AI concentration risk in index funds

The arithmetic

Cap-weighted index funds hold companies in proportion to market value. That is usually a feature. It becomes a risk when a handful of names get very large relative to everything else.

Measure	Figure	Context
S&P 500 top-10 weight	~37–38%	Early September 2026, after peaking above 40% during 2025
Top-10 weight, 1990–2015	~18–23%	The range that held for a quarter century
Top-10 weight at dot-com peak	~25–27%	March 2000. Today's level is materially higher
Largest single holding	~8%	Nvidia, as of September 2026 — the biggest single weight in decades
Top-10 share of index earnings	~32%	End of 2025, against ~40% of index weight

That last row is the one worth sitting with. The leaders are genuinely, enormously profitable — this is not 1999, when many leaders had no earnings at all. But the market is paying for roughly 40% of the index with companies producing roughly 32% of its earnings. The premium is the risk, not the profitability.

Why your funds are more correlated than they look

The common mistake is counting funds rather than exposures. Holding five funds feels diversified. Looking through to what they actually own often tells a different story:

• S&P 500 fund. The exposure, directly.
• Total US market fund. Adds thousands of small companies at tiny weights. The top names remain nearly the same share, because they are weighted by the same market values.
• Nasdaq-100 fund. The same leaders, at even heavier weights.
• Target-date fund. Typically a wrapper around broad index funds, so it inherits the concentration and changes only the stock/bond split.
• International developed fund. Genuinely different companies, but many are in the same AI supply chain — semiconductor equipment, foundries, power and cooling.
• Employer stock, if you happen to work in tech. Your salary, your options and your index fund can all depend on the same capital-expenditure cycle.

The useful exercise is a look-through: list every account, pull each fund's top-10 holdings, and total up what you actually own of each underlying company. Most people are surprised. That number, not the number of funds, is your concentration.

Four historical episodes, and what actually happened

History does not repeat on schedule, and none of these is a forecast. They are useful mainly for calibrating how wide the range of outcomes has been.

The Nifty Fifty (1972–1974)

A group of large, genuinely excellent American companies — Xerox, Polaroid, Avon, IBM, Disney, McDonald's — became known as “one-decision” stocks you could buy and never sell. Many traded above 50 times earnings. In the 1973–74 bear market Xerox fell about 71%, Avon about 86%, and Polaroid about 91% from their highs. The businesses were real. The prices were the problem. Several of these companies went on to decades of solid operation while their shareholders spent years underwater.

Japan (1989–2024)

The Nikkei 225 rose more than 450% in the eight years to its December 1989 peak near 38,957, trading around 60 times earnings. It did not surpass that level again until February 2024 — roughly 34 years. This is the tail case that matters most for anyone with a fixed retirement date: not that markets fall, but that “stocks always recover” can mean a timeframe longer than an investing lifetime.

Dot-com (2000–2015)

Information technology reached roughly 33% of the S&P 500 by June 2000. The Nasdaq Composite then fell about 78% into October 2002 and did not reclaim its March 2000 closing high until April 2015. The broader S&P 500 fell roughly 49% and made new highs in 2007. The gap between those two recoveries — seven years versus fifteen — is the entire practical argument for diversification.

The counterexample nobody mentions

Concentration has also persisted far longer than skeptics expected, repeatedly. Commentators called technology overvalued in 1996, four years and enormous gains before the peak. People who exited early on valuation alone did real damage to their own returns. “This looks expensive” and “this will fall soon” are completely different claims, and conflating them is how cautious investors underperform for a decade.

What is genuinely different this time

A fair treatment has to include the strongest version of the other side:

• Real profits. Today's leaders generate enormous free cash flow. Most 1999 high-flyers never earned a dollar. This is a substantive difference, not a talking point.
• Real revenue from the theme. AI spending is showing up in actual reported results, not just projections.
• Fortress balance sheets. The largest names hold net cash, unlike the debt-funded telecom buildout that amplified the 2000 bust.

And the honest rebuttals: much AI revenue is capital expenditure by a small circle of companies buying from each other, which resembles the vendor-financing dynamic of 2000 more than anyone likes. Depreciation schedules on AI hardware are an assumption, not a fact, and if useful life is shorter than assumed, reported earnings are overstated. And profitability has never prevented a valuation contraction — ask a 1972 Xerox shareholder.

Metrics you can actually watch

These describe conditions. None of them times a top — most were flashing for years before 2000, and acting on them early was itself costly. Treat them as a thermometer, not an alarm clock.

Signal	What it tracks	Why it matters
Top-10 index weight	Share of the index in its ten largest names	The direct measure of how much of your fund is one bet
Weight vs earnings gap	Top-10 share of weight minus share of earnings	A widening gap means price is outrunning fundamentals
Market breadth	Equal-weight versus cap-weight index performance	Persistent divergence means a narrow group is carrying the market
Capex to operating cash flow	How much of hyperscaler cash flow goes into buildout	Rising ratios mean today's earnings depend on tomorrow's payoff
Customer concentration	Share of a supplier's revenue from a few buyers	Circular revenue among a small group is fragile
Your own look-through	Total exposure to each company across all accounts	The only one of these that is specific to you

You can examine the concentration and momentum characteristics of a strategy yourself in the builder, or look at how individual names score on the Buffett screener and the momentum screener.

Mechanics passive investors use, and what each costs

These are descriptions of common approaches and their tradeoffs, not recommendations. Every one of them can underperform for years, and several will look foolish for as long as the leaders keep leading.

• Measure first. Do the look-through before changing anything. Some people find they are far less exposed than feared, and the correct action is none.
• Scheduled rebalancing. The quiet de-concentrator. Rebalancing on a fixed calendar trims winners mechanically, without requiring a forecast. Cost: it lags in a sustained momentum regime and can trigger taxable gains outside a retirement account.
• Equal weight. An equal-weight version of the same index caps any single name's influence. Cost: higher turnover and fees, a structural tilt toward smaller companies, and long stretches of underperformance when megacaps lead.
• Broaden beyond US large-cap. International developed, emerging markets, and small-cap value hold genuinely different businesses. Cost: US large-cap has beaten all of them for most of the past fifteen years, and the diversification benefit weakens precisely during global panics, when correlations converge.
• Factor tilts. Value and quality behaved very differently from growth in 2000–2002. Cost: value spent much of the 2010s as a punchline, which is exactly how long a tilt can hurt before it helps.
• The bond and cash allocation. The only reliable diversifier in an equity crash is not being fully in equities. This is unglamorous and is usually the largest single lever an ordinary investor controls.
• Explicit hedges. Puts, collars, and inverse funds exist. They carry ongoing cost, expire, and require timing to work. They are generally poorly suited to a passive plan, and are listed here for completeness rather than as an option most people should reach for. You can model option payoffs in the options explorer to see the cost structure for yourself.

The question worth asking instead

“Is this a bubble?” is unanswerable in advance and mostly unproductive. A better question is: if the top ten fell 50% and stayed down for seven years, what would I be forced to do?

If the answer is “keep contributing and ignore it,” your concentration is probably survivable regardless of what happens next. If the answer is “sell at the bottom to pay for something,” the problem is not the AI trade — it is a portfolio sized to a timeline it cannot support. That is fixable today, without predicting anything.

This is also why the honest version of risk management is boring. It is about the size of a position you can hold through a bad decade, not about spotting the top.

Frequently asked questions

How concentrated is the S&P 500 right now?

As of early September 2026, the ten largest companies were roughly 37–38% of the index by weight, having peaked above 40% during 2025. The top ten were about 25–27% at the 2000 dot-com peak and generally 18–23% from 1990 to 2015. These figures move daily; check a current source.

Is my index fund still diversified if ten stocks are 38% of it?

It holds many names, but its risk is concentrated. A total-market fund does not help much, since it holds the same leaders at similar weights. Many international funds carry exposure to the same AI supply chain, so holding several funds can still mean holding one bet.

Is the AI boom a bubble?

Nobody can answer that in advance. The narrower observation: today's leaders are highly profitable, unlike most dot-com leaders, but carried roughly 40% of index weight against roughly 32% of index earnings at the end of 2025. That gap raises the cost of being wrong without indicating when, or whether, it will matter.

What happened to investors in past concentrated markets?

Outcomes varied widely. The Nasdaq fell about 78% after 2000 and took until April 2015 to recover; the S&P 500 fell about 49% and recovered by 2007. Nifty Fifty leaders like Avon and Polaroid fell more than 85% in 1973–74. Japan's Nikkei took from December 1989 until February 2024. Broadly diversified investors who kept contributing recovered considerably faster than concentrated ones.

What can a passive investor actually do about concentration risk?

Measure true look-through exposure, rebalance on a schedule so winners are trimmed automatically, consider whether equal-weight, value, or international exposure fits the plan, and size the bond and cash allocation to a drawdown that would otherwise force a sale. Each carries real costs and can underperform for years. None of this is advice, and concentration metrics have never reliably timed a top.