The 12 AI bubbles, ranked by empirical realness

The QuantAbundancia taxonomy starts with twelve editorial "bubbles" — thematic groupings of stocks the AI supercycle is supposed to lift together. They're chosen by hand to express specific theses (compute, memory, power, photonics, robotics, space, quantum, etc.).

The taxonomy is honest about one thing most thematic dashboards aren't: not every editorial bubble passes the empirical test. Some are real co-movement blocs. Others are noise dressed up in narrative. The data tells us which is which, and we publish both answers without hedging.

Below: the twelve, ranked by their residualized intra-cluster correlation over a 252-day window — the methodology test we walk through in detail in the residualization article. The short version: strip $SPY beta from every constituent, then measure how tightly the residuals still cluster. If that number stays high, the bubble is a real bloc you can trade as one position. If it collapses near zero, the apparent co-movement was just everyone moving with the market, and there's no thematic edge.

Tier 1 — strongest, validated bubbles (residualized > 0.70)

These are the marquee blocs. Not many editorial themes survive the test this cleanly. When they do, the trade is to size the bubble as a single exposure, because the constituents will move together regardless of which name you pick.

🟢 Quantum — 0.76 residualized

The strongest bubble in the entire taxonomy. $QBTS, $RGTI, $IONQ, $ARQQ, $QUBT co-move on news that has nothing to do with broad-market sentiment. Pre-revenue speculation on a single shared thesis (post-GPU compute optionality) means they trade synchronously when that narrative moves.

Pure binary outcome at the bubble level: if any of these companies cracks fault-tolerant qubits before 2030, the entire bloc gets revalued upward. If none does, they deflate together. Position sizing should reflect that — small basket, treat as one bet.

Live Quantum bubble dashboard · or read the Quantum deep-dive.

🟢 Semi Equipment — 0.82 residualized

The single tightest bloc we measure. $ASML, $AMAT, $LRCX, $KLAC. There are roughly five companies on Earth that make the machines that make the chips that make AI, and four of them trade as one stock.

The reason is structural: when TSMC orders one EUV scanner, ASML wins. When Intel adds 100k wafers/year of capacity, AMAT + LRCX + KLAC all win pro-rata. Same buyer, same demand cycle, same earnings response. Duopoly economics, synchronized.

The cleanest single expression of "AI capex is real and accelerating" — it's the picks-and-shovels for the entire stack.

Semi Equipment bubble dashboard.

🟢 Memory / HBM — validated, ~0.71

$MU, plus SK Hynix and Samsung via ADRs, with $WDC and $STX as secondary tier. Three companies make ~95% of HBM stacks, and HBM is the pacing constraint on AI compute. Oligopoly pricing means when one raises HBM3E ASPs, the others follow within weeks — synchronized signal.

The cleanest "AI is real" trade because it doesn't require picking the right GPU vendor or betting on any specific datacenter. Pure exposure to AI compute volume × HBM ASP.

Memory / HBM bubble dashboard.

Tier 2 — validated, mid-tier (residualized 0.55–0.70)

These pass the test but with looser coherence — usually because one or two constituents have idiosyncratic stories that pull the bloc apart, even though the demand signal is shared.

🟢 Compute / GPUs — moderate, ~0.65

$NVDA, $AMD, $AVGO, $TSM, $INTC, $MRVL, $MU, $ARM. The bloc exists. Demand signal (AI capex) is real and shared. But NVDA's idiosyncratic moves — margin extension, MI300 vs Blackwell competitive dynamics — distort the bloc and prevent perfect co-movement.

If you're bullish AI capex broadly but agnostic on NVDA-specific margin trajectory, $SMH is a cleaner expression than concentrating on NVDA alone.

🟢 Networking / Optical — validating, ~0.60

$AVGO, $ANET, $MRVL, $COHR, $CIEN, $LITE. The bandwidth layer that connects AI clusters. Lags semis by 1–2 quarters in earnings response — translation: when NVDA's order book swells, the optical bloc's earnings haven't caught up yet, opening a window where the read is ahead of the price.

🟢 Datacenter Power — validating, ~0.55

$VST, $CEG, $NRG, $GEV, $ETR, $SO, $AEP. AI eats electricity, and these are the utilities + IPPs supplying the AI campuses. Looser correlation than Memory because regulatory regimes are local — Texas (VST) ≠ Mid-Atlantic (CEG) ≠ South (SO) — but the demand signal (multi-GW PPAs from hyperscalers) is uniform.

🟢 Nuclear / SMR — speculative, 0.70+ on the pre-revenue subset

$OKLO, $NNE, $SMR, $NPWR. Pure-play SMR speculation. The pre-revenue names trade tighter than the incumbents ($LEU, $CCJ, $BWXT) because they have nothing else going on except a single shared thesis: AI baseload. Volatility is the price of admission. Right-tail expected payoff if SMRs commercialize on schedule (~2030).

Tier 3 — emerging or quiet compounders (residualized 0.55–0.75)

🟢 Cooling / DC Infra — validated, ~0.70

$VRT, plus Schneider Electric (SU.PA) and $TT, $EMR, $JCI, $GNRC. The most underrated bloc in the taxonomy. Liquid cooling is mandatory at AI rack densities (~30 kW/rack vs 10 kW for traditional). The bloc trades tightly because every name sells different products to the same customer (a hyperscaler) at the same time.

Boring. Wins. Cooling / DC Infra dashboard.

🟢 Space / Sat Comms — validated, 0.55–0.65 with brutal volatility

$RKLB, $ASTS, $IRDM, $ATRO, plus $LMT, $NOC. The only bubble that has both validated bloc behavior AND triple-digit annualized volatility. The bloc is real — they all bid for the same orbital infrastructure — but binary milestones (rocket launches, FCC approvals, capital raises) make individual moves whip the entire group.

Position size matters more than entry timing here.

Tier 0 — failed bubbles (residualized < 0.10)

We publish these too. The taxonomy doesn't gain anything by hiding the bubbles that don't work.

🔴 AI-Capex Hyperscalers — collapses to ~0.05 residualized

$MSFT, $GOOGL, $AMZN, $META, $ORCL. Looks like a bubble at first — raw 252d correlation around 0.65. Residualize and the number collapses to near zero. The raw correlation was driven entirely by these names being mega-caps with similar betas — they move with the market, not with each other on any AI-specific signal.

The methodology-honesty piece on this is in Hyperscalers: the failed bubble.

🔴 AI Software & Data — collapses to ~0.10 residualized

$PLTR, $ORCL, $NOW, $SNOW, $DDOG, $MDB, $NET, $CRM, $IBM, $SAP. Too heterogeneous. PLTR moves on political risk. ORCL on legacy infrastructure refresh. The data-infra subset (SNOW + DDOG + MDB) does cluster more tightly when you isolate them — the real bloc is two layers deeper than the editorial label.

The lesson: editorial taxonomies that group by story (e.g. "all AI software") fail when constituent revenue exposures actually differ by factor.

🔴 Robotics / Automation — collapses to ~0.20 residualized

$TSLA (Optimus), ABB, Fanuc, $SYM, $NDSN. Different end markets entirely. TSLA on FSD news, SYM on Walmart guidance, Fanuc on Asian factory capex. No common signal. Could become a real bloc post-2027 if a humanoid platform commercializes — until then it's a bag of unrelated bets.

What this means in practice

Three actionable conclusions:

1. Position by bubble, not by name. When the residualized correlation is high (Quantum, Semi Equipment), constituent risk and reward are highly substitutable — your sizing budget is for the bubble, not the name. The opposite holds for failed bubbles: each constituent of Hyperscalers is its own idiosyncratic bet that needs its own analysis.

2. Hedge with the right ETF. Tight residualized clusters can be hedged with the bloc's primary ETF ($SOXX for Semi Equipment exposure, $QTUM for Quantum). Failed clusters can't be hedged because they don't move together coherently.

3. The brand-narrative trap. "AI software" sounds like a unified bloc. The data says it isn't. "Cooling" sounds like a boring infrastructure category. The data says it's a tightly-correlated, fast-compounding bloc. The reverse of what most retail intuition would predict — and that gap is the alpha.

Methodology footnote

All correlations: 252 trading days of daily log returns, residualized against $SPY, averaged pairwise across all bubble members. Sample sizes vary by bubble (some have 4 members, some have 10) so the standard error on the average correlation differs — generally tighter for larger bubbles, looser for the speculative ones with 4–5 names.

We update the matrix nightly after the US close. The numbers above are point-in-time as of the most recent publication; the live values on each bubble page may have drifted since.

---

If you want the daily snapshot of which bubbles are leading and which are lagging, the free email digest ships nightly with bubble-level moves, top constituent gainers, and Fib-level approaches across the universe. No spam, no upsell, no paywall.