Dresden, with Bosch/Infineon/NXP): Targeting N22/N28 node production in 2027, supported by a €5 billion EU Chips Act subsidy.

The critical caveat: every fab TSMC is building outside Taiwan is on older, less sensitive nodes — N4 at the most advanced. N3, N2, and the future A16 node production stays in Taiwan through at least 2030. The geographic diversification is real; the advanced-node concentration is not resolving on any near-term timeline.

---

The Geopolitical Discount

TSMC trades at a persistent valuation discount relative to the US-listed semiconductor companies whose products it manufactures. AMD has traded at 35–45x forward earnings. NVIDIA has ranged from 30x to over 50x. TSMC's forward P/E typically sits at 20–25x — a 30–40% discount to its major fabless customers.

That discount has an informal name in analyst models: the Taiwan Strait premium. It is the probability-weighted haircut for scenarios ranging from US-China trade restriction (tighter export controls that reduce TSMC's China revenue, which represented roughly 12% of sales) to more severe disruption scenarios. Every institutional model for TSM includes this adjustment, even if the exact probability assigned varies widely.

The irony is sharp: the same characteristic that makes TSMC uniquely valuable — its irreplaceable role in AI infrastructure — also makes it the highest-stakes geopolitical asset in the global semiconductor supply chain. The more essential it becomes, the larger the potential downside if that essentialness becomes contested.

This creates a structural anomaly. You can own NVIDIA at 35x forward earnings and get indirect TSMC exposure with the geopolitical risk partially obscured. Or you can own TSMC at 22x forward earnings and get direct exposure to the same AI capex cycle — at a 30–40% multiple discount — with the geopolitical risk fully visible in the price.

---

Fair Value on TSM: Three Scenarios

Base earnings per ADR: approximately $52 USD. This is derived from TSMC's NT$334.65 EPS estimate (NT$/USD at 32:1, multiplied by the 5-share ADR ratio), which gives ~$52.3 per ADR. Current ADR price: approximately $418.

Fair values are estimated using a 10-year DCF with a 10% discount rate and 3% terminal growth. Buy-below prices reflect a 50% margin of safety to intrinsic value estimate. These are estimates, not projections.

At approximately $418, TSM has nearly converged with the base-case fair value estimate of $420 — suggesting the market has largely priced in the AI capex supercycle thesis. This is a meaningful re-rating from the $366 level where the stock traded earlier in 2026, and it removes most of the margin of safety that made TSM a compelling asymmetric opportunity at lower prices. Investors entering at current levels are essentially paying for the base case with limited buffer against a deteriorating geopolitical environment.

The bear case is worth taking seriously. An $180 fair value under the bear scenario implies roughly the same ~50% discount the market applied to Chinese tech stocks after the 2021 regulatory crackdown. It is not the base case, but it is not an implausible tail.

---

The Investment Question

Here is the question every AI bull should ask before reviewing their portfolio: do you own TSMC?

The answer is almost certainly yes — indirectly. Owning NVDA, AMD, AAPL, or QCOM means you have exposure to every cycle TSMC participates in. But you are getting that exposure at the elevated multiples of fabless companies that have outsourced the concentration risk to TSMC. You are paying a premium for the abstraction.

Owning TSM directly provides:

• 60% gross margins and 51% operating margins on the world's most essential manufacturing process
• A 30–40% valuation discount to US semiconductor peers on a forward P/E basis
• Direct exposure to AI infrastructure capex from every hyperscaler that orders NVIDIA or AMD chips
• Geographic diversification that is expanding, even if the advanced-node timeline is long

The downside is real and should not be discounted. The geopolitical risk premium embedded in TSM's multiple could widen further — not only from Taiwan-related scenarios, but from US-China export policy tightening that limits TSMC's China revenue or restricts technology transfer to its US fabs. Multiple compression can occur independently of earnings performance.

The case for TSM is not that the geopolitical risk is low. It is that the market is already pricing in a substantial discount for that risk — and that the underlying business, at 60% gross margins and 33% revenue growth, is exceptional enough that even a discounted multiple generates attractive returns in the base case.

---

Want to run your own TSM valuation or compare it against NVDA and AMD?

→ Try EvidInvest free

---

This post is for informational purposes only and does not constitute financial advice. Fair value estimates are model outputs based on publicly available data and stated assumptions. Actual results may differ materially. Always conduct your own due diligence and consult a qualified financial advisor before making investment decisions. EvidInvest holds no position in any of the securities mentioned at time of publication.