Goldman Sachs published a number that should stop any macro watcher cold: strip AI-related stocks from the S&P 500, and the index's six-month gain collapses from 18.3% to 6.7%. The 11.6-percentage-point gap means roughly 63% of the market's entire advance came from a single thematic cluster. This is not diversification. This is concentration dressed up as a bull market.
The index in question, SPXXAI, was launched in February 2025 explicitly as a hedging tool for clients worried about AI exposure. By August, it had peaked and rolled over, sitting 6.4% below its high while the S&P 500 hovered just 0.3% from its own peak. Two indices, same underlying universe, diverging trajectories. That divergence is the signal.
The macro view reveals what the micro ledger hides. At the micro level, individual AI winners look unstoppable. At the macro level, the market has become a single-factor bet on hyperscaler capital expenditure, and the breadth data confirms it.
The Breadth Problem Nobody Wants to Price
Schwab's Kevin Gordon flagged a statistic that deserves more attention than it received: the average S&P 500 stock has fallen 14% from its peak since early August. Not the index. The average stock. While the market-cap-weighted S&P 500 sits near all-time highs, the typical constituent is in a drawdown deeper than many corrections.
This is a narrow rally. Historically, narrow rallies are late-cycle phenomena. The 1999-2000 dot-com period had the same signature: a handful of names carrying the index while the median stock bled. The 2021 meme-stock era repeated it. Both ended the same way.
Goldman classifies roughly 45% of S&P 500 constituents as AI-driven. That number is the one that should worry anyone holding SPXXAI as a hedge. When nearly half the index is the thing you are trying to hedge against, the hedge is mathematically compromised. You are not neutralizing AI exposure. You are buying a slightly different flavor of the same risk.
The AI 2.0 Narrative and Its Structural Flaw
The market has already moved to the next story. Van Eck's "AI 2.0 trade" framework splits the theme into two phases: chips first, then power producers and nuclear energy. On the surface, this makes sense. Data centers consume enormous electricity. If AI scales, power demand scales with it.
But here is where I apply my pre-mortem framework. In 2022, I spent four weeks reverse-engineering TerraUSD's collapse mechanism. The lesson was not that algorithmic stablecoins fail. The lesson was that when a system's survival depends on reserve assumptions that cannot hold under stress, the market prices the happy path until the exact moment it cannot.
The AI 2.0 narrative has the same structure. Van Eck, which manages nuclear and uranium ETFs, describes the power phase as ending a "drought period." Prediction markets assign less than 10% probability to nuclear approval within the year. These two data points cannot both be true. Either the narrative is early, or the narrative is marketing.
Code does not lie, but it often obscures intent. ETF issuers publish research that aligns with their product lineup. That is not conspiracy. That is incentive. The question is whether the market is pricing the incentive or the substance.
The Physical Constraint That Breaks the Thesis
I designed a zero-knowledge proof settlement layer for autonomous AI agent transactions in 2026, processing 50,000 transactions per second. The bottleneck was never computation. It was always the physical layer, specifically energy density and cooling. This experience shapes how I read the AI infrastructure trade.
The constraint has shifted. GPU supply was the binding limit in 2023-2024. Now it is grid interconnection, transformer lead times, and nuclear regulatory cycles. A nuclear plant takes 7 to 15 years from approval to grid connection. AI demand growth operates on a 1-to-3-year horizon. That mismatch is not a gap. It is a chasm.
The macro consequence is counterintuitive: the power bottleneck can strangle the chip trade it supposedly complements. If hyperscalers cannot secure enough electricity to power the GPUs they have already ordered, the marginal demand for new chips declines. The AI 2.0 trade is not additive to the chip trade. It is potentially cannibalistic.
Alphabet's power agreement with Constellation Energy, which sent Constellation up 12.3% in a single session, signals that tech giants are locking in energy supply. But the deal terms are undisclosed. The duration, the volume, the price, none of it is public. The market celebrated a headline without seeing the contract. In my audit work, I learned that the details are where the risk lives. A multi-signature wallet with an integer overflow vulnerability looks fine until someone exploits it.
What the Hyperscaler Earnings Actually Test
The next critical node is hyperscaler Q3 earnings, specifically capital expenditure guidance. This is the binary event. If capex guidance is raised, the AI trade continues. If it is lowered or even held flat, the market loses its primary support.
The reason is simple. AI chip stocks like AMD, Marvell, Synopsys, and Cadence, which gained over 20% in 20 days, are second-derivative bets on hyperscaler spending. EDA tool vendors like Synopsys and Cadence benefit from chip design activity, not end-user AI adoption. Their rise reflects supply-side exuberance, not demand-side validation. This is the diffusion phase of a thematic trade, where capital rotates from leaders to adjacent names because the leaders are already crowded.
In an 18-month period, I analyzed over 10 million on-chain transactions to map institutional deposit patterns around the Bitcoin ETF approval. The insight was that ETF inflows acted as a liquidity sink, not a price driver. The same logic applies here. The AI trade is absorbing capital, not generating returns. The difference between a sink and a source matters when the flow reverses.
The Valuation Anchor Is Missing
The S&P 500 closed at 7,818.93. No price-to-earnings multiple is provided. No earnings growth rate is cited for the AI-weighted constituents. No comparison to historical valuation percentiles is offered. This is not an oversight. It is a structural feature of how thematic trades are covered.
The valuation question is the one that determines whether this is a durable trend or a bubble. If AI earnings are growing fast enough to justify multiples, the concentration is rational. If multiples have expanded without corresponding earnings growth, the concentration is a risk.
I do not have the answer. But the absence of the question in mainstream coverage is itself informative. When a trade becomes consensus, the questions that challenge it stop being asked.
The Contrarian Read: Decoupling Is the Real Story
The dominant narrative frames this as "AI is carrying the market." The contrarian framing, and the one I find more useful, is "the market has no other engine."
Remove AI, and the S&P 500 returns 6.7% over six months. That is not a crash. But it is a market that depends on a single sector to generate alpha. The 45% AI-classification threshold means the "non-AI" market is not a diversified basket. It is a residual. It is what is left after you subtract the thing that actually moves.
This matters for macro positioning because it changes the correlation structure. In a normal market, sectors provide diversification. In an AI-concentrated market, most sectors are correlated to the same underlying factor. When that factor turns, everything turns together.
The crypto market learned this lesson in 2022. DeFi protocols appeared diversified by yield source but were correlated through shared liquidity and shared collateral. When Terra collapsed, the correlations went to one. I published a warning on liquidity fragmentation three months before the first major exploits. The mechanism was identical: interconnected protocols lacked isolation. The equity market is now running the same playbook with AI as the shared collateral.
What the Power Constraint Means for Crypto Infrastructure
My recent work on AI-agent payment protocols has reinforced a structural view: the next wave of crypto utility will not come from speculation. It will come from machine-to-machine settlement. Autonomous agents need non-custodial, high-throughput, low-latency payment rails. That demand is real, but it depends on AI scaling, which depends on power, which depends on regulatory timelines measured in years.
The power bottleneck is therefore not just an equity market issue. It is a constraint on the entire AI stack, including the crypto infrastructure layer that will serve it. The protocols that survive will be the ones that do not assume unlimited compute growth.
The AI trade is not dead. But the risk-reward has deteriorated. The 11.6-point spread between the AI-stripped index and the full index is the cleanest measure of how much the market depends on a single thesis. When the thesis wobbles, the index wobbles. There is no buffer.
The macro view reveals what the micro ledger hides. And right now, the micro ledger is showing a market with thinner support than the headline number suggests.
Watch the hyperscaler capex guidance. Watch the SPXXAI divergence. Watch the equal-weight versus cap-weight spread. These are the leading indicators. The lagging indicator will be the correction that everyone says was obvious in hindsight.
It always is.
