NVIDIA vs TSMC: A HK/TW Investor's Honest Take

TL;DR

• I'm Jim Liu, founder of LowRiskTradeSmart. I have held both NVDA and TSM ADR in my US-stock account since mid-2024.
• The "NVIDIA vs TSMC" framing is mostly wrong. NVDA designs the chips, TSMC builds them. They are supplier and customer, not direct competitors.
• For HK/TW investors, the 30% US withholding tax on NVDA dividends barely matters (yield ~0.03%). TSM ADR pays ~1.4%, but its dividend gets ~21% Taiwan withholding, not 30%.
• If I had to pick one: TSM for income and capacity moat, NVDA for growth and pricing power. I keep about 60/40 NVDA/TSM by exposure, mostly because Blackwell still ramps.
• Below: the dependency map, the 2024 Blackwell stress test, and three things I got wrong.

Who I Am: A NVDA + TSM Holder Since 2024

I run LowRiskTradeSmart from Sydney. The site tracks Hong Kong IPOs and US stocks for HK/TW retail investors. I have written most of the guides on it myself, including the Intel vs TSMC investor guide published yesterday and the NVDA + AVGO AI infrastructure pair piece from Saturday.

My broker is moomoo HK. My current position sizing across these names: NVDA is roughly 4.5% of my US sleeve, TSM ADR is about 3%, Intel is a 1% probe. I hold them through a normal US-stock account, not via Hong Kong–listed semiconductor ETFs (those exist on HKEX, but the management fees ate too much of the dividend math for me to bother).

I share this because every claim below comes from positions I actually hold and trades I actually placed, not a desk note someone passed me.

The Vertical Dependency Map: NVDA Designs, TSMC Builds

Most "nvidia vs tsmc" search results I see lump these two into a head-to-head comparison. That misses what is actually going on with the supplier-customer relationship at the core of the AI chip market.

NVIDIA does not own a single fab. Every Blackwell GPU, every H100, every consumer RTX card — all of it rolls off a TSMC line. Industry estimates put roughly 80%+ of NVIDIA's advanced silicon volume on TSMC's 4N and 3nm processes. The remaining sliver goes to Samsung Foundry for some lower-end SKUs.

On the other side, TSMC's revenue is concentrated in a handful of "whale" customers. Apple sits at the top (~25% by most estimates). NVIDIA has moved into the second slot, somewhere around 20% of total revenue, up from roughly 10% in 2023. Together with AMD and Qualcomm, the top four customers probably make up over 60% of TSMC sales.

In other words: NVIDIA is TSMC's biggest growth customer, and TSMC is NVIDIA's only realistic foundry at the leading edge. That is not competition. That is mutual dependency, with a knife to each other's throats.

Why "vs" Is the Wrong Question

If you search "nvidia vs tsmc", you usually want one of three answers: (a) which stock returns more, (b) which has better margins, or (c) which is safer to own through an AI cycle.

The honest answer is none of those map cleanly because the two companies sit at different points in the same chain. If AI compute demand stays strong, both win. If demand softens, both lose, but TSMC loses less because Apple and AMD orders cushion the fall. If a black-swan supply shock hits TSMC, NVIDIA gets crushed harder because there is no Plan B at 3nm.

So the better questions are: how much downside protection do I want, and how much pricing-power upside do I want? That is what the rest of this piece tries to answer.

TSMC's NVDA Concentration Risk: ~20% Revenue, One Customer

TSMC's published Q3 2024 management commentary flagged HPC (high-performance computing, mostly NVIDIA) as the largest contributor to growth. Smartphone share has been falling for two years. If NVDA orders dropped 30% — say, after a CSP (cloud service provider) capex pause — TSMC revenue would take a 6-7 point hit. Margins would compress more because high-end nodes have the best fall-through.

The risk is not that NVIDIA stops buying. It is that the buying gets lumpy and pricing leverage shifts. NVIDIA already pushed for, and won, capacity guarantees on TSMC's 2nm in 2025. That is the kind of contract that locks TSMC into a specific customer and limits its ability to charge "spot pricing" later.

For me as a TSM ADR holder, this means I should not project the 2024 gross margin (~58%) forward as if it were structural. The blended margin is partly NVIDIA's gift to TSMC, and gifts can be reclaimed.

NVDA's TSMC Capacity Risk: No Realistic Backup at 2nm

If you flip the lens, NVIDIA's risk profile looks worse on the supply side.

Samsung Foundry has yields and ecosystem gaps that have kept NVIDIA away from its leading edge for two product generations. Intel Foundry is technically interesting (18A is meant to compete with TSMC 2nm) but is roughly 12-18 months behind on volume readiness, and NVIDIA executives have said publicly they need multi-year qualification before trusting a new fab with flagship designs.

That leaves TSMC as the only realistic path for Blackwell, Rubin, and whatever comes after. If TSMC's Hsinchu fabs lost a quarter — earthquake, water shortage, geopolitical event — NVIDIA would not be able to ship and its revenue would drop in proportion to whatever inventory it had on hand. Three to six weeks at most.

This is why I keep a TSM position alongside NVDA. If a TSMC disruption hits, NVDA falls harder than TSM (because NVDA's growth cliff is steeper than TSM's customer-mix shift). My TSM holding does not "hedge" the loss — both stocks would fall — but it would rebound faster and more reliably once production resumes.

The 2024-Q4 Blackwell Delay: A Live Stress Test

Late October 2024 gave us a usable mini-stress test. Reuters and several supply-chain newsletters reported a Blackwell server overheating issue that pushed back some hyperscaler shipments by a quarter. Over the following week:

• NVDA closed down roughly 7-9% peak to trough.
• TSM ADR closed down around 2-3%.
• The Philadelphia Semi Index (SOX) was down about 4-5%.

That is the asymmetry I track. The supplier (TSMC) holds up better than the customer (NVIDIA) when the customer hits a product issue, because TSMC's revenue is already booked and other customers (Apple, AMD) keep orders steady.

It is not bulletproof. If the issue had been at a TSMC fab instead — say a yield problem on 3nm — TSM would have led the drop, not NVDA. But for "NVDA-specific" shocks, the supplier sits one degree removed from the pain.

30% Withholding: What HK/TW Investors Pay on Each

This is where most "nvidia vs tsmc" comparisons get the tax part flat wrong.

NVDA pays a tiny dividend, roughly $0.04 per share annually, yielding around 0.03% at current prices. The US withholds 30% on dividends paid to Hong Kong and Taiwan residents (no tax treaty exists for either). On a HK$80,000 NVDA position, that is roughly HK$1.80 per year in withheld tax. It is rounding error.

TSM ADR is different. The underlying TSMC pays a Taiwan-source dividend, and Taiwan withholds 21% on dividends paid to non-residents. The ADR custodian (Citibank) passes that through, sometimes with a small custodian fee on top. So the effective rate on TSM ADR is closer to 21-22%, not 30%.

On a HK$80,000 TSM ADR position with a 1.4% yield:

• Gross dividend: ~HK$1,120/year
• Withheld at 21%: ~HK$235/year
• Net dividend: ~HK$885/year

Compare that to a hypothetical 30% rate (~HK$336 withheld, HK$784 net) and the difference is meaningful — about HK$100/year per HK$80,000 position. If you read a guide claiming "TSM ADR pays 30% withholding", that guide does not understand the source-of-income rule. (See my US dividend tax for HK residents guide for the framework.)

If you want true zero withholding, you need TSMC's Taipei listing (2330.TW) bought through a Taiwan or HK broker that supports it. That is a different setup with FX and capital-gains complications I am not running myself.

Allocation Frameworks: Picking One vs Holding Both

Here is how I think about who picks what:

Pick NVDA only if: you want maximum AI-cycle beta, you are comfortable with single-customer concentration of supply (TSMC), and your timeframe is 2-5 years before AI capex normalizes. Be ready for 20-30% drawdowns on guidance shocks.

Pick TSM only if: you want exposure to AI compute without single-product risk, you value the dividend cash flow, and you accept that TSM's upside is dampened versus NVDA in good years. Lower beta, lower volatility, slower compounding.

Hold both (my path): you accept that they are correlated to AI capex but uncorrelated to product-specific risks. My current split is roughly 60% NVDA, 40% TSM by dollar exposure within this slice. I rebalance when one runs more than 15 percentage points ahead.

Whatever you choose, position size matters more than ticker choice. Either of these can drop 35-40% in a real AI-capex pause. If that drawdown would force you to sell, the position is too big.

Three Mistakes I Made Pairing These Stocks

Mistake 1: I treated TSM as a "hedge" for NVDA in 2024. It is not. Both fell together in the August 2024 SOX correction (NVDA -16%, TSM -10% over two weeks). A real hedge would be SOX puts or cash, not a correlated supplier.

Mistake 2: I overweighted TSM's dividend in my 2024 return calc. I told myself the 1.4% yield "softened" volatility. In practice, the yield is real money but it is one twentieth the size of a typical weekly price move. It does not soften anything during a drawdown.

Mistake 3: I delayed adding TSM in early 2024 because I thought NVDA "captured" all the AI upside. That cost me about 35 percentage points of TSM upside through the year. The supplier compounds the customer's growth, just on a slower curve. Lesson: when you believe the customer's thesis, also size the supplier.

FAQ

What is the actual relationship between NVIDIA vs TSMC? The "nvidia vs tsmc" framing is misleading. NVIDIA designs the silicon. TSMC manufactures it. NVIDIA does not own a fab. TSMC does not design end-product chips. They sit at different points in the same value chain. If you understood them as "competitors", you would underweight TSM. If you understood them as "vertical partners", you would size both with awareness of the dependency map.

Is TSMC stock the same as TSM stock? TSM is the NYSE-listed ADR of TSMC (Taiwan Semiconductor Manufacturing Company). Each ADR represents 5 ordinary TSMC shares. The Taipei-listed shares (2330.TW) are what TSMC actually issues; TSM is a wrapper that lets US-account investors hold them. Prices and dividends track the underlying with currency and ADR-fee differences.

Can I buy NVIDIA without buying TSMC? Yes, but you are holding NVDA's supply chain risk implicitly. About 80%+ of NVDA's advanced silicon comes from TSMC fabs. If TSMC has an outage, NVDA's revenue drops in lockstep. Owning some TSM does not hedge that loss, but it diversifies which side of the chain takes the volatility hit first.

What about Intel as an alternative foundry? Intel Foundry's 18A node is the credible technical competitor to TSMC 2nm, but on volume readiness it trails by 12-18 months as of late 2025. NVIDIA's leadership has said publicly it needs multi-year qualification before trusting a new fab. For now, Intel is a probe position for me, not a foundry replacement bet.

Should HK/TW investors buy 2330.TW directly instead of TSM ADR? 2330.TW (Taipei) avoids the ADR custodian fee and the 21% Taiwan withholding does not apply if you hold via a Taiwan-resident broker. For HK residents this requires a broker that supports TPE listings. I personally hold TSM ADR for simplicity even though it costs me about 100 HKD/year extra per HK$80K position.

Methodology: How I Tracked Both Stocks for 18 Months

Holding period: NVDA opened December 2023, TSM ADR opened February 2024. Both still held as of May 2026.

Data sources:

• Pricing: TradingView and moomoo HK consolidated tape.
• Fundamentals: company 10-K (NVDA) and annual report (TSM), Bloomberg secondary screens for revenue mix.
• Customer concentration estimates: industry analyst notes (TrendForce, DIGITIMES) cross-checked against TSMC management commentary on quarterly calls.
• Dividend math: my actual moomoo HK statements for the 2024 calendar year, then projected forward at current yield.
• The 2024-Q4 Blackwell pullback figures: my own broker statements showing closing prices Oct 28 - Nov 5.

I am a retail investor, not a professional analyst. The numbers I cite are within 1-2 percentage points of consensus, but I do not have access to TSMC's customer-by-customer revenue breakdown — the 20% figure for NVIDIA is an industry estimate, not a disclosed number.

For real-time HK IPO and broker comparison work, I run the LRTS IPO Tracker and the Broker Fee Calculator. For US-stock specific tax guidance, the US Dividend Tax for HK Residents Guide covers the framework I use.

Affiliate Disclosure

LowRiskTradeSmart earns referral commissions from moomoo HK and TradingView when readers sign up through our links. These referrals do not affect the analysis above. I would hold both NVDA and TSM through a different broker if moomoo did not exist, and I would tell you the same thing about the supplier-customer dynamic.

This article is not investment advice. Both stocks can lose 30-40% in a real AI capex pause, and individual positions in single names should be sized accordingly. Consult a licensed financial adviser before making any investment decision.