Reflexivity and the AI Supercycle: What Happens When the Narrative Drifts

Belief is financing the largest capex cycle in history. An analysis of index concentration, debt-funded AI infrastructure, and who actually holds the risk when the reflexive loop breaks.

George Soros' theory of reflexivity makes a claim that sits uncomfortably with textbook finance: markets do not merely discover prices, they help manufacture the fundamentals those prices are supposed to reflect. Participants act on biased perceptions; their actions move prices; prices change the underlying reality, including credit conditions, capex decisions, and reported growth; and the changed reality appears to validate the original bias. The loop runs until the gap between narrative and underlying utility becomes too wide to sustain, at which point the correction is not gradual but violent.

Applied to the AI investment cycle of 2025 to 2026, the framework is uncomfortably precise.

The Loop, Operationalized

Hyperscaler capital expenditure is tracking above $600bn in 2026, a 36% increase over 2025, with roughly three quarters of it targeting AI infrastructure. That spending is not a sideshow to the US economy: it is the US economy at the margin. AI-related investment accounted for roughly 74 to 75% of Q1 2026 GDP growth; without it, growth would have been near 1% annualized, and total corporate capex would have been negative. AI infrastructure spending now represents about 5% of US GDP, a share last seen in the late-1990s telecom buildout.

This is the reflexive mechanism in its purest form. The belief that AI will transform the economy triggers the investment; the investment shows up in GDP and in supplier revenues; those data points are then cited as evidence that the belief was correct. The demand side adds a second-order loop: more than $800bn in circular financing arrangements, including chip makers investing in AI labs that spend the proceeds on chips, and cloud providers booking revenue funded by their own vendor commitments. OpenAI alone has infrastructure commitments of roughly $1.15tn across seven vendors against a projected 2026 loss of about $14bn. Revenue validates capex; capex generates revenue. Whether this is a virtuous circle or accounting alchemy depends entirely on end-demand materializing, which is precisely the variable the loop cannot generate internally.

Concentration: The Index Is the Trade

The equity market expression of this loop is historic concentration. The top 10 stocks now account for approximately 43% of S&P 500 market capitalization, the highest reading on record, versus roughly 27% at the 2000 peak and an 18 to 23% norm from 1990 to 2015. The Magnificent 7 alone are about a third of the index; Nvidia, at a market capitalization near $5tn, carries roughly an 8% weight by itself.

The practical consequence: passive allocation to "the market" has become a leveraged bet on a single thesis. A pension plan holding a plain S&P 500 tracker is running a roughly 40% position in AI-adjacent mega-caps, whether its investment committee ever approved that view or not.

Notably, this is not a story about absurd index multiples. The S&P 500 trades near 20x forward earnings, elevated against its 5- and 10-year averages, but nowhere near dot-com trailing extremes. What distinguishes this cycle from 1999 is not the price of earnings; it is the concentration of the bet and, more importantly, how the underlying capex is financed.

The Financing Layer: Who Is Actually Paying

The first phase of the AI buildout was funded from operating cash flow. That phase is over. Hyperscalers raised over $100bn in debt in 2025; projections for 2026 hyperscaler issuance run at $250 to $300bn, with cumulative AI-related bond issuance approaching $570bn by year-end. Meta's capex is projected at roughly 54% of sales, Microsoft's at 47%, Oracle's at 86%: a ratio historically associated with utilities and shipyards, not software franchises. Oracle's credit default swaps have roughly tripled since September; its rating sits two notches above high yield.

Beyond the bond market, the structures grow more opaque: special purpose vehicles, GPU-backed lending, private credit funds, securitizations. The chain of exposure runs from the technology company through the SPV to the private credit originator, into rated tranches, and ends with pension funds, insurers, and asset managers. New York and Pennsylvania state pension plans are invested in the $7bn digital-infrastructure fund behind Meta's off-balance-sheet data center SPV and multiple Oracle financings. The capital funding the AI narrative is, at the end of the chain, retirement savings: largely without the beneficiaries' knowledge, and partly in structures designed to sit outside anyone's balance sheet.

The Drift Scenario

Reflexive loops do not end because someone disproves the narrative. They end when the narrative exhausts itself, when the marginal data point stops validating the story. The mechanism to watch is semantic, not fundamental: the moment consensus re-labels AI capex from innovation to leverage, every number changes meaning without changing value. $600bn of annual spending reads as visionary while the story holds and as a debt problem the day it drifts. Depreciation schedules, take-or-pay contracts, and off-balance-sheet vehicles that were growth infrastructure on Monday become fixed obligations on Tuesday.

The transmission channels in that scenario are mechanical. First, equity: a re-rating of a 43% index block cannot be diversified away within the index. Second, credit: AI-adjacent issuance has grown fast enough that a sentiment turn hits investment-grade spreads, private credit marks, and securitized structures simultaneously. Third, the real economy: if roughly 75% of GDP growth is AI capex, a capex pause is not a sector story but a macro event. And fourth, the reflexive amplifier: weaker growth undermines the very earnings that justified the multiples that funded the capex.

Soros' own playbook is instructive here, and widely misread. He is not positioned against this cycle: his fund increased its Nvidia stake by 61% and TSMC by 49% in Q1 2026. That is consistent with his stated method: "When I see a bubble forming, I rush in to buy, adding fuel to the fire. That is not irrational." The theory does not say bubbles are shortable early; it says the gap between price and utility is rideable until the inflection point, and that the wider the gap has grown, the more pronounced the snapback.

Allocator Implications

For portfolio constructors, the central analytical question is not whether AI infrastructure investment will ultimately prove productive. On sufficiently long horizons, it very likely will. The question is whether the financial architecture around that investment, the concentration, the leverage, the circular financing, and the SPV structures, is stable enough to allow capital to remain patient through the years between current capex and monetized utility.

The concentration problem is not addressable within broad index strategies. A portfolio that seeks diversification through weighting to non-AI sectors remains substantially exposed through the index mechanism itself; reducing that exposure requires active underweighting of the largest constituents, which most mandates do not permit. The credit exposure problem is less visible and arguably more acute: allocators with private credit or infrastructure debt sleeves may be holding AI-adjacent paper they have not identified as such, priced at marks that assume continued narrative support.

The historical comparator most often cited is the late-1990s telecom buildout, which ended in roughly $2tn of stranded assets and a credit cycle that took five years to clear. The differences are instructive: that cycle was concentrated in one sector; this one runs through the index, through private credit, through SPVs, and into pension portfolios. The scope of transmission, if the narrative drifts, is broader.

The Question That Remains

If a reflexive loop is identifiable in real time, why does it not self-correct? The answer is embedded in the mechanism: participants who exit early are penalized by the performance of those who stay, and the loop reinforces itself precisely because it is profitable to remain inside it. The correction is not triggered by awareness; it is triggered by an external break in the self-referential data chain. The relevant question for allocators is not whether the break is coming, but whether the portfolios they hold are structured to distinguish between a position inside a reflexive loop and a position in productive long-term capital formation. In most cases, at present, they are not structured to make that distinction.

Sources: CreditSights; MUFG AI Chart Weekly; BEA Q1 2026 third estimate; St. Louis Fed; Apollo Academy; Bloomberg AI circular deals analysis; Fortune; FactSet Earnings Insight; Soros Fund Management 13F Q1 2026 via TheStreet; George Soros, The Alchemy of Finance.

Photo: George Soros, World Economic Forum Annual Meeting 2011, Davos. Credit: World Economic Forum / Remy Steinegger. License: CC BY-SA 2.0.

This is not investment advice.