AI Picks and Shovels (Arc 1: Forging the Brain): Taiwan Semiconductor Manufacturing (NYSE: TSM) — The Foundry

Introducing Our New “Picks and Shovels” Series

Every gold rush has two ways to profit: dig for gold, or sell the shovels. The gold, this time, is artificial intelligence. The shovels are the companies that build, connect, power, and keep alive the infrastructure underneath it.

This nine-part series — Picks and Shovels: The Anatomy of an AI Cluster — maps that stack from the ground up. Three arcs, three stocks each:

• Arc 1 — Semiconductors: the chips that get built, customised, and challenged
• Arc 2 — Networking, memory, and data centres: the wiring and storage that hold the intelligence together
• Arc 3 — Power and cooling: the unglamorous systems that keep billion-dollar compute clusters running

Nvidia designs the AI brain. A design is nothing without a factory. Before that chip can train a model or serve a query, it has to be manufactured — atom by atom — at a precision only one company can currently deliver at scale.

That is where we begin.

TSM — The Foundry

Before AI can think, it has to be built.

The Factory the World Depends On

Taiwan Semiconductor Manufacturing Company (NYSE: TSM) does not design chips. It does not brand them, market them, or sell them to end users. It takes other companies’ designs and fabricates them into physical silicon at tolerances measured in nanometres — billionths of a metre.

That sounds narrow. It is the single most important chokepoint in the global technology supply chain.

Nvidia, AMD, Apple, Qualcomm, Broadcom, and almost every other company building leading-edge processors share one constraint: they cannot get those chips made without TSMC. As of the second quarter of 2026, TSMC held about 73% of the global pure-play foundry market, according to Counterpoint Research — up from the 60–70% range still cited in a lot of older commentary. Samsung’s foundry arm sat at roughly 7%. Intel’s foundry push is real, and still not in that top tier.

The more important monopoly is at the frontier. TSMC is the only foundry producing the most advanced nodes at commercial scale with yields customers will bet a product cycle on. In Q2, 2-nanometre wafers were already 3% of wafer revenue, 3-nanometre 30%, 5-nanometre 33%, and 7-nanometre 11%. Technologies at 7nm and below accounted for 77% of wafer revenue. Samsung is still working through 2nm yield. Intel remains years behind on volume.

That is not a branding advantage. It is structural dependency — and it cuts both ways.

The Numbers Behind the Position

TSMC has just posted the best quarter in its history. It was a very good quarter. It was not a perfectly clean one.

For the second quarter of 2026, the company reported:

• Revenue: US$40.20 billion (+33.7% year on year in US dollars; +36.0% in NT dollars)
• Net profit: NT$706.56 billion (+77.4% year on year)
• EPS: US$4.31 per ADR (NT$27.25 per local share)
• Gross margin: 67.7%
• Operating margin: 60.3%
• Net margin: 55.6%

Profit grew more than twice as fast as revenue. Mix and pricing power are a large part of that. Every wafer that moves from an older node to 3nm or 2nm carries a richer margin, and the mix is moving that way. High-performance computing — primarily AI chips — accounted for 66% of Q2 revenue, up 20% quarter on quarter. In 2023, HPC was 43% of full-year revenue. The shift is structural.

It is not the whole story. Non-operating income jumped to NT$95.83 billion from NT$28.83 billion in Q1. Management said NT$63 billion of that was disposal and mark-to-market gains on Vanguard shares, contributing about NT$2.24 to quarterly EPS. Strip that out and the operating engine is still exceptional. Leave it in unremarked and the “pricing power” headline overclaims.

The forward look matters more than the print. For Q3 2026, management guided revenue of US$44.6 billion to US$45.8 billion — another record if it holds — with gross margin 65–67% and operating margin 56–58%. That implied step-down is disclosed, not a surprise: the 2nm ramp is expected to dilute gross margin by roughly 3–4 points in the second half, with overseas fabs adding further dilution over time. Full-year 2026 revenue growth was raised to slightly above 40% in US-dollar terms.

Capital expenditure was raised with it, to US$60–64 billion, from prior guidance of US$52–56 billion. About 70–80% of that budget is earmarked for advanced process technologies, about 10% for specialty nodes, and 10–20% for advanced packaging, testing, mask making and other work. That last bucket is the CoWoS story in a line item.

The spend is enormous. The cash engine still clears it. In Q2, TSMC generated NT$783 billion of operating cash flow, spent NT$496 billion on capex, and still produced NT$287 billion of free cash flow — about US$9.1 billion at the quarter’s average rate. This is a reinvestment machine, not a cash-starved one.

As of the 28 August 2026 close, TSM was US$417.52. Depending on the earnings definition used, that is roughly the high-20s on trailing earnings and the mid-20s on forward estimates. Recent analyst-target samples cluster around the mid-$500s; a figure near US$540 implies something like high-20s percentage upside from here. Treat that as a snapshot, not a house target. This piece is a map of the asset, not a discounted-cash-flow initiation.

The CoWoS Bottleneck

Understanding TSMC’s value to the AI industry requires understanding a manufacturing step most investors have never had to care about: CoWoS — Chip on Wafer on Substrate.

Modern AI accelerators, including Nvidia’s latest data-centre GPUs, are not a single chip. They are a stack of processor dies, high-bandwidth memory, and logic bridges assembled so tightly they function as one compute unit. TSMC is still the primary supplier of that packaging at the scale the AI buildout requires.

For the past two years, CoWoS capacity — not the chip design, and often not even the silicon wafer — has been the binding constraint on AI accelerator supply. Nvidia has said so publicly. Industry reporting through mid-2026 still had TSMC’s CoWoS lines sold out through the year, with monthly output heading toward about 120,000–130,000 wafers by year-end and still short of demand. Nvidia is estimated to hold around half to 60% of that allocation. AMD and the custom-ASIC crowd are taking a rising share as the pool expands.

That is the real story. TSMC is not only the world’s leading chipmaker. It is simultaneously the bottleneck and the expansion plan for the most acute supply constraint in the AI hardware stack. The capex increase is the company trying to be both at once.

Geopolitical Concentration Risk

No serious analysis of TSMC can skip the obvious: almost all of its critical leading-edge capacity still sits in Taiwan.

The geopolitical risk is real. Taiwan Strait tension is a long-standing concern, and the concentration of global advanced semiconductor production on a small island in a contested region is a structural vulnerability that investors, governments, and TSMC itself are trying to reduce. They cannot eliminate it quickly.

Diversification is underway, at extraordinary cost. The company has now pledged US$265 billion in total to its Arizona campus after an additional US$100 billion announced with the Q2 results. It is building in Japan and has explored Europe. Early Arizona output was N4/N5, then N3. The newest US money is aimed at 2nm and below, plus more packaging. That is a genuine shift from the first-generation “trailing-edge overseas” narrative.

It is still a decade-scale project. Even if Arizona eventually holds a meaningful slice of 2nm-and-beyond capacity, Taiwan will remain the centre of gravity for volume at the frontier for years, not months. Geographic diversification is real and accelerating. In the near to medium term, the concentration risk has not gone away.

Wall Street has a habit of discounting that risk until it becomes impossible to discount. This is not a reason to avoid the stock by default. It is a reason not to let TSMC become the entire AI sleeve of a portfolio. Size it as a core compounder with a tail risk you cannot model, not as a geographically diversified industrial.

Customer concentration sits underneath the same heading. TSMC does not name names in its filings. The 2025 annual report showed the largest customer at 19% of revenue and the second-largest at 17%, with the top ten at 78%. Those first two are widely understood to be Nvidia and Apple. Together they were more than a third of the company last year. The factory can make money whichever design runs through it — until one of those customers has a bad cycle at the same time as the other.

The Aussie Angle — Tax, Treaties, and the ADR Mess

Australian investors can buy TSM easily. The ADR trades on the New York Stock Exchange and is available through brokers with US market access — Stake, Interactive Brokers, CommSec International, and others.

The tax treatment is more nuanced than the usual “US stock” shorthand, and it is worth getting right.

TSMC is a Taiwanese company. Taiwan’s statutory withholding tax on dividends paid to non-residents is 21%. Australia does have a tax agreement with Taiwan: the 1996 Agreement between the Australian Commerce and Industry Office and the Taipei Economic and Cultural Office. The ATO lists Taiwan as a tax-treaty jurisdiction. Under that agreement, Taiwan’s tax on portfolio dividends is generally limited to 15% (10% only where a company holds at least 25% of the payer).

That does not mean an Australian buying the NYSE ADR automatically receives the treaty rate.

Two complications matter in practice:

1. ADR versus underlying shares. TSMC’s own filings have long warned it is unclear whether holders of ADSs are treated as owning the underlying common shares for treaty purposes. Depositary processes have often paid the dividend net of the full 21%, with any step-down toward 15% available only through a long-form reclaim — paperwork, deadlines, fees, and no guarantee of a clean result.
2. FITO still exists. A tax agreement is not the on/off switch for Australian relief. Resident investors are taxed on worldwide income and can generally claim a foreign income tax offset for foreign tax actually paid, up to the Australian tax payable on that foreign income. The treaty affects the withholding rate and the documentation path. It does not mean “no offset is available.”

Compare that with a typical US common stock, where the Australia–United States treaty and standard broker withholding are cleaner. TSM is messier, not untreatable.

The dollar stakes are also smaller than the complexity suggests. TSM’s ADR yield is about 1% forward and lower on a trailing basis — not the 1.5% often rounded into commentary. This is a capital-growth asset. Investors buying it for income are in the wrong stock. Investors buying it for the AI infrastructure thesis are at least in the right argument.

FX applies on top. Returns are in US dollars. AUD/USD at entry and exit shapes the real result. A hawkish Federal Reserve — the tone Fed Chair Kevin Warsh struck at Jackson Hole on 28 August 2026 — tends to support the USD. That can help an Australian holder on the way out and make the initial conversion more expensive. You are long the dollar as well as the foundry.

What TSM Is, and What It Isn’t

TSMC is not a bet on which AI application wins. It is not a bet on Nvidia specifically, or AMD, or any particular model architecture. It is a bet that AI chips will keep being manufactured at advanced nodes, at massive scale, and that nobody is better placed to do that work than TSMC.

That is a cleaner thesis than picking a single AI winner. The factory can make money whichever design runs through it.

It is also a reinvestment machine, not a cash cow. US$60–64 billion of annual capex is what it costs to stay at the frontier. Q2 still left NT$287 billion of free cash flow after that kind of spend. The business has to keep writing cheques of this size indefinitely if the lead is going to hold.

The risks are not decorative:

• Taiwan concentration that diversification will not fix on a retail investor’s timetable
• Customer concentration — two buyers near 36% of 2025 revenue, ten buyers at 78%
• Capital intensity that can dilute margins while new nodes and overseas fabs ramp
• Packaging and HBM constraints that can cap volume even when wafer yields are fine

None of those is automatically a dealbreaker. None of them is trivial.

The thesis starts to look wrong if leading-edge alternatives at Samsung or Intel begin shipping in volume with acceptable yields, if CoWoS and the nodes behind it are overbuilt into an AI capex pause, or if a Taiwan shock takes the physical network offline. Those are the falsification tests. Everything else is noise around a very expensive, very profitable factory.

What’s Next in This Series

Next week we move one step along the chip supply chain, to a company that takes silicon like TSMC’s and turns it into purpose-built processors for the world’s largest AI infrastructure platforms. It has become one of the most important players in the buildout not by selling general-purpose chips, but by designing exactly what its customers need and nothing else.

It is a business model with pricing power that most investors still price like a parts supplier.

Data Sources:

• TSMC Q2 2026 earnings release and prepared remarks, 16 July 2026
• TSMC Q2 2026 earnings call transcript
• TSMC 2025 Form 20-F / annual report (customer concentration: largest 19%, second 17%, top ten 78%)
• TSMC 2023 earnings materials (HPC 43% of full-year revenue)
• Counterpoint Research, global pure-play foundry share, Q2 2026
• ATO / Taipei Agreement (Australian Commerce and Industry Office–Taipei Economic and Cultural Office tax agreement, 1996)
• TSMC Form 20-F disclosures on ROC withholding tax and ADS treaty treatment
• Market prices and consensus targets as of 28 August 2026
• Industry reporting on CoWoS capacity and estimated Nvidia allocation through mid-2026
• TSMC / official announcements on the Arizona expansion to US$265 billion

Wall St. Down Under | Australia

Subscribe | wallstdownunder.com.au

Disclaimer: Wall St. Down Under is an independent financial newsletter for informational and educational purposes only. Nothing published here constitutes financial advice, a recommendation to buy or sell any security, or a solicitation of any investment decision. Always do your own research and consult a licensed financial adviser before making investment decisions. Australian investors should consider their own financial situation, objectives, and risk tolerance. Past performance is not indicative of future results.