Since Jackson Hole, the S&P 500's rate-sensitive cohort — utilities, real estate, consumer discretionary, and a slice of the banks — has surrendered roughly 10%. Over the same window, Citi's year-end target for the index sits at 8,100, raised from 7,700 in June on a single justification: AI-driven earnings. The arithmetic is unforgiving. That target needs two things at once: AI earnings keep beating, and rate-sensitive sectors rebound. Right now, only the first is true. The market has quietly become a single-factor trade, and the factor is capital expenditure. The ledger doesn't lie about which side is carrying the index — and it doesn't lie about what is missing. This is not a prediction problem. It is an accounting problem, and the accounts are public.
Citi's Scott Chronert has been explicit. Two groups of stocks exist inside the S&P 500, and only one of them — the AI cohort — is in good shape. To reach 8,100, the other half has to join the move. He calls this a diffusion problem. I call it a breadth problem, and breadth problems are measurable.
The framing is worth stating precisely because it is slippery. "AI-related companies" are described as slightly more than half of the S&P 500. If that means weight, AI is now a macro variable, not a sector. If it means company count, the number is close to meaningless — an index of 500 names cannot be half-composed of one theme without that theme swallowing the whole. The reporting places the AI trade at roughly 50% of S&P 500 companies. Either way, the definition is doing enormous work and hiding enormous variance. Semiconductors, hyperscale cloud, enterprise software, power generation, and data-center REITs are not the same business, do not share margins, and do not share capex cycles. Bundling them is a convenience, not a category.
The mechanical driver is hyperscaler capital expenditure. These firms are simultaneously the largest buyers of AI compute and the largest sellers of AI cloud services. Their capex guidance is the demand signal for the entire upstream chain — chips, networking, cooling, electricity. Aya Yoshioka's framing is the honest one: the outlook depends on whether cloud growth and return on investment hold. If they do not, investors reset 2027 expectations, and the reset is where drawdowns live.
Hyperscaler capex also has an oligopoly structure worth naming. If one firm holds back, it risks losing cloud share. If all firms invest, industry returns compress. That is a prisoner's dilemma with a multi-year payback period, and the payback period is precisely what is not being disclosed.
The earnings arithmetic reinforces the caution. Q3 consensus sits in the upper-20s percent year-over-year; Q2 printed in the low-30s. Growth is still high, but it is decelerating, and deceleration is the enemy of multiple expansion. The 8,100 target was justified by AI-driven earnings strength. It was not accompanied, in anything I can find, by the EPS assumption or the multiple that gets you from here to there.
Concentration has a mechanical amplifier, and it is not sentiment. Passive funds must buy the index's largest weights regardless of price. The bigger the AI cohort grows as a share of the index, the more passive inflow it captures, which grows it further. That is a reflexive loop, not a valuation judgment. It means the index can drift higher on flows alone — and it means the unwind, if it comes, is equally mechanical.
Breadth is the entire discussion, and the source is candid about it. The 8,100 target is not a bet on AI. It is a bet that the rest of the index catches up. Chronert says it plainly: the target depends on the other half joining the advance. That is a conditional forecast, and conditionals are only as good as the second condition. Right now the second condition is failing, and the failure is visible in the same rate-sensitive cohort that the Fed's path controls.
Two macro variables sit underneath all of this. The first is the rate path. The second is the midterm elections, which the source material lists as a year-end obstacle — a reasonable flag, because tax, regulatory, and energy policy all feed directly into capex timing and into the discount rate applied to long-duration assets.
One data-hygiene flag before we go further. The source material timestamps a 25 basis point Fed hike to September 16 and references an October 28 meeting. Against the actual 2024–2025 easing path, that is a contradiction, not a detail. If the rate direction in the source is wrong, every rate-sensitive conclusion built on top of it inherits the error. I flag it because I have watched too many models get validated on inputs nobody audited.
Here is where I stop reading strategist notes and start reading the chain. My method is unchanged since I audited oracle price feeds in 2017: follow the transmission path, not the headline. The question is not whether AI earnings are strong. It is whether the capital that would have to fund a diffusion rally is actually present, and where it is sitting.
Start with stablecoin supply. Net minting and burning is the cleanest proxy we have for dollar-denominated risk appetite entering and leaving the system. In the weeks since Jackson Hole, the pattern has been one of rotation, not expansion. Supply has held roughly flat while shifting between chains and between venues. That is not the signature of new capital arriving to broaden a rally. It is the signature of existing capital repositioning. When I ran the same lens in 2022, after the Terra collapse, retail panic showed up on-chain only after whale balances had already moved to cold storage. The flow preceded the sentiment. It always does.
The next series is spot ETF custody. I spent part of 2024 auditing cold-wallet movement against reported reserve ratios for major issuers, and I found discrepancies that corrected public figures by roughly 15%. That experience made me distrust any "institutional demand" narrative that is not tied to a specific wallet, a specific block, and a specific timestamp. The relevant signal now is not the headline net inflow. It is the dispersion — how concentrated creations and redemptions are among a handful of authorized participants. If a small number of desks is doing most of the moving, the "institutional bid" is liquidity provision, not conviction.
The most reflexive series is funding rates and perpetual open interest across the major venues. This is where the AI-capex trade and the crypto trade physically touch. Both are duration-sensitive, both are rate-sensitive, and both are priced off the same macro discount rate. When funding compresses while price holds, leverage is leaving without price conceding — a coiled spring, not a broken one. When funding compresses and price falls, you are watching forced de-risking. The distinction is the whole game. Order-book depth tells the same story from the other side: depth that thins while spreads hold is market makers stepping back, not sellers stepping in.
None of this requires proprietary data. It requires discipline. I run the same three checks every week: net stablecoin issuance, ETF custody dispersion, and funding against open interest. When all three agree, the signal is real. When they disagree, the disagreement is the signal, and it usually means price is running ahead of capital.
I want to be concrete about what "dispersion" looks like at the block level, because vague language is how narratives survive. Concentration shows up as a small number of large transfers moving between a handful of addresses in tight time windows, with fee spikes that mark urgency. Broad participation shows up as many mid-sized transfers across many addresses with flat fees. I have pulled both patterns from public data more times than I can count. The first is a desk. The second is a market.
Now overlay the AI-capex variable. The on-chain expression of the AI trade is not the S&P's semiconductor complex. It is the DePIN and decentralized-compute cohort — the tokens that sell GPU time, bandwidth, and storage against the same demand curve the hyperscalers are riding. These tokens are a higher-beta, lower-liquidity derivative of the same capex narrative, which makes them a cleaner read on sentiment than the underlying. When hyperscaler capex guidance is strong, this cohort bids first and hardest. When guidance wobbles, it bleeds first. If you want to know what the market believes about AI capex before the earnings call, watch the compute tokens, not the index.
There is a caveat on that cohort, and it comes from my own work. In 2021 I traced a 50-wallet cluster wash-trading a collection's floor price by matching gas patterns and mint timestamps. Manufactured volume has a fingerprint. Compute tokens have their own version: incentive programs that subsidize usage. Reported demand and organic demand are two different numbers, and only one of them survives an incentive cliff. Any read on this cohort that does not separate the two is a read on a marketing budget.
There is also a slow variable underneath the fast one. AI capex converts into depreciation on a multi-year schedule. Revenue has to arrive before the depreciation does, or margins compress mechanically regardless of demand. The 2027 reset Yoshioka warns about is not a sentiment event. It is an accounting event with a calendar, and every quarter that capex outruns revenue pushes that calendar closer.
Market structure compounds the risk. When a single factor drives the index, options dealers hedge against that factor, and their hedging becomes part of price discovery. The tail risk is not that AI earnings miss. It is that AI earnings merely meet expectations, and the mechanical hedges unwind into a vacuum of buyers.
Put the series together and the picture is coherent. Capital is present but not expanding. Institutional flow is concentrated, not broad. Leverage is being managed, not abandoned. None of that supports a diffusion rally. All of it supports a narrow one. And a narrow rally, held up by roughly half the index, is the definition of concentration risk. You cannot audit a target that does not show its work.
The lazy conclusion is that narrow equity breadth means narrow crypto breadth, and that both roll over together. Correlation is not causation, and here the correlation is partly an artifact of a shared discount rate, not a shared mechanism. Two things can move together because the same macro variable drives both, and still diverge violently when that variable stops dominating.
The AI-capex complex and the crypto AI-token complex are correlated in price and different in mechanism. Hyperscaler capex is a balance-sheet decision made by a handful of committees. Decentralized compute demand is a spot decision made by thousands of buyers. The first is slow and lumpy; the second is fast and reflexive. If capex guidance disappoints, the equity complex reprices on a quarterly cadence. The token complex reprices in hours, often overshooting in both directions. Treating them as one trade is the error.
There is a second blind spot. Utilities sit in both buckets at once — rate-sensitive and AI-power beneficiary — and the source material never separates them. A position that is short rate-sensitive sectors and long AI infrastructure may be long and short the same cash flows. That is not a hedge; it is a hidden doubling.
A third blind spot is the label itself. "The AI trade" is not one position. It is a semiconductor position, a power position, a software position, and a credit position wearing a single name. When breadth narrows, the label hides which leg is actually carrying the weight. Strip the label and the exposure becomes auditable. Keep the label and you are trading a narrative with four different risk profiles.
Watch the compute tokens into the next hyperscaler print. If they bid while funding stays flat and stablecoin supply does not expand, the market is expressing conviction without new capital — a narrow tape that can run further than it should. If they fade while ETF creation concentrates in a few desks, the diffusion the 8,100 target requires is not coming this quarter. The next seven days will separate a positioning market from a conviction market. The tape is sideways; the flows are not. That gap is where the next move is decided. The chain will tell you first. It usually does.