On the final session before the third-quarter earnings window, the S&P 500 closed at 7,818.93 — a fraction below its all-time high. Over the same stretch, the average constituent of that index fell 14% from its own peak. Both figures are accurate. Neither is noise. The index sits at a record; the market beneath it is in a drawdown.
I have learned to treat that kind of divergence as a structural signature rather than a statistical curiosity. When I audit a DeFi pool, the headline is the last number I trust. The first thing I pull is the distribution — not total value locked, but who holds it. A vault showing $400M in deposits, where three wallets control $320M of it, is not a $400M vault. It is an $80M vault carrying a $320M exit overhang. The headline and the substance are different objects, and conflating them is how auditors miss the failure that is already priced into the structure.
Goldman Sachs built an index — SPXXAI — that performs precisely this audit on the US equity market, and it does so with numbers rather than adjectives. Over the trailing six months, the S&P 500 returned 18.3%. Strip out the AI complex and the same market returns 6.7%. The gap is 11.6 percentage points. Approximately 63% of the index's entire half-year return came from a single narrative cluster. That is not diversification. That is a concentrated position wearing an index's clothing, and the market's vocabulary is not built to name it.
I want to be surgical about the claim, because the lazy version of it is wrong. I am not arguing that AI is a bubble, or that the companies driving this are fraudulent, or that the technology will not deliver. I am arguing something narrower and fully auditable: the structure of the current US equity market exhibits the same concentration pathology I have spent years measuring on-chain, and the tools the market uses to describe itself — "the market is up," "stocks are rallying" — are structurally incapable of expressing the risk that concentration creates.
What SPXXAI actually measures
To read the index properly, you have to know what it excludes. Goldman constructed SPXXAI in February by classifying roughly 45% of S&P 500 constituents as AI enablers — beneficiaries of the AI capital cycle. That classification is deliberately generous. It captures the obvious names: the hyperscalers, the GPU supply chain, the semiconductor designers. But it also sweeps in second-derivative plays — EDA tool vendors like Synopsys and Cadence, networking silicon like Marvell, and increasingly the power producers who sell electricity into data centers. Remove 45% of a cap-weighted index and you are not removing a sector. You are removing the load-bearing wall and asking the plaster to hold the roof.
That the residual — the ex-AI 55% — still returns 6.7% tells you something specific about what is left. That remainder is defensives, financials, healthcare, industrials: the part of the market that is valued on earnings rather than on a capital-expenditure forecast. It is not falling, but it is not carrying the market either. The market's return and the market's fundamentals have decoupled, and the decoupling is measurable.
There is a thesis migrating across two markets that explains why power names now count as AI. Van Eck — an ETF issuer, a fact I will return to — frames the trade in two stages. Stage one was silicon: the chips that train and serve models. Stage two, which it calls the "AI 2.0 trade," is electricity: the generation and, specifically, the nuclear capacity required to feed the data centers. The framing is tidy. It is also, I will argue, a rotation story dressed as an infrastructure thesis.
The mechanism underneath the framing is real, and it is the part worth respecting. AI's binding constraint has migrated from model capability to physical infrastructure — power, cooling, land, and grid interconnection. This is not speculation; it is the same physics I documented in Bitcoin mining economics after the 2024 halving. Compute without power is a paperweight. Power without compute is a stranded asset. The two must be matched, and the matching is governed by physics and permitting timelines, not by valuation multiples or by the enthusiasm of the people quoting them.
So the story is not really an AI story. It is a market-structure story that happens to be wearing an AI mask. And market structure is the one domain where on-chain analysis gives me an unfair advantage, because crypto is the only market in which the entire cap table, every flow, and every concentration metric is public by default. Equities hide what blockchains expose. The skill transfers: I have spent years watching concentrated, reflexive, mechanically-driven structures unwind in real time, and I recognize the shape when I see it.
The arithmetic of concentration
Index returns are weighted averages, and weighted averages can be dominated by a minority of inputs to a degree that intuition systematically underestimates. I wrote a short model to quantify how much of a headline index return a subset of constituents can generate when capital is distributed the way real cap-weighted indices distribute it — as a power law.
import numpy as np
rng = np.random.default_rng(7)
n = 500
w = rng.pareto(1.16, n) + 1 # heavy-tailed cap weights
w = w / w.sum()ai = w >= np.percentile(w, 55) # top ~45% = "AI complex" returns = np.where(ai, 0.25, -0.02)
print(f"cap-weighted: {np.sum(w*returns):+.2%}") print(f"equal-weighted: {np.mean(returns):+.2%}")
# effective breadth: inverse Herfindahl-Hirschman index print(f"effective names: {1/np.sum(w**2):.0f} of {n}") ```
Run it and two things fall out. The cap-weighted index prints a number in the high teens — consistent with the S&P's actual +18.3% — while the equal-weighted average of the same 500 names prints roughly half that. The gap is not a modeling artifact; it is the arithmetic of concentration. A cap-weighted index can post a double-digit gain while the majority of its constituents decline, and nothing in the headline reveals it. The S&P's real data confirms the model: the average stock is down 14% from its peak while the index sits 0.3% from a record.
The second output is the one I find more useful. The inverse Herfindahl-Hirschman index — the effective number of names — collapses a 500-stock index to an effective breadth in the low dozens. A 500-stock index that behaves like a forty-stock portfolio is not a diversified index. It is a concentrated portfolio with 460 passengers. The equal-weighted versus cap-weighted spread is the cleanest real-time monitor of a narrow rally, and it is the equity analog of a metric I track constantly on-chain: the divergence between a token's price and the breadth of its holder base. When a token's price rises while the number of unique profitable holders falls, you are watching distribution — early holders exiting into late buyers. The equity version is the same event at index scale.
Now the classification problem, which is where the hedging logic breaks.
If 45% of the index is AI, then SPXXAI — the ex-AI index — is not a hedge. It is the residual, and the residual is being asked to carry the market's beta with 45% of its mass removed. To hedge an AI-concentrated exposure by shorting the ex-AI remainder is to short the part of the market that is not the trade in order to protect against the trade. The correlation does not cooperate. In DeFi terms, it is like hedging a long ETH position by shorting everything that is not ETH: you have not reduced your risk, you have manufactured a new one and given it a ticker. Goldman built SPXXAI so clients could hedge AI exposure; the same concentration that makes the hedge necessary also makes it ineffective. The tool and the disease are the same object.
The passive-flow mechanism
The deeper mechanism is passive flow, and it is the part the equity commentary consistently underweights.
Index funds and ETFs do not evaluate. They allocate mechanically by weight. When AI names constitute 45% of the index, every passive dollar is a 45% AI dollar, and there is no active decision at the point of purchase — no analyst, no thesis, no exit trigger. The buyer believes they hold a diversified basket. Structurally, they hold one position, replicated across thousands of accounts. This is the precise architecture of a DeFi vault that auto-compounds into a single asset: the depositor reads "basket" and receives "position." And when the position unwinds, the exit is not gradual. It is a queue, ordered mechanically by index weight, which means the largest holders are forced to move first.
This is not a theoretical concern, because a large fraction of that passive flow is retirement money — 401(k) contributions, target-date funds, the automated payroll deductions of households that have never made a discretionary equity decision in their lives. The wealth effect that policymakers rely on to transmit market strength into consumer confidence runs through the same concentrated channel, in reverse. When the AI block reprices, it does not just move an index. It moves the retirement accounts of the median household, and it does so in proportion to their exposure to a single theme they did not choose and cannot see. That is the transmission path from a narrow rally to a broad recession, and it runs through the same wire that produced the rally.
I have audited this failure mode before, and it is worth naming because the equity market keeps rediscovering it. In 2022 I pulled apart the LUNA stabilizer contract and the Mirror Protocol oracle vector — two hundred lines of incentive design that looked robust until you modeled the reflexive loop. The collapse was not a sentiment event. It was a mechanism: a yield that depended on its own price, an oracle that could be pushed, a queue that formed the instant the loop turned. The lesson was not "crypto is fragile." It was that concentrated, mechanical, reflexive structures unwind faster than their participants model them, because participants model the steady state and the structure fails in the transition. The AI concentration in the S&P is a reflexive, mechanical structure. It has not failed yet. But the failure mode is drawn, and it is drawn in the same hand.
The physics constraint
The energy dimension is where the story stops being about sentiment and starts being about physics, and physics is the only kind of analysis that survives a change in narrative.
Data centers and Bitcoin miners are now direct competitors for the same commodity: firm, cheap, dispatchable power. This is not an analogy; it is a bidding war on the same interconnection queues. After the 2024 halving, miner revenue per unit of hash collapsed, and the only operations that survived were those holding the lowest-cost power contracts — often legacy, often stranded, often the product of a decade of negotiation. AI hyperscalers bidding into those same markets raise the clearing price for everyone. Two consequences follow, and both cut against the comfortable version of the AI trade.
First, AI capex is now partially a power-procurement race, and power procurement is governed by permitting, not by capital. You can write a check for a gigawatt. You cannot cash it for seven years. Second, miners without cheap contracts get squeezed out, which accelerates hashrate concentration. My long-standing position — that post-halving hash power concentrates into a shrinking set of pools, hollowing out the decentralization the consensus mechanism is supposed to guarantee — is not a crypto-only story. It is the same story as AI's power race, running on the same grid, driven by the same economics. When two industries bid for a fixed physical resource, the resource wins, and both industries reorganize around it.
The arithmetic is unforgiving, and it is the same arithmetic that governs pool concentration on-chain.
import numpy as np
# post-halving miner survival by firm power cost
cost = np.linspace(0.02, 0.09, 200) # $/kWh
rev_per_kwh = 0.048 # rough post-halving revenue
print(f"surviving hashrate share: {(rev_per_kwh > cost).mean():.0%}")# pool concentration -> effective pool count pool_w = np.array([0.34, 0.22, 0.14, 0.09, 0.06, 0.15]) print(f"effective pools: {1/np.sum(pool_w**2):.1f}") ```
Six pools, an effective count near five. The nominal number of participants is always larger than the effective number, and the effective number is the one that determines behavior. This is why I stopped trusting participant counts years ago. A network with fifty pools and an effective count of five is a network with five pools.
The nuclear leg of the "AI 2.0" thesis runs straight into a permitting wall, and the wall does not care about the narrative. Nuclear capacity, from approval to grid connection, runs on a seven-to-fifteen-year timeline. AI's demand horizon is one to three years. That mismatch is not a footnote to the trade. It is the trade. Prediction markets put the probability of near-term nuclear approval below 10%. The instrument the thesis depends on cannot be delivered inside the window the thesis requires. What can be delivered inside that window is natural gas, storage, and demand response — real assets, but not the ones being marketed.
Which brings me to the source, and to the incentive audit the original reporting declined to perform.
Van Eck is an issuer of nuclear, uranium, and power-themed ETFs. Its "AI 2.0 trade will end the drought" framing is not a forecast issued from neutrality; it is a product statement issued from a position. This is the architecture of trust in a trustless system applied to equity research: when a claim arrives, you do not evaluate the claim in isolation. You audit the incentive behind the claim, because the incentive determines which claims get made and which get buried. A uranium ETF issuer telling you nuclear is the next leg is not lying. It is speaking from inside a position, and reporting that quotes it without disclosing the position has laundered a sales pitch into an analysis. On-chain, we learned this the hard way: the token most loudly branded "community-driven" is usually the one with the most concentrated insider allocation. The tell is never in the claim. It is in the cap table.
And there is the quiet counterexample that the RWA narrative keeps refusing to see.
Alphabet signed a bilateral power purchase agreement with Constellation — a legal contract between two private parties, cleared off-chain, denominated in megawatt-hours and dollars. Constellation rose 12.3% on the day. Not one megawatt-hour was tokenized. Not one settlement touched a public chain. This is the three-year RWA story in miniature, and it points the wrong way: when the stakes are real and the counterparties are institutional, the settlement layer is a contract and a bank, not a smart contract and a block. Traditional institutions do not need your public chain. They need enforceable agreements, and they already have a system for those. The RWA-on-chain narrative assumes institutions want to settle on-chain; the Alphabet-Constellation deal is what institutions do instead.
I have made this argument before and it keeps being confirmed by behavior rather than by rhetoric. In 2021, I spent a week tracing hash collisions in a 500-file sample of Bored Ape metadata and found that roughly 15% of attributes resolved through centralized servers, contradicting the "decentralized" marketing. The lesson was not that NFTs are fraudulent. It was that infrastructure reality and marketing narrative diverge by default, and the divergence is only visible if you inspect the substrate rather than the pitch. The AI trade has the same substrate problem. It markets itself as a broad infrastructure buildout. It resolves, under inspection, to a handful of concentrated dependencies — a few hyperscalers, a few power regions, a few permitting queues.
The blind spot
Here is the part the market is getting backwards, and it is the reason I am writing this rather than a summary.
Everyone is reading the "AI 2.0" trade — the migration from chips to power — as diversification. It is the opposite. Chips and power are not two bets. They are one bet with two legs, and the correlation between the legs is close to one, not close to zero. If hyperscaler capital expenditure decelerates, chip demand falls and power-procurement commitments get renegotiated — simultaneously, because they are funded by the same cash flow and justified by the same forecast. The market is treating a vertically integrated risk as a horizontally diversified one, and pricing the two legs as if they can diverge. They cannot. They are the same cash flow viewed from two ends of a wire.
I have seen this exact error in crypto portfolios during every alt rotation. When Bitcoin dominance peaks and capital rotates into alternative layer-ones, holders believe they have diversified. They have not. They have added beta to the trade they already held, under a new ticker, with a new story. When the trade reverses, the "diversification" reverses with it, because it was never diversification — it was leverage with better branding. The "AI 2.0" rotation is the equity market's alt season: a late-cycle diffusion in which capital, unable to find value in the crowded core, invents a periphery and calls the invention a thesis.
The peripheral surge confirms the stage. Synopsys and Cadence — EDA tool vendors, a second derivative of chip design activity, one step removed from the silicon itself — rose more than 20% in twenty days. No EDA vendor re-rates 20% in a month on fundamentals. That is flow. When the toolmakers who sell shovels to the shovel-sellers rally 20% in a month, you are no longer in the discovery phase of a theme. You are in the distribution phase, and the distribution phase has a name: it is the moment the early buyers need late buyers, and the late buyers need a story that sounds like the early one.
I want to be honest about my own bias here, because an audit that does not disclose its own priors is not an audit. My framework is built to find fragility, because that is what auditing is for — you do not audit a protocol to confirm it is safe, you audit it to find where it breaks. So I am dispositionally inclined to read concentration as risk. The discipline against that bias is to demand a mechanism rather than a mood. The mechanism here is not "AI is overvalued." It is: a 45%-concentrated, passively-replicated, mechanically-rebalanced index, carrying a hedging instrument that cannot hedge, anchored to a physical constraint that cannot be resolved on the thesis's timeline, funded by a cash flow whose deceleration is the single input the entire structure depends on. That is a mechanism. It has a failure mode. The failure mode is not a crash forecast; it is a structural fact about how the thing is built.
Your portfolio has the same shape
If you hold crypto and you are reading this only for the equity angle, you are missing the part that applies to you. The concentration pathology is not an equities-only disease. Pull the holder distribution of any top-20 token and you will find the same power law the S&P exhibits: a small number of addresses controlling a large share of supply, a headline market cap that overstates the float, and an exit queue ordered by size. The stablecoin market is the clearest case — two issuers dominate the entire category, which means the liquidity that the rest of DeFi depends on is itself concentrated into two balance sheets, two regulatory jurisdictions, and two redemption mechanisms. In a bear market, survival is a function of what you are actually exposed to, not what your portfolio is labeled. The reader who wants to know whether their assets are safe should start by measuring effective breadth, not by reading the price.
Where the floor is
The binary catalyst is close and it is specific: hyperscaler third-quarter capital expenditure guidance. Guidance up, and the trade extends another leg, with the fragility compounding underneath it. Guidance down, and the AI complex — chips, power, EDA, the whole correlated block — reprices as one instrument, into a market whose breadth has already thinned to minus fourteen percent and whose index has no second engine. That is the finding, and it is not really about AI. It is about what happens when a market forgets it is concentrated, and then remembers. Where logic meets chaos in immutable code, the remembering is not a process. It is an event.
The question I keep returning to is not whether AI is real — it plainly is. The question is whether anything else is holding the floor. When the load-bearing wall is removed and the roof still stands, you are not looking at strength. You are looking at a structure that has not yet been asked the question.
