TSMC Is the Most Important Company in AI and It Has Zero AI Products

The most important company in artificial intelligence does not make a large language model. It does not have a chatbot. It has never published a research paper on transformer architecture or scaling laws or reinforcement learning from human feedback. It does not compete in benchmarks. It does not have a waitlist.

It has a $560 billion market cap, a 92% share of the most critical manufacturing process in the global economy, and a single point of geographic concentration that represents an existential risk to the entire AI industry — a risk that almost no one in the industry is seriously pricing.

Taiwan Semiconductor Manufacturing Company is the company the AI boom forgot to think about. And the longer it takes the industry to think about it, the worse the eventual reckoning will be.

The Invisible Chokepoint

Here is a simple exercise. Name the company that manufactured the Nvidia H100 GPU that trained GPT-4. Name the company that made the chips inside every MacBook, iPhone, and iPad running on-device AI. Name the company that fabricated the Google TPUs powering Gemini. Name the company that produced the AMD MI300X accelerators that Microsoft, Meta, and Oracle are deploying at scale.

Same answer every time: TSMC.

This is not a coincidence or a historical artifact. It is the structural reality of semiconductor manufacturing in 2026. Advanced chip fabrication — the production of transistors at 3, 4, and 5 nanometer process nodes — requires a combination of capital intensity, process expertise, equipment relationships, and intellectual property accumulation that took TSMC thirty years and hundreds of billions of dollars to build. No competitor has replicated it. No government program has shortcut it. And no geopolitical risk analysis has adequately priced what happens if it is disrupted.

Every company in that table is a tier-1 technology company with market caps ranging from $300 billion to $3 trillion. Every one of them has a single manufacturer for their most critical chips. That manufacturer is on an island 100 miles from China.

The Concentration Numbers

TSMC's market share figures are so extreme they read like a typo.

At process nodes below 7nm — the territory where every modern AI chip lives — TSMC holds approximately 90-92% of global production capacity. Samsung Foundry claims 6-8%, nearly all of which is consumed internally or by a small number of customers who cannot get competitive allocation from TSMC. Intel Foundry Services has shipped meaningful external volume on zero advanced nodes to date.

The trailing-edge and mature-node foundry market is competitive. GlobalFoundries, UMC, SMIC, and others provide significant capacity for chips that do not require cutting-edge process nodes — power management ICs, microcontrollers, automotive chips, analog sensors. None of this helps the AI industry, which specifically requires leading-edge nodes to achieve the transistor densities that make modern GPU and accelerator designs feasible.

The competitive map at advanced nodes looks like this:

That "struggling with 3nm yields" note next to Samsung is the key qualifier. Samsung announced its 3nm Gate-All-Around (GAA) process in 2022, ahead of TSMC's N3. The announcement was met with significant industry attention. The production reality has been significantly more complicated. Multiple customers who trialed Samsung's advanced nodes have reportedly moved workloads back to TSMC due to yield and performance issues. Samsung's foundry division reported operating losses throughout 2024 and into 2025 as it invested in process improvements. For the AI industry, Samsung is not an alternative to TSMC — it is a cautionary tale about how hard advanced node manufacturing is to execute.

The Arizona Math

In May 2020, TSMC announced it would build a semiconductor fabrication plant in Phoenix, Arizona — the first TSMC fab on US soil. The political reception was rapturous. "Made in America" chips. Semiconductor sovereignty. A hedge against the Taiwan risk everyone claimed to be worried about.

The math was always harder than the politics.

TSMC's Arizona investment has grown through successive announcements to over $65 billion committed across three planned fab buildings. Fab 1, targeting TSMC's N4 (4nm) process, went into limited production in 2024 — roughly 18 to 24 months behind the original timeline. Fab 2, targeting the more advanced N3 (3nm) process that Nvidia's next-generation GPUs will use, has slipped from 2026 to 2028-2029. Fab 3, targeting future nodes, has no firm schedule.

The delays are not primarily political or financial. They are fundamentally human.

TSMC's manufacturing process depends on a workforce culture that has been developed over decades in Taiwan. Engineers work in shifts that maximize fab uptime. Process discipline is maintained through layers of institutional knowledge that lives inside people, not documentation. When TSMC began hiring American workers for the Arizona fab, they encountered a workforce with different expectations about hours, hierarchy, and working conditions — not worse expectations, just different ones than TSMC's operational model was built around.

TSMC's leadership publicly acknowledged the cultural friction. CEO C.C. Wei cited a "lack of skilled workers" as a key challenge, prompting significant pushback from American semiconductor workers who argued the real issue was that TSMC expected workers to operate under conditions more similar to Taiwan than to US labor norms. Both things are probably true simultaneously.

The yield gap is the financial manifestation of these challenges. Industry sources have estimated that Arizona fab yields — the percentage of chips per wafer that meet specification — lagged TSMC's Taiwan baseline by 15-25 percentage points in early production runs. For context: a 20-point yield gap on a leading-edge node can represent hundreds of millions of dollars per year in lost output. TSMC is working to close this gap. It will close. But it illustrates how much tacit operational knowledge is embedded in TSMC's Taiwan operations that cannot simply be transplanted to a new geography.

Even at full planned capacity, TSMC's Arizona operations will represent approximately 10% of TSMC's total advanced-node output. The Taiwan concentration risk does not go away. It gets marginally reduced.

The CHIPS Act: Billions Spent, Years of Runway

The CHIPS and Science Act allocated $52.7 billion for US semiconductor manufacturing and research. The announcement in August 2022 was the most significant US industrial policy intervention since the postwar period. The reality of what it buys, and when, is considerably more sobering.

TSMC received $6.6 billion in CHIPS grants for its Arizona facility. Intel received $8.5 billion. Samsung received $6.4 billion for its Taylor, Texas fab. Micron received $6.1 billion for memory manufacturing expansion. In aggregate, these grants leverage roughly $450 billion in private investment commitments — a significant multiplier.

But semiconductor manufacturing does not run on announcements. It runs on operational fabs producing qualified chips at competitive yields. By that measure, CHIPS Act output in early 2026 is approximately: a single TSMC fab in Phoenix producing N4 chips at below-Taiwan yields, Intel's Ohio fab in early qualification, and Samsung's Texas facility still ramping.

The timeline reality is that meaningful CHIPS Act production output — fabs at full capacity, producing chips that major customers are actually specifying into products — is a 2028-2030 story. The political narrative implies the US is rebuilding semiconductor independence now. The manufacturing reality is that the first meaningful results are 4-6 years out, and full strategic independence in advanced semiconductors is a 10-15 year project if everything goes well.

Politicians who claim that semiconductor reshoring is on track are measuring announcements. Engineers measure wafer starts. Those numbers are very different.

Why Intel Is Not the Answer (Yet)

No narrative about semiconductor concentration is complete without addressing Intel, which spent decades as the world's most advanced chip manufacturer and is now attempting to reinvent itself as a foundry business capable of serving external customers.

Intel's foundry ambitions are real, well-funded, and probably achievable — on a timeline that does not help the AI industry in the next five years.

Intel's 18A process node, its most advanced, has been under development for years and has faced repeated delays. As of early 2026, Intel has announced a small number of external customer wins for 18A — most notably a wafer supply agreement with an undisclosed customer that analysts believe is Qualcomm or Amazon. But Intel has not yet demonstrated the sustained high-yield advanced-node production at external scale that would make it a credible alternative to TSMC for Nvidia, AMD, or Apple.

The challenge Intel faces is structural, not just technical. TSMC's foundry model — never competing with customers, focusing entirely on manufacturing, building long-term relationships with chip designers — took three decades to optimize. Intel is attempting to build a comparable model while simultaneously managing a product business, cutting costs to improve profitability, and navigating significant leadership and strategic uncertainty. These are not impossible challenges, but they are not solvable in 36 months.

The realistic scenario for Intel Foundry is that it becomes a credible, competitive option for some categories of advanced chip manufacturing by 2028-2030. It does not become a TSMC substitute. It becomes a meaningful second source for some customers, which reduces — but does not eliminate — the concentration risk.

These are optimistic timelines. Intel's track record on manufacturing schedules over the past decade suggests adding 12-18 months to every announced date as a baseline assumption.

The Financial Picture Everyone Ignores

While the AI industry lavishes attention on Nvidia's revenue growth and OpenAI's valuation, TSMC's financial profile sits largely outside mainstream AI discourse. This is a significant analytical gap, because TSMC's financials reveal both the enormous value of its position and the nature of the risk concentration embedded in it.

TSMC's 2025 revenue came in at approximately $90 billion, up roughly 35% year-over-year driven by AI chip demand. Gross margins are approximately 54-56% — semiconductor industry margins that reflect genuine monopoly pricing power on advanced nodes. Operating income margins are around 44-46%. Return on equity is consistently above 30%.

For context: TSMC's revenue from high-performance computing — the segment that includes AI chips — grew from 46% of total revenue in 2023 to approximately 52% in 2025. AI chip demand has become the primary growth driver of the world's most critical manufacturing company. This is the same company that the mainstream AI narrative treats as infrastructure, not story.

The market cap comparison is revealing. Nvidia, whose AI chip business depends entirely on TSMC manufacturing capacity, is worth roughly 5x TSMC. The company that makes the picks and shovels is worth 5x less than the company whose picks and shovels enable. The AI narrative has priced the design layer at a significant premium to the manufacturing layer — which might be correct if manufacturing were a commodity, but which looks increasingly questionable given the concentration of manufacturing in a single company on a single island.

The Unpriced Risk

Here is the question the AI industry is not asking with sufficient seriousness: what happens to AI infrastructure costs and capacity if TSMC faces a sustained disruption?

Not a catastrophic military conflict. Start smaller. A major earthquake in the Hsinchu Science Park area — where TSMC's most advanced fabs are concentrated — comparable to the 1999 Chi-Chi earthquake, which killed more than 2,400 people and caused significant factory damage. A severe typhoon that disrupts power supply to fabs that require ultra-stable electricity. A political escalation in the Taiwan Strait that stops short of military action but causes customers to begin hedging allocations and creates booking uncertainty. A single critical water supply disruption — semiconductor fabs consume enormous quantities of ultrapure water — in a region experiencing increasingly severe drought conditions.

Any of these scenarios, none of which involve military action, would stress the global AI chip supply chain in ways that current industry planning does not adequately address.

The inventory buffer in the AI supply chain is thin. Major hyperscalers — Microsoft, Google, Amazon, Meta — hold GPU inventory measured in months, not years. If TSMC's Taiwan output were curtailed by 30% for six months — a scenario far short of a complete shutdown — the ripple effects would include delayed data center builds, extended GPU lead times (already stretched to 9-12 months for H100s in recent procurement cycles), and meaningful increases in the cost of AI compute.

What does a 30% TSMC capacity disruption do to Nvidia GPU pricing? To cloud compute costs? To AI startup runway calculations? These questions have answers, and those answers should be informing risk assessments at every company whose business depends on AI infrastructure. The number of companies that have seriously modeled this scenario is very small.

The probability figures in that table are illustrative estimates, not actuarial certainties. But the distribution of outcomes — heavy tails on the downside, no upside case — is the defining characteristic of a risk that deserves a more serious pricing conversation than it is currently receiving.

Why the Market Is Not Pricing This In

The AI industry's failure to price TSMC concentration risk is not irrational given the incentive structures of the people running the analysis.

Venture capitalists are paid to find the upside. Their models are built around "what if this works?" not "what if the chip supply gets disrupted?" Portfolio companies are burning cash and need to show growth; they are not optimizing for geopolitical tail risk scenarios that have a low probability in any given year.

Hyperscalers have procurement teams that think about supply chain risk, but their GPU stockpiling behavior suggests they are hedging against allocation shortages, not manufacturing disruption — a different problem with different solutions. You cannot hedge against a TSMC Taiwan disruption by buying more GPUs ahead of time. You can only reduce your dependence on a continuously producing TSMC, which is structurally impossible given the competitive landscape.

AI researchers and executives are — correctly — focused on model capability, product development, and market share. The semiconductor supply chain is three levels of abstraction away from their daily work.

And TSMC itself, to its enormous credit, has been reasonably transparent about the concentration risk. TSMC's leadership has consistently acknowledged that their Taiwan fabs represent geographic concentration that they are working to address through Arizona and Japan investments (TSMC is also building fabs in Kumamoto, Japan, with support from the Japanese government and Sony). The company is not hiding the risk. The industry is choosing not to look at it.

What Repricing Looks Like

The scenario worth modeling is not a Taiwan military conflict, which is low-probability in any given year even given elevated tensions. It is a more gradual repricing of TSMC concentration risk driven by accumulating near-miss events, tighter US export controls affecting TSMC's ability to serve Chinese customers (which represent approximately 10-15% of revenue), or sustained geopolitical pressure that causes enterprise procurement teams to start asking questions they have not historically asked.

Once that repricing begins, several dynamics follow:

GPU costs increase. If demand for AI compute stays high but supply-side uncertainty enters the calculus, the effective cost of GPU access rises. This affects every AI company's unit economics, from the cost of a training run to the price of a cloud API call.

Alternative chip architectures gain attention. AMD, Intel, and a range of AI ASIC startups benefit from any narrative that questions Nvidia/TSMC concentration. The market for non-TSMC-dependent compute alternatives — including Intel's Ohio fabs when they mature, or domestic ASIC designs on slower nodes — becomes more strategically interesting.

Insurance and hedging instruments develop. Just as cyber insurance matured after major incidents, semiconductor supply chain risk insurance will develop as a product category. This will not be cheap, and the pricing of those instruments will serve as a real-time market signal on how institutional risk managers are assessing TSMC concentration.

Regulatory scrutiny of AI infrastructure concentration increases. The US government is already heavily involved in TSMC's US expansion. A concentration risk narrative creates political pressure for additional policy interventions, some of which may distort market outcomes in unpredictable ways.

None of these dynamics are priced into AI infrastructure costs, AI company valuations, or the strategic planning of most organizations betting heavily on continued AI capability growth.

The Uncomfortable Conclusion

The AI industry has a geological problem. Not metaphorically — literally geological. The advanced semiconductor manufacturing capacity that makes modern AI possible is concentrated in a seismically active island in one of the world's most geopolitically contested regions. The programs designed to reduce that concentration are real but will take a decade to produce meaningful results. The alternatives are years from competitiveness. And the industry is treating this as a background condition rather than an active risk.

TSMC will probably continue producing chips without a major disruption. The Taiwan Strait will probably not become an active conflict zone in the next five years. Probably.

But "probably" is not a risk management strategy. And the AI industry — which has priced Nvidia at $2.8 trillion, which is spending hundreds of billions on data center buildout, which is making multi-decade infrastructure bets — is largely operating without a serious answer to the question of what happens if the company that makes all its chips has a bad year.

TSMC is the most important company in AI. It has zero AI products. And the AI boom has essentially zero plan for what happens if TSMC cannot deliver.

That is the conversation the industry needs to start having before the conversation is forced on it.