The One-Factor Portfolio

The index stopped tracking itself

Start with two correlations that are rarely put side by side.

The first is the twelve-month rolling correlation between the S&P 500 and the Nasdaq-100. By March 2026 it reached 0.98, an all-time high, according to CME Group's research desk. The broad-market index and the technology index have become, for practical purposes, the same series.

The second is the correlation between the S&P 500 and its own equal-weighted version, which over the same period has drifted the other way, toward roughly 0.8, with unusually wide swings. Same five hundred companies. Different weights. Weaker relationship.

Put plainly: the S&P 500 now tracks the technology index more tightly than it tracks itself. Whatever the index is measuring, it is no longer the average American listed company. The mechanism is not mysterious. The information technology sector's weight in the index has gone from 6.7% in 1990 to about 39.6% today. The top ten names reached a record 40.7% of the index in 2025, against roughly 27% at the dot-com peak. Three companies, Alphabet, Amazon and Meta, account for something like 70% of expected 2026 earnings growth for the whole index.

This is the part everyone already knows, and on its own it is just concentration. Concentration is old news and not automatically a problem. The interesting question is what the concentrated thing is exposed to, and whether that exposure has quietly been sold to you more than once.

Five labels, one exposure

Consider a portfolio that a reasonable allocator would call diversified. Public equities. Investment-grade corporate credit. A utilities or infrastructure sleeve. A private credit allocation. Some securitised product. Five buckets, five risk labels, five different lines in the committee deck.

Now look at what is inside them in 2026.

The equity sleeve is the index described above, which is a levered claim on the AI capital cycle. The investment-grade sleeve is increasingly the same names on the other side of the capital structure. Goldman's credit work puts AI-related debt at roughly 23% of dollar investment-grade issuance and about 20% of US high-yield supply, with something near $500bn issued so far this year. Only about 40% of that came from the hyperscalers themselves. The rest is the ecosystem around them, which is the part that is not rated AA.

The utilities sleeve is not a defensive rates proxy any more either. Utility issuance has been the largest single sector in investment-grade corporates since 2024, running at roughly 22% to 25% of year-to-date supply, and the driver is data centre load growth. The names with the most attractive rate-base growth are the ones with the most concentrated data centre exposure. That is the same trade with a regulated wrapper on it.

The private credit sleeve is where the buildout actually gets financed, and the securitised sleeve is where it gets distributed. Data centre asset-backed issuance has gone from around $4bn outstanding in 2020 to roughly $61bn, now about 12% of the esoteric ABS market against 3% five years ago.

So the five labels describe one exposure. That is the claim, and I want to be careful about it, because it is exactly the claim that gets asserted and not tested. What I have described is co-movement in composition, not a measured factor loading. Those are different things, and the difference is the entire piece of work worth doing. I will come back to it.

Credit disagrees with equities, and the disagreement is the signal

Here is where the popular version of this argument goes wrong.

The headline has been that AI credit is cracking. Oracle's five-year credit default swaps have widened to their highest in well over a decade. S&P cut Oracle to BBB− on 9 July 2026, one notch above speculative grade, after fiscal 2026 free cash flow came in at negative $23.7bn. CoreWeave's protection trades at levels that imply a serious probability of default. Net notional CDS on major technology names has hit a record, and Oracle alone is a large share of it.

Now strip those two names out. On the available spread data, the other large hyperscalers have been broadly flat, trading at or inside the index of AA-rated corporates. There is no systemic repricing of AI credit. There is one badly financed issuer, one junk-rated neocloud, and a peer group the market still treats as very high quality.

That is not a smaller story. It is a different one, and for a risk manager it is the more useful of the two. The equity market has collapsed this whole complex into a single factor. The credit market is still pricing it name by name, on balance sheets and cash flows. Two markets are looking at the same capital cycle and disagreeing about whether it is one risk or twenty. When that happens, at least one of them is mispricing something, and the gap between them is where the dispersion trade lives.

Where the exposure hides from the risk system

There is a second reason the labels have drifted from the exposure, and it is more mechanical than sentiment. A growing share of the financing is structured so that it does not appear where a risk system would look for it.

The clearest example is Meta's Hyperion facility in Louisiana, financed through a joint venture with Blue Owl. The special purpose vehicle raised roughly $27bn in loans from a lender group including Pimco, BlackRock and Apollo. Meta holds a minority equity stake and leases the facility. The debt is not Meta's debt on Meta's balance sheet.

The part worth reading twice is the residual value guarantee. Meta has committed to compensate the vehicle's investors if the asset falls below a threshold value, a commitment reported at up to $28bn, described in the footnotes to its annual report with no liability recorded against it.

That is not an accounting scandal. It is ordinary treatment for a contingent obligation. But think about what it does to measurement. The exposure is real, it is large, it is contingent on the resale value of depreciating specialised hardware and the buildings around it, and it is invisible to anything that reads reported leverage. A factor model fed on reported balance sheets will systematically understate how much of this risk any given issuer carries.

The distribution channel has been getting easier too. On 29 July 2026 the SEC's Office of Structured Finance issued an interpretive letter confirming that securities in certain data centre securitisations are not asset-backed securities under the Exchange Act test, which removes the risk-retention requirement for those sponsors. Risk retention exists to keep an originator's skin in the game. Removing it at the point in the cycle where issuance is running hardest is a choice with a history.

None of the official sector is asleep here. The IMF's April 2026 Global Financial Stability Report names concentration in AI-related sectors as a key vulnerability, and pairs it with private credit and the interconnectedness between banks and non-bank lenders. A Federal Reserve study earlier in the year traced bank exposure running indirectly through lending to the very funds that originate data centre debt. The BIS view is calmer, calling the macro-financial risks moderate, while noting that the whole thing depends on AI firms meeting high earnings expectations. That last clause is doing an enormous amount of work.

How you would actually measure it

Everything above is circumstantial. It is a good argument. It is not a measurement, and I would rather say so than dress it up.

The measurement is not exotic. Build an AI factor, either as a long basket of the complex or as the first principal component of it. Regress the total returns of each sleeve, sector and asset class on that factor. Report the loadings, and report the R² through time rather than at a point.

The R² path is the whole answer. If the share of variance in investment-grade credit, or utilities, or a private credit proxy that is explained by one AI factor has been climbing for three years, then diversification has genuinely decayed and you can say by how much. If it has not, then this is a composition story about index weights and not a risk story, and the five labels are doing their job. Either result is worth knowing. Nobody publishing on this has shown either.

Then stress it. Take the allocation, shock the factor, and read the loss. That is a different question from "what do I think about AI," and it is the question a risk engine can actually answer. It is most of what I built Aegis to do.

The reason to do the work rather than assert the conclusion is that the assertion is unfalsifiable and the measurement is not. A number invites someone to check it.

What would make this wrong

The strongest counter is not that the exposure is diversified. It is that the exposure is fine.

JPMorgan's work has hyperscaler capital spending rising above $1.1tn in 2027 against operating cash flow above $900bn, which is a demanding but not obviously broken coverage picture. The bear framing, capital spending compounding far faster than cash flow until free cash flow crosses zero, assumes no acceleration in monetisation. If the fleet converts into durable high-margin revenue, that crossover slides out indefinitely. And revenue in this cycle has arrived faster than in any comparable infrastructure buildout.

Notice, though, that none of that touches the argument. Whether the buildout pays off is a view. How much of your portfolio moves on the same news is a measurement. A portfolio can be correctly positioned for an AI boom and still be carrying four times the exposure its risk report shows. Those are separate failures, and only one of them is fixed by being right.

That is the part I keep coming back to. The interesting risk here is not that AI capital spending disappoints. It is that a lot of allocators will find out how much of it they owned on the same afternoon, having been told all along that they owned five different things.