Seven structural signals ordered along the AI capex chain — the marginal buyer and the supply chain crack before the flagship does.
General crash-precursor gauges, calibrated against every ≥12% drawdown since 2004 (plus the 1990/2001/2008/2020 cycles for the curve). Ordered by measured lead: the left gauges historically move before tops; the right ones fire after — they grade severity, never timing. Known blind spots: valuation, positioning, and rate-shock bears where the curve inverts after the top (2022).
Signals size positions; they don't time tops. Act on confirmation across categories — credit plus volatility plus breadth plus the upstream chain — never on one line alone.
Stay the plan. Keep contributing on schedule, keep the RSP/international tilt building, sell covered calls into elevated IV on mega-cap positions. This is when hedges are cheap — a good time to price collars, a bad time to panic-buy them.
Review, don't trade. Identify which category is flashing. Equity-only stress with calm credit is usually rotation. Upstream flashing first (CRWV, TSM) while NVDA holds is the pattern worth respecting — check whether credit starts confirming within two weeks. Pre-write your de-risk orders now so execution later is mechanical.
Execute the tiers. Tier 1: stop reinvesting into cap-weighted index funds; new money to T-bills. Tier 2: trim top-10 overlap (SPY/QQQ) toward equal-weight and international sleeves — tax-advantaged accounts first. Tier 3: roll covered calls to lower strikes or buy 3–6 month index puts if VIX is still under ~25. Do not liquidate wholesale — the goal is cutting concentration, not calling the exact top.
The white pin shows distance to the tripwire. Tiles are ordered by the thesis's presumed breakdown sequence — marginal buyer → supply chain → financing → index fragility → institutional hedging → flagship → regime confirm. How far right the red reaches is how far the cascade has progressed.
CoreWeave is the debt-financed, single-purpose GPU landlord — no diversified cash flows to hide behind. If AI compute demand softens even slightly, it hits CRWV before NVDA's order book feels anything. The weakest balance sheet in a boom always cracks first; in 2000 the CLECs broke well before Cisco. Wide thresholds because the stock is violently volatile by nature.
TSMC fabricates essentially all of NVIDIA's chips and reports revenue monthly, not quarterly. Its stock plus its monthly number is NVDA's next earnings report visible a quarter early. A TSM drawdown while NVDA holds up means the ecosystem is repricing before the flagship.
Extra yield junk-rated borrowers pay over Treasuries. Credit investors get paid to be paranoid, so spreads widen before equity accepts bad news. Speed matters more than level: +75bps in a month means lenders are repricing risk fast. With the AI buildout heavily debt-financed, financing stress shows up here while the flagship still looks fine.
If equal-weight falls hard while cap-weight sits at highs, fewer stocks carry the index — fragile. If RSP holds while mega-caps wobble, money is rotating within the market — healthy. This gauge classifies every other drawdown on the board.
When 1-month volatility costs more than 3-month (backwardation) for 3+ days, big money is paying up for near-term protection. Same structure you screen in single names, applied to the whole index.
The most-owned expression of the AI trade. A 15% drawdown alone has happened repeatedly this cycle and recovered. Its job on this board is confirmation: NVDA cracking while CRWV/TSM already flashed and credit widens is the sequence starting. NVDA cracking alone is a dip.
Same spread measure for blue-chip debt, including bonds funding Oracle and Meta data centers. IG barely moves in normal corrections; IG widening alongside HY is regime change, not noise — the last domino in the chain.
These were calibrated empirically, and the honest finding is that most classic indicators confirm rather than predict. Treat the left half as early-warning and the right half as severity-grading once a drawdown is already underway.
Inversion alone is far too early — the 2022-24 inversion ran 535 trading days. The historical pre-top signal is a deep inversion (10yr−3mo trough ≤ −0.40pp) that then re-steepens rapidly (+0.40pp over 60 days, back above zero) as the market prices Fed cuts. That combination fired 1-4 months before the recession-linked tops of 1990, 2007, 2020 — and in Dec-2024, two months before the Feb-2025 top. Blind spot: rate-shock bears like 2022, where the top comes before the inversion.
Copper is a bet on activity, gold a bet on fear. The ratio dumping ≥9% in a month — while gold itself holds — is a growth scare with a safety bid. Only counts when gold holds: if both fall, it's a commodities selloff, not fear.
Global liquidity drains at the edge first. EEM underperforming SPY by 5%+ in a month while SPY sits near highs is the index-level version of "the marginal buyer cracks first." Caveat: inverts when EM itself is the bubble (2007).
SKEW prices deep out-of-the-money puts; VVIX prices volatility-of-volatility; the gauge is whichever is richer versus its own 2-year range. Half-weight by design: it often peaks a month or two before the top then fades — a falling reading is not an all-clear. The value text also shows the VIX/VIX1Y ratio: front-month panic vs out-year pricing.
Discretionary/staples rotation. Empirical surprise: this ratio rose into 7 of 10 tops — mild weakness means nothing. Only a violent ≥8% monthly collapse counts, and by then the correction is usually underway. Severity signal, not timing.
Global crashes are dollar-funding events: when leveraged books unwind, everyone needs dollars at once. A ≥4%/month dollar spike has essentially never been a false alarm (Lehman, eurozone, 2020, 2022) — but it comes after the equity top. It tells you the drawdown you're in is global deleveraging, not rotation.
Chicago Fed composite of 105 money-market, debt and equity measures. Mechanically loose at tops. Warn = ordinary correction underway; alert = credit event in progress. Grades severity of what's already happening.
Initial claims 4-week average vs its one-year low, sustained two weeks. +10% is a wobble; +20% has flagged every NBER recession since 1968 — but it lags equity tops by weeks to months. When this fires during a drawdown, the drawdown has a recession behind it: deepest tier of the playbook.