We don't just publish what we won't predict. We adversarially audited our own engine against eight real shocks and are publishing exactly where it would break, degrade, or hold — plus the public data that would tell you before we noticed.
Our engine models who wants AI's inputs. It has almost no model of who supplies them. Across eight independent shock tests — each checked by an adversarial skeptic hunting for flattering verdicts — 18 of 40 mapped failure paths came back genuinely vulnerable, 17 partial (the claim survives, the numbers go stale), and only 5 fully absorbed by the model's existing design.
The single deepest finding: our calibration is armored in the direction the world already moved from 2023-2026, and comparatively naked to a reversal. That's not a bug we're hiding — it's the reason this page exists.
Every shock below got the same treatment: name the real-world event, name the exact engine fields it would hit, and grade the honest verdict. Vulnerable means the shock would silently produce a wrong answer. Partial means the claim class survives but the calibrated numbers go stale until we refresh them. Absorbed means the model's existing design already handles it. Every verdict below was independently re-checked by a skeptic instructed to argue the opposite case first.
Our matrix maps who wins from chip demand; it does not yet model what happens when the one island that fabs the chips goes offline — a Taiwan supply shock would flip our strongest industry linkages faster than any field we currently track would catch it.
Watch: TSMC monthly revenue releases, CoWoS/HBM allocation reports, and Taiwan Strait war-risk insurance premia.
Our rankings assume policy arrives at institutional speed; a tariff round, an export-control reversal, or an impulsive strike arrives in days — and our cross-effects web contains no oil-price, defense, or tariff geometry, so the first such shock would grade us on a damage-ordering we never actually modeled.
Watch: USTR Section 301 filings and the Section 122 150-day statutory clock; BIS entity-list actions; tanker war-risk premia on Hormuz-sensitive routes.
Our fragility flags already name the crowded AI industries as the ones that fall hardest in an unwind, but our near-term rankings assume that unwind is gradual — a fast positioning cascade would hit our top-ranked industries in an order our cross-effects don't currently model, and we'd catch it at the next calibration, not in real time.
Watch: Prime-broker leverage and crowding in the Fed's Financial Stability Report, the NAAIM exposure index, and absorption-ratio spikes.
Our industry tape anchors are cap-weighted ETFs, so where one company effectively IS the index — NVIDIA in chip design — a single earnings print can rewrite what we believe the industry did, and we currently grade ourselves against benchmarks one name can move.
Watch: Any single name's implied one-day move (options straddle pricing) exceeding roughly 8% when that name is more than 10–20% of an industry's proxy ETF weight.
Our read treats power scarcity as durable and monetizable by utilities and grid-equipment makers; if datacenter moratoria turn that scarcity into demand destruction, or an SMR/interconnection surprise makes it abundant, our energy-linked rankings would be wrong in ways the model does not currently self-detect.
Watch: State PUC large-load tariff dockets, PJM/ERCOT interconnection-queue withdrawals, and firm SMR purchase orders (not MOUs).
Our timeframe columns assume AI diffuses on roughly an industrial-revolution clock — if capability suddenly jumps or hits a wall, the 1–2 year rankings are the first output to distrust, because the world re-times faster than we can recalibrate and nothing in the engine currently measures clock speed.
Watch: Agent task-horizon doubling time breaking trend (METR-style benchmarks) and enterprise weekly usage share re-accelerating past roughly 85%.
Our chips-clouds-AI-software rankings treat the dollars flowing between those industries as independent demand — to the extent the same vendor-financed dollar is going in a circle (NVDA-into-neoclouds, Oracle/OpenAI-style contracts), an unwind would hit our whole top tier as one block, and our customer-health inputs, like everyone's in 1999, are calibrated on financed appearances.
Watch: NVDA related-party/investee disclosures, neocloud secured-debt spreads, and Oracle's RPO concentration in one AI counterparty.
Our top-ranked winners assume Western pricing power at the chip, tool, model, and grid-equipment layer — we encode China as regulatory friction, not as the solar-2011 overcapacity flood, so a gradual Chinese price collapse is the shock most likely to make our rankings wrong before we see it.
Watch: SMIC/Hua Hong capex guidance and mature-node wafer pricing, and Chinese tool-maker (Naura/AMEC) revenue growth.
Read across the eight and one shape repeats: the engine is a demand-side model with almost no supply-side representation, and its firmest, highest-conviction rankings sit on top of its most correlated hidden assumptions — Western pricing power, single-supplier dependency, institutional-speed policy, and financing counted as independent demand all stack on the same handful of industries. That's not a reason to distrust the rankings today. It's the reason we keep this list current, and the reason a Taiwan shock or a financing unwind would be treated as evidence that our premise broke, not quietly absorbed into next quarter's numbers.
AI Stock Market Impacts maps AI disruption across 31 industries and roughly 200 cross-industry effects — rankings, direction, and fragility, recalibrated continuously since February 2026, with every parameter change logged and every failure mode named in public.
See the 31-industry matrixWant the full audit, not just the summary? Read What We Refuse to Predict for the forecasting-research half of this page's honesty, see the demand-side half of shock #7 in How Our Matrix Would Have Scored the Metaverse, or see shock #3's retail-scale mechanism in The Ouroboros Trade. Or get the "what moved" note for when a rank changes.