Buy NVIDIA, or the Semiconductor ETF? Flagship Stock VS Sector ETF

Say you’ve become convinced semiconductors look good. One fork remains: buy the flagship stock, or buy the sector-ETF basket? Everyone says the ETF is “safer” — but few have checked how much safer, and in what sense, with actual numbers. So we set up the same experiment twice: NVIDIA versus SMH in the US, SK hynix versus KODEX Semicon in Korea — same five years of data, same two models, side by side. One market followed the script. The other tore it up.

A fair comparison starts with one yardstick. All four assets use 61 month-end observations of adjusted-price returns over the same five years, June 2021 to June 2026, and the risk-free rate behind every Sharpe ratio is the same US 13-week T-bill (3.67%/yr). Now, the lineup.

One thing to note in the table before we start: NVIDIA is about 15% of SMH — meaning buy SMH and 15% of your money goes to NVIDIA anyway. The monthly-return correlation between the two is 0.78; for the Korean pair, a striking 0.88. A flagship and its sector ETF are not strangers — they are the same story at different concentrations. Which raises the question: if only the concentration differs, how different can the numbers really be?

US — less return, same efficiency

Start with the US pair. SMH’s return (+39.0%/yr) is clearly below NVIDIA’s (+62.9%) — necessarily so, since an ETF’s return is the weighted average of its holdings and can never beat its best stock. But shift your eyes to the volatility column and the story changes: σ drops from 50.8% to 33.3% — risk cut to two-thirds. As a result the Sharpe ratio, which measures return earned per unit of risk, barely moves (1.16 → 1.06), and the worst-year loss estimate (95% VaR) shallows from −20.7% to −15.8%. The ETF gave up return but shed more risk — a trade that came out ahead.

This is Markowitz’s “free lunch.” Bundle several stocks and the return comes out exactly at the weighted average — but the risk comes out below it. As long as the holdings don’t move in perfect lockstep (correlation < 1), their wobbles partially cancel. A well-built basket is therefore not “average return, average risk” but “average return, below-average risk.” That is precisely why SMH’s Sharpe held up.

Korea — the ETF loses on Sharpe and VaR?

Yet the same logic flips in Korea. KODEX Semicon’s volatility (45.3%) is indeed lower than SK hynix’s (66.4%) — that part is normal. The problem comes next: the Sharpe ratio drops from 1.15 to 0.82, and the 95% VaR deepens from −29.3% to −33.7%. Putting the money in the basket made both risk-adjusted efficiency and the worst-year loss estimate worse.

The culprit is on the return side. Most of Korean semis’ gains over these five years belonged to a single stock — SK hynix (+79.9%), holder of the HBM crown. Yet hynix occupies only about a quarter of KODEX Semicon; the rest is filled by Samsung Electronics, sluggish over this window, and small/mid-cap suppliers. The result: return was cut in half (+40.8%) while risk only fell to two-thirds (45.3%). When return is diluted more than risk is reduced, Sharpe must fall — and with a lower mean, VaR digs deeper. Diversification didn’t break; it’s just that what the basket lost by underweighting the flagship exceeded what it gained by spreading out — for these five years.

So how does the “risk cut to two-thirds” magic actually happen inside the ETF? To find out, we lifted SMH’s lid: its top ten holdings, bundled at their real (renormalized) weights, fed straight into a Markowitz analysis — watching diversification happen on site.

The secret lies in the correlations between the holdings. “Same sector, surely they all move together” — but measure it and the pairwise correlations spread from 0.23 (NVIDIA–Intel) to 0.84 (Lam Research–KLA), averaging near 0.5. Designers, foundries, equipment makers, memory houses — even within one industry, companies march to their own beats. Those offbeats, added up, produced the 25.19% diversification effect.

Yet the same number also shows the ceiling of sector-ETF diversification. In earlier posts we measured 39.3% for the National Pension Service of Korea’s top-100 equity book, and even the genre-mixing “attention basket” managed 38.5%. Within a single sector, mix as cleverly as you like and you get 25% — there is a band of risk reduction that only opens when you cross industries. A sector ETF is not the finish line of diversification; it is the first step.

Where Markowitz measured the size of risk, the Fama–French five-factor model asks about the source of return. It splits an asset’s return into the shares owed to five common factors — market, size, value, profitability, investment — leaving two numbers behind: the fraction of the movement those factors explain (R²), and the excess return nothing explains (alpha). The US pair is regressed on North America factors; the Korean pair on our own Korea factors, built from domestic market data.

R² — bundling makes an asset explainable

Start with the R² row: NVIDIA 62.5% → SMH 68.6%; SK hynix 83.6% → KODEX Semicon 87.7%. Different markets, same direction — the moment an asset becomes a basket, factor explainability rises. No coincidence. A single stock’s price is full of events that happen only to that company — product launches, lawsuits, earnings surprises — and that idiosyncratic risk lives in the unexplained residual (1 − R²). In a basket, those firm-level events offset one another, leaving mostly the common factors. The σ reduction in Section 3 and the R² rise here are the same phenomenon seen from two angles — proof that the risk being erased was precisely single-company risk.

One thing stands out — the Korean assets’ R² (83.6%, 87.7%) is actually higher than the US pair’s (62.5%, 68.6%). Explainability that high for single stocks comes from the fact that the Korean market itself is heavily concentrated in a handful of chip giants, so SK hynix’s swings largely coincide with the market factor’s. The flagship doesn’t so much “track” the market as constitute a large slice of it.

Alpha — drama on one side, a sum of factors on the other

The alpha row is the coldest line in this piece. NVIDIA’s alpha is +44.2%/yr and statistically significant (p=0.008) — a genuine, company-specific explosion that five factors combined cannot account for. But the same market’s ETF, SMH, comes in at +17.7% (p=0.062), not distinguishable from zero. The flagship’s drama, blended into the basket, is pulled down toward ordinary factor-explained returns — diversification erases not just the bad idiosyncratic risk but the good explosions too.

The Korean pair reads differently. Neither SK hynix (+25.8%) nor KODEX Semicon (+15.3%) has a significant alpha. Even the flagship leaves no statistical trace of excess return. HBM sent hynix soaring, but that surge was a sum of factors — a high market beta (1.24) and a pronounced profitability tilt (RMW +1.19) — rather than company-only magic. Where NVIDIA’s AI explosion was singular enough to defy even the whole North America factor set, hynix’s HBM cycle was closer to the Korean market’s own big wave. Diversification didn’t so much erase the idiosyncratic drama as find there was little to erase.

A caution, though: do not read that big, significant NVIDIA alpha as “replicable edge.” We picked a flagship that had already won and replayed its past, so the alpha is soaked in after-the-fact selection. Intel in 2021 had its own glorious trailing alpha once.

Factor fingerprints — two semiconductors, two characters

Read the betas line by line and each Korean asset’s character emerges. SK hynix loads −0.80 on size — a large cap behaving like one — −1.25 on value (a strong growth tilt) and +1.19 on profitability (a distinctly high-quality name): the identity of an ultra-profitable large-cap growth stock, the HBM margin leader, stamped straight into five numbers. KODEX Semicon, by contrast, sits near zero on size (+0.06) — its large caps like hynix and Samsung and its small/mid-cap suppliers cancel out each other’s size color. The basket blurs even the flagship’s sharp factor fingerprint toward the average.

① What it erases — single-company risk

Three numbers pointed the same way: σ fell (50.8→33.3%), R² rose (62.5→68.6%), and alpha’s amplitude died (+44.2%→+17.7%). Compressed into one sentence — an ETF erases the risk of “something happening to that one company.” But the jackpot only that company could hit gets erased with it. The two always come as a set.

② What it can’t erase — the sector itself

And yet, after all that erasing, SMH’s market beta is still 1.55 and its volatility 2.1× the S&P 500’s; KODEX Semicon runs 1.4× KOSPI with a −33.7% VaR. When the semiconductor cycle turns, the whole basket turns with it — a risk no amount of slicing within the sector removes. Buying a sector ETF doesn’t mean you finished diversifying; it means you boarded the one ship called “semiconductors,” merely spreading out across its cabins. When the ship rolls, every cabin rolls together.

③ Two “semiconductor ETFs” can be different animals

Finally, the lesson the Korean twist leaves behind. By name, SMH and KODEX Semicon sound like the same product — but one holds 25 global large caps and the other 20 Korean names, and over these five years the former kept pace with its flagship on Sharpe while the latter fell far behind. An ETF’s report card is written not by the ETF format but by what it holds and at what weights. Read the holdings table, not the name.

Nothing here was exotic — public price data and two models, that’s all. If there’s a flagship you’re eyeing and a sector ETF that holds it, put both tickers in and run them side by side. How much the return dilutes, how much the risk drops, whether the Sharpe holds or collapses — whether that ETF really is the safe choice comes out in numbers, not vibes. Turning “it’s an ETF, so it must be safer” into “this ETF is safer by this much.” That is the real use of this piece.

Data · prices: yfinance month-end adjusted close (Jun 2021–Jun 2026, 61 months) · factors: US pair = Ken French Data Library FF5 North America; Korean pair = our own Korea FF5 factors (from public Korea Exchange, DART and Bank of Korea data, from 2016-07), 59–61 months · risk-free rate: US 13-week T-bill (3.67%/yr) · SMH holdings: yfinance fund data (retrieved Jul 2026) · KODEX Semicon: tracks the 20-stock KRX Semicon index, per Samsung Asset Management disclosures.

This article was written with the help of AI.