The Earnings Are the Bubble

This week the beat stopped working. Palo Alto Networks beat and raised, and the stock faded. CrowdStrike beat on nearly every line, announced a four-for-one split, and fell double digits after hours. Broadcom beat, more than doubled AI revenue year-over-year, guided next-quarter AI revenue up sharply — and dropped more than eight percent because the CEO declined to raise a full-year number he'd already given. Three companies cleared the bar and all three were sold. That is not a fundamentals story. That is an expectations story, and it is the first honest tell I've seen in months that the binding constraint on this market has quietly shifted from how good are the numbers to how good did the price already assume they'd be.

I want to be precise about what I think is mispriced, because the lazy version of this view — "stocks are expensive, watch out" — has been wrong for two years and will get you run over. My claim is narrower and, I think, more falsifiable: the bubble right now is in the earnings, not in the multiple. The valuation is extreme, but the valuation is downstream of an earnings base that is itself inflated by an accounting asymmetry the AI capital cycle has not yet been forced to close. I'll show the mechanism, the financing that confirms it, and the single variable — total addressable market — that decides whether this ends in a crash or in something arguably worse.

Three analytical frames carry this piece — the capex super-cycle staging, the spend-versus-expense earnings point, and the TAM bifurcation — and I've assembled each from my own tracking of the buildout over the past months. Where I'm most accountable is the accounting: the heart of this thesis is a matching-principle problem I can take apart line by line.

01What the tape said this week

Start with the facts, because the reaction function is the story. Palo Alto Networks posted Q3 adjusted EPS of $0.85 against a $0.80 estimate on revenue of $3.00B (+31% y/y) and raised the full-year guide — and after a roughly 60% run in May into the print, faded from a sharp after-hours spike to flat-to-lower. CrowdStrike's Q1 beat on nearly every line — adjusted EPS $1.10 vs $0.88, revenue $1.39B (+26% y/y), record net-new ARR of $256M, a raised FY27 ARR guide and a 4-for-1 split — and the stock fell ~11–13% after hours from a $747.61 close, because billings of $1.35B (+18%) read light against a price up ~65% year-to-date. Broadcom's Q2 cleared EPS ($2.44 vs $2.40) on revenue of $22.19B (+48% y/y, a hair light), more than doubled AI revenue to $10.8B and guided next-quarter AI to ~$16B — and fell more than 8% because management did not raise the standing ~$100B full-year AI target after a ~13.6% five-day pre-print run.

The index math frames it. The S&P 500 set a record close at 7,609.78, the Nasdaq Composite at 27,093.90 and the Dow at 51,307.79 on 2 Jun 2026. Then, on 3 Jun, the S&P fell ~0.74%, the Nasdaq ~0.89%, and the Dow 620 points (−1.21%) — snapping a nine-day win streak on the tech reactions above plus a fresh Iran flare-up I'll come to. And the whole thing happened with the VIX at 15.77. That is the part that should bother you: a 1.2% Dow day that closed with the volatility index under 16. The market is not braced. My own read going into this week was the same: at the 31 May close, implied vol was sitting at essentially zero on the 13-week look — I haven't seen less risk priced in over a comparable stretch. Complacency is not a timing signal, but it is a fragility signal, and fragility plus an expectations-bound tape is how air pockets start.

Both reads on this market — the bull's "earnings are great, look at the beats" and the bear's "valuations are insane" — are half-right, and together complacent. The beats are real. The valuations are real. What neither camp is pricing is that the earnings themselves are borrowing from a future expense line.

02The mechanism: a matching-principle problem at national scale

Here is the accounting, and it is the center of the whole thesis. When a hyperscaler or a neocloud spends on AI infrastructure, that spending shows up almost immediately as revenue and gross margin for the seller — Nvidia, Broadcom, Micron, the equipment makers, the software layer that rents the compute. But the matching depreciation expense, which should offset that revenue over the life of the asset, lands slowly, on a schedule, in later periods. So in the installation phase of a capital super-cycle, the income statements of the enablers look spectacular for a structural, not a fundamental, reason: 100% of the revenue is recognised now, and only a sliver of the eventual expense is recognised against it. The earnings base is, quite literally, ahead of its own cost.

Run the numbers I've worked through, because they scale cleanly. Take 2026 AI-related capex at the midpoint of the range — call it $500B (2026 hyperscaler capex is estimated near ~$725B, ~75% AI-specific or ~$545B; the $500B figure is conservative). Split it the way the buildout actually splits: ~50% silicon depreciated over ~6 years, ~25% equipment over ~15, ~25% structures over ~30.

Year one, ~$290B of gross margin shows up against ~$54B of depreciation. That gap is why the enablers' earnings are screaming. But depreciation is cumulative and revenue is not guaranteed to be. Do the same $500B again in 2027 — even at zero growth — and you are now carrying 2026's depreciation plus 2027's on top of it. At a six-year life the schedule compounds to roughly $108B; if the right useful life is four years rather than six, it's ~$150B. Against a 2026 AI revenue run-rate of ~$70B, that means AI revenue would have to nearly double just to cover the depreciation expense — not to be profitable, just to cover the schedule. The depreciation clock is the timing device for this entire cycle, and it has barely started ticking on the income statement. (That ~$70B denominator is the softest figure in this piece; I reconcile it to a defensible ~$110–140B net in the Skeptic's Appendix, which shrinks the gap without closing it — and, more importantly, re-points the fragility from depreciation to revenue quality.)

This is where my training does the work the macro framing can't. The fight over useful life is not academic — it is the whole game, and it is an auditor's fight. CoreWeave depreciates GPUs over ~6 years; Nebius ~4; Michael Burry argues 2–3. The spread between a two-year and a six-year life on a $250B silicon base is the difference between an industry that prints money and one that is structurally unprofitable on a fully-costed basis. Every year a hyperscaler extends a useful-life assumption, it flatters current earnings by pushing expense further out — and every shortening does the reverse. So the cleanest early-warning indicator I have is not a price. It is a footnote. A hyperscaler or neocloud shortening its server depreciation life is admitting the assets wear out faster than the earnings assumed. I'll be honest that this is the weakest link in my chain, and I don't want to oversell it: Amazon has already trimmed a subset of server lives from six years to five with no reckoning, and the 2026 GPU-rental data cuts against the simple version of the claim. I take that apart properly in the Skeptic's Appendix; for now, treat the footnote as a signal to watch, not a smoking gun.

03The financing is telling on itself

If the earnings were as clean as the headline beats suggest, the financing would be boring. It is not boring. It is getting circular and off-balance-sheet, which is precisely the late-cycle tell. Roughly $120B of AI infrastructure debt has moved off balance sheets via special-purpose vehicles — Oracle ~$66B, Meta ~$30B, xAI ~$20B, CoreWeave ~$2.6B. The structure is always the same, and it has ancestors: an external SPV raises the debt, builds the data center, and leases it back to the tech company, which then books a lease rather than a borrowing. The Meta Louisiana vehicle is the cleanest specimen — a $30B project, ~$27B debt (PIMCO/BlackRock/Apollo) plus $3B equity (Blue Owl), off Meta's balance sheet, with Meta retaining 20% and obliged to cover losses if the facility value falls below a floor and the lease isn't renewed. That last clause is the point: Meta has retained a contingent liability and a residual-value risk, and neither shows up as debt. The SPV carried an investment-grade rating against a sponsor — CoreWeave — that is itself high-yield. An auditor's eye snags on every one of those seams: the consolidation question (who really controls the VIE?), the residual-value guarantee, the lease-versus-debt characterisation, the IG-rating-on-a-HY-sponsor arbitrage.

And then there is the circular financing — Nvidia taking equity stakes in customers who turn around and buy Nvidia hardware. Strip out the AI vocabulary and that is vendor financing, and the last time vendor financing was this fashionable in technology it was Nortel and Lucent lending their customers the money to buy their own switches into 2000. The revenue was real until the customer couldn't pay, at which point it was never revenue at all. I am not saying we are at that point. I am saying the financing structures that appear near the top of capital cycles are here, on schedule, and they corroborate the earnings-asymmetry read rather than contradict it. Michael Burry holds puts on SOXX, QQQ, Nvidia, Oracle and Palantir (Jan-2027 expiry) and flags Nvidia's $119B non-cancellable TSMC obligations and customer concentration. You don't have to agree with his timing to respect the exposure he's pointing at.

04The one variable that decides everything: TAM

Whether this is a bubble at all comes down to a single question — what is the total addressable market the $500B-a-year is chasing?

Here is the part I want to sit with, because it's the version of this thesis that should frighten you more than a tech correction: if scenario two is right, the bubble isn't the AI infrastructure — it's the wage bill. Ten percent of the cognitive-labor market is $2T of income that disappears from households, which is $2T of demand removed from the economy. And the two scenarios are not mutually exclusive. You can get scenario one's infrastructure crash first — the digestion, the impairments, the bankruptcies of the weak neoclouds — and then, on the cheap overbuilt capacity that survives, scenario two's slow repricing of labor. That is the nightmare ordering for a government already running structural deficits: a market crash that demands fiscal support, followed by a labor shock that raises transfer payments while shifting income from highly-taxed individuals to lightly-taxed corporations. The fiscal arithmetic gets hit on both sides at once.

05The overlay the framework didn't have: an energy shock and a boxed-in Fed

The macro tape has hardened since my last full read at the end of May, in one specific way: the Iran situation re-escalated, and it matters for inflation and therefore for the Fed. The Strait of Hormuz is effectively blockaded, traffic at ~5% of pre-conflict levels since mid-May. A paper ceasefire keeps breaking — Iran suspended US talks on 1 Jun over Israeli action in Lebanon; Netanyahu warned Israel and the US were "ready" to strike again on 3 Jun; a Kuwait airport was struck by drones. Oil jumped on it — WTI settled ~$96.02, Brent ~$97.81 on 3 Jun, easing to WTI ~$94.46 the next day. On the year crude is actually down (WTI −7.5% YTD) because it's unwinding from a ~$126 Brent peak in March — but it is +~49% year-over-year, and that y/y comparison is what feeds the inflation prints.

And the prints are split in a way that is the whole policy problem. Headline CPI is 3.8% y/y — highest since May 2023. But core PCE is 0.2% m/m / 3.3% y/y and, critically, the Dallas Fed trimmed-mean PCE is 2.3% (12-mo). The energy shock is doing the damage in the headline while the fat middle of the distribution — the part the new Fed chair built his career arguing you should watch — is essentially at target. Consumer sentiment printed 44.8, a record low, even as ISM Manufacturing hit 54.0, its highest since May 2022. Firms see demand; households feel the pump. That is a textbook supply-shock signature, and it is the worst possible backdrop for a central bank because the two halves of the dual mandate point in opposite directions.

Into that, the Fed has changed hands. Kevin Warsh was sworn in as the 17th Fed Chair on 22 May 2026 (term to 2030); Powell remains a governor. The framework had this right, and it's now fact, not forecast. The market reads the June 17 meeting as ~70% hold / ~28% cut / ~2% hike — and the live debate, contrary to the "rate-hike risk" of a few weeks ago, is hold-versus-cut, not hold-versus-hike. My read: the right first move for a new chair facing a supply shock is no hikes, no cuts, and no forward guidance — establish that the reaction function has changed, let the energy spike wash out of the y/y comps, and watch the trimmed mean rather than the headline. The thing to watch on the 17th is not the rate; it's whether Warsh signals the end of the residual Treasury reinvestment and lays out a balance-sheet path. QT formally ended 1 Dec 2025, so the framing is no longer "aggressive runoff" — it's how fast he chooses to let reserves drift down from ~$3T+.

06The consumer underneath the record highs

One more layer, because it's the one the index level hides. 90+ day delinquencies are climbing across the board: credit cards ~13–14% (approaching GFC levels), auto loans the highest in the series back to 2003, student loans surging, mortgages ticking toward ~1%. With personal income at 0% growth in Q1 2026 and spending still rising, the gap is being financed on the most expensive consumer credit in a generation — the market has already delivered the equivalent of nearly two rate hikes to the five-year (auto) and one to the 30-year fixed mortgage at 6.53%. The household balance sheet is not why this market is at a record. It is why the next negative surprise won't be cushioned.

07Where I come out

I am not bearish on price into year-end, and I want to be honest about that because it's the uncomfortable part of holding this view. The story is real — the buildout is real, the technology works, and from a game-theoretic seat every hyperscaler's dominant strategy is to keep spending past the point of marginal return because cutting back cedes the field. Overbuilding is the rational equilibrium, and it can persist for quarters. My blunt version: be very careful about shorting anything here. The beat-and-fade tape doesn't change that; it just tells you the fuel is running lower than the index suggests. The euphoria most likely carries into the end of 2026. The reckoning is a 2027 event, and the depreciation schedule is what brings it.

So I hold two things at once: the direction of travel is up for now, and the quality of the earnings carrying it is deteriorating beneath the surface. The way I'd express that — and Brighthedge writes expressions, never trade calls — is to stop paying up for the enablers after 60–80% three-month runs (the PANW/CRWD/AVGO reactions just repriced the cost of that bravery), to treat any hyperscaler depreciation-life extension as a yellow flag and any shortening as a red one, to watch the SPV credit-spread complex for the first sign the off-balance-sheet machine is jamming, and to read every "AI revenue" figure as gross-of-double-counting until proven otherwise. The capex-to-revenue ratio is the single number that adjudicates the whole thesis: ~7:1 on the headline figures (~$500B capex vs ~$70B revenue) vs ~2.4:1 at the 2011–12 cloud buildout peak. I'll flag immediately that the 7:1 is the generous-to-my-own-case version — quality-adjusted for double-counting, the net base is nearer ~$110–140B and the ratio nearer ~3.5–4.5:1 (the reconciliation is in the appendix). On either cut the question is the same: if that ratio compresses toward 3:1 on genuine, profitably-priced revenue catch-up, I'm wrong and this is early innings. If it widens, or if the revenue closing it turns out to be subsidised, the clock wins.

This also keeps me honest against my own live book. My standing macro view in The Dollar's Borrowed Strength (22 May 2026) is bearish-USD; the dollar has firmed to DXY ~99.4 (3 Jun) on safe-haven and solid data, and USD/JPY 159.92 (4 Jun) is pressing the 160 intervention watch. That thesis is intact — EUR/USD ~1.165 is nowhere near its 1.1300 falsifier — but a sustained energy-shock dollar bid is the kind of thing that drifts a position toward its invalidation, and I'm flagging it rather than burying it.

08What makes me wrong

09Skeptic's appendix — the strongest case against me

I ran this thesis through an adversarial pass before publishing, and it landed two hits hard enough that I'd be dishonest to bury them. I'll state them at full strength, because a thesis you can't see through is a thesis you don't really hold.

The first hit: the depreciation error may run the opposite way. My piece implies depreciation is under-recognised, so earnings are over-stated. The 2026 tape on GPU economics argues the reverse. Older A100 (2020-vintage) rental prices rose ~15% in 2026, with aging GPUs "trading like a scarce commodity" per Nvidia's CFO; A100s remain cost-efficient for sub-13B-parameter inference at <40% utilisation, and replacing them early with H100s incurs unrecoverable double capital cost. If a six-year-old chip is appreciating in the rental market, a six-year depreciation life is conservative, not aggressive. And the useful-life judgments are not a soft footnote — they are audited estimates that have moved in both directions: Amazon cut a subset of servers from 6 to 5 years on a study citing faster AI/ML obsolescence, while the bulk extended 3–4 → 6 years (~$18B/yr aggregate effect). The considered industry read is that "AI factories bend, but don't break, useful-life assumptions." Burry's ~$176B understated-depreciation claim for 2026–2028 is a hypothesis about economic obsolescence that the rental data is actively refuting. So I concede: the useful-life version of my argument is the weak version, and I am de-emphasising it.

The second hit, and the one that re-points the whole thesis: my denominator is probably too low — which would trigger my own falsifier. I anchored on a ~$70B net AI-revenue run-rate. But Anthropic alone reached a ~$47B run-rate in 2026 (from $30B ARR in April; $1B→$30B in 15 months) and OpenAI ~$25B. That is ~$72B from two model vendors before Google Cloud AI, Azure/Copilot, AWS Bedrock, or Broadcom's $10.8B AI quarter (~$43B annualised). Sum the visible layers and the gross AI-economy tally runs $150–200B; the honest net base is lower, and it's worth pricing the haircut precisely rather than gesturing at it. To be concrete: the ~$72B of model-vendor run-rate is the cleanest layer, but Anthropic books revenue on compute it rents from AWS and Google — so part of its top line is already someone else's cloud revenue — and the downstream SaaS "AI revenue" stacked on top is the model vendors' output resold, not new dollars. Strip the resold-compute pass-through (~15–20%), exclude the app-layer double-count entirely, and add back only genuinely first-party hyperscaler AI services, and I reconcile to a non-double-counted base near ~$110–140B — still roughly double the $70B I used, which puts capex-to-revenue around ~3.5–4.5:1 rather than 7:1: not yet through my 3:1 invalidation line, but far closer than my headline framing implied. I cannot wave that away: if $70B is right, I owe an explanation for why two private firms alone report run-rates summing to more than the whole industry's net revenue.

So here is where the adversary actually left me — not retreating, but re-pointed. The fragile part of the AI earnings story is not hidden depreciation. It is revenue quality. There is a token-pricing crisis behind the OpenAI/Anthropic revenue race — much of that fast-compounding revenue base is bought with below-cost, capacity-grab pricing. That rescues the spirit of my thesis while killing its original mechanism: the danger isn't that depreciation is concealed; it's that the revenue catching up to the capex is itself subsidised and won't survive rational pricing. A ~$130B base earned at negative unit margins is not the same asset as the same base earned at a profit, and only the second one absorbs the depreciation schedule when growth flattens. The compounding math still bites — three stacked years of $500B+ capex run their depreciation cohorts simultaneously — so I don't need Burry's two-year-life claim; I only need 2028 revenue growth to decelerate while the 2025–27 cohorts all depreciate at once.

Where I'll concede the financing point too: the off-balance-sheet fragility concentrates in Oracle ($66B), xAI ($20B) and CoreWeave ($2.6B) — the weak hands — not in the hyperscalers funding the bulk of the ~$725B 2026 capex from operating cash flow. A depreciation drag on Microsoft or Alphabet is a margin story absorbed by buybacks, not an insolvency cascade. The systemic-crash version of my view requires the weak hands to be systemically important, and they aren't yet. That lowers my probability on a 2027 infrastructure crash and raises it on a slower margin-and-multiple de-rate.