AI Infrastructure Sector: Comprehensive Quantitative Research
AI Infrastructure Sector: Comprehensive Quantitative Research
| Field | Value |
|---|---|
| Date | 2026-03-09 |
| Researcher | Quant Researcher (Claude Opus 4.6) |
| Status | Complete |
| Script | analysis/quant-research/scripts/ai-infra-sector-2026-03-09.py |
| Charts | charts/ai_infra_normalized_prices.png, ai_infra_risk_return.png, ai_infra_drawdowns.png, ai_infra_correlations.png, ai_infra_volatility_boxplot.png, ai_infra_scorecard.png, ai_infra_dendrogram.png, ai_infra_monthly_drops.png |
1. Hypothesis
Primary hypothesis: The AI Infrastructure watchlist universe contains distinct sub-sectors (miners/compute, semiconductors, power/infrastructure, networking) with meaningfully different risk-return profiles. Identifying which sub-sector and which specific names offer the best risk-adjusted entry points during the current market pullback is the primary objective.
Secondary hypotheses:
- H2: Miners/compute names (WULF, CLSK, CIFR, IREN, WGMI) exhibit extreme volatility that renders them poor risk-adjusted vehicles despite high absolute return potential.
- H3: Power/infrastructure names (VST, PWR, FIX, VRT, BE, LEU) offer superior Sharpe ratios due to more stable business models with AI-driven demand.
- H4: The sub-sectors cluster statistically (by return correlation) in a way that validates treating them as distinct groups rather than one homogeneous "AI infra" basket.
- H5: ETFs (SOXX, SOXL, COMM, PRYMY, MNRS, AAOI) provide meaningful diversification with lower tail risk.
Context: User is evaluating this 30-name watchlist during a market pullback to identify which names merit close monitoring for buy-point entries vs. which should be avoided.
2. Data
Source: Local OHLCV CSVs from yfinance, stored in data-samples/ohlcv/.
Universe:
| Sub-sector | Tickers | Description |
|---|---|---|
| Miners/Compute | WULF, CLSK, CIFR, IREN, WGMI | BTC miners / AI compute providers |
| Semiconductors | NVDA, AVGO, TSM, MU, ALAB, COHR, GLW, LITE, SNDK, WDC | Chip designers, foundries, memory, photonics |
| Power/Infrastructure | VST, PWR, FIX, VRT, BE, LEU | Nuclear, electrical contractors, cooling, energy |
| Networking | APH, JBL | Connectors, contract manufacturing |
| ETFs | SOXX, SOXL, COMM, PRYMY, MNRS, AAOI | Sector ETFs + leveraged |
| Benchmarks | SPY, QQQ | Broad market references |
Date ranges: Most names have full 2018-01-02 to 2026-03-09 history (~2,055 bars). Notable exceptions:
- ALAB: 493 bars (IPO 2024-03-20) -- short history, results less reliable
- WGMI: 1,023 bars (ETF launched 2022-02-08)
- CIFR: 1,351 bars (IPO 2020-10-20)
- IREN: 1,077 bars (IPO ~2021-11)
- SNDK: 265 bars only -- risk metrics unreliable, included with caveat
- MNRS: 277 bars -- very short, included with caveat
Excluded (< 252 bars): SOLS (96 bars, IPO 2025-10-20)
Common overlap period: Used 80% ticker threshold; overlap varies by analysis section.
Data quality notes: - SNDK has only ~1 year of data and shows an extreme annualized return (306%) driven by a short high-momentum sample. This is not representative of long-run behavior and should be treated as an outlier. - CLSK has extreme kurtosis (505) driven by one -52% daily move, distorting all distribution-based metrics. - AAOI is classified as "ETF" per the watchlist but is actually a single stock (Applied Optoelectronics). Its behavior is fully idiosyncratic.
3. Methodology
All computation performed in Python via .venv/bin/python. No statistics computed mentally.
Analyses performed:
- Full-history return statistics: Annualized return, volatility, Sharpe, Sortino, Calmar, max drawdown, skewness, kurtosis
- Drawdown analysis: Peak-to-trough drawdowns from common period, time underwater, current drawdown depth
- Correlation analysis: Intra-subsector, cross-subsector, and ETF-to-company pairwise correlations
- Volatility profiling: 20-day realized volatility distribution, median, 95th percentile, vol-of-vol
- Regime-conditional performance: Bull/bear regime split using SPY 60-day return moving average
- Tail risk: VaR (95%), CVaR (95%), worst day, frequency of >5% daily drops
- Relative strength: Rolling 1M/3M/6M/12M cumulative return vs. SPY
- Liquidity analysis: 20-day average dollar volume, high-low spread proxy
- Stop-loss frequency: How often 10%, 15%, 20% trailing stops would fire per year
- Sub-sector clustering: Hierarchical clustering on return correlations to validate sub-sector labels
- Composite pullback opportunity scoring: Weighted combination of Sharpe, drawdown depth, bear resilience, volatility, liquidity, and relative strength
Regime definition: Bull = SPY 60-day average daily return > 0; Bear = SPY 60-day average daily return <= 0.
4. Results
4.1 Risk-Return Summary
Ranked by Sharpe ratio (full history):
| Ticker | Sub-sector | Ann Return | Ann Vol | Sharpe | Sortino | Max DD | Calmar |
|---|---|---|---|---|---|---|---|
| SNDK* | Semis | 306.0% | 100.5% | 3.05 | 4.70 | -47.5% | 6.44 |
| FIX | Power/Infra | 52.6% | 43.8% | 1.20 | 1.62 | -49.7% | 1.06 |
| PWR | Power/Infra | 39.1% | 34.8% | 1.12 | 1.57 | -45.5% | 0.86 |
| NVDA | Semis | 57.0% | 51.0% | 1.12 | 1.64 | -66.3% | 0.86 |
| VRT | Power/Infra | 58.6% | 54.5% | 1.07 | 1.32 | -71.2% | 0.82 |
| AVGO | Semis | 42.3% | 41.3% | 1.02 | 1.44 | -48.3% | 0.88 |
| TSM | Semis | 36.1% | 36.1% | 1.00 | 1.56 | -56.5% | 0.64 |
| LEU | Power/Infra | 84.7% | 86.2% | 0.98 | 1.56 | -78.2% | 1.08 |
| APH | Networking | 27.6% | 29.1% | 0.95 | 1.19 | -37.6% | 0.74 |
| JBL | Networking | 35.3% | 38.4% | 0.92 | 1.22 | -57.3% | 0.62 |
| VST | Power/Infra | 38.5% | 43.9% | 0.88 | 1.12 | -53.3% | 0.72 |
| LITE | Semis | 47.1% | 55.9% | 0.84 | 1.10 | -66.9% | 0.71 |
| ALAB* | Semis | 76.2% | 91.8% | 0.83 | 1.31 | -63.7% | 1.20 |
| QQQ | Benchmark | 19.8% | 23.9% | 0.83 | 1.08 | -35.1% | 0.57 |
| SOXX | ETF | 28.7% | 35.2% | 0.82 | 1.16 | -45.8% | 0.63 |
| MU | Semis | 39.0% | 49.9% | 0.78 | 1.17 | -57.6% | 0.68 |
| SPY | Benchmark | 14.7% | 19.3% | 0.76 | 0.93 | -33.7% | 0.44 |
| GLW | Semis | 24.7% | 33.7% | 0.73 | 1.03 | -48.8% | 0.51 |
| SOXL | ETF (3x) | 75.9% | 104.2% | 0.73 | 1.04 | -90.5% | 0.84 |
| BE | Power/Infra | 65.6% | 94.4% | 0.70 | 1.15 | -92.5% | 0.71 |
| PRYMY | ETF | 24.9% | 37.4% | 0.67 | 0.93 | -53.9% | 0.46 |
| IREN | Miners | 78.3% | 119.5% | 0.66 | 1.26 | -95.7% | 0.82 |
| CIFR | Miners | 69.9% | 113.5% | 0.62 | 1.05 | -97.2% | 0.72 |
| WDC | Semis | 30.6% | 49.8% | 0.61 | 0.86 | -70.5% | 0.43 |
| WULF | Miners | 67.5% | 110.0% | 0.61 | 0.97 | -98.5% | 0.69 |
| AAOI | ETF** | 58.6% | 99.5% | 0.59 | 0.96 | -96.9% | 0.60 |
| CLSK | Miners | 104.4% | 192.7% | 0.54 | 1.27 | -98.6% | 1.06 |
| WGMI | Miners | 41.6% | 82.3% | 0.51 | 0.91 | -85.8% | 0.49 |
SNDK (265 bars) and ALAB (492 bars) have short histories. Their metrics are less reliable.
*AAOI is a single stock, not an ETF. Listed under ETFs per watchlist categorization but should be evaluated as a single name.
Key findings:
- Power/Infrastructure dominates on risk-adjusted basis. FIX (1.20), PWR (1.12), VRT (1.07), VST (0.88) all have Sharpe ratios above 0.85 with moderate volatility (34-54%). This sub-sector offers the best risk/reward in the AI infra universe.
- Semis mega-caps are strong. NVDA (1.12), AVGO (1.02), TSM (1.00) all cross the 1.0 Sharpe threshold with manageable volatility relative to their returns.
- Networking is quietly excellent. APH (0.95) and JBL (0.92) deliver strong Sharpes with the lowest volatility among non-benchmark names (29-38%).
- Miners are return-rich but Sharpe-poor. Despite annualized returns of 42-104%, miners have Sharpes of only 0.51-0.66 because their 82-193% annualized volatility overwhelms the return premium.
- CLSK is the most extreme name: 104% annualized return but 193% annualized vol and a 98.6% max drawdown. The kurtosis of 505 (driven by a single -52% day) makes this a lottery ticket.
4.2 Drawdown Analysis
Current drawdown depth and historical severity:
| Ticker | Sub-sector | Max DD | Current DD | % Time >5% Underwater |
|---|---|---|---|---|
| CLSK | Miners | -98.6% | -87.4% | 98% |
| WULF | Miners | -98.5% | -61.9% | 98% |
| CIFR | Miners | -97.2% | -42.1% | 97% |
| AAOI | Single* | -96.9% | -6.8% | 88% |
| IREN | Miners | -95.7% | -52.0% | 90% |
| BE | Power/Infra | -92.5% | -22.6% | 90% |
| SOXL | ETF (3x) | -90.5% | -25.8% | 77% |
| WGMI | Miners | -85.8% | -43.4% | 85% |
| LEU | Power/Infra | -78.2% | -55.1% | 74% |
| VRT | Power/Infra | -71.2% | 0.0% | 63% |
| WDC | Semis | -70.5% | -17.3% | 68% |
| LITE | Semis | -66.9% | -18.2% | 65% |
| NVDA | Semis | -66.3% | -14.1% | 56% |
| ALAB | Semis | -63.7% | -51.4% | 61% |
| JBL | Networking | -57.3% | -10.8% | 61% |
| MU | Semis | -57.6% | -15.4% | 59% |
| TSM | Semis | -56.5% | -3.4% | 53% |
| VST | Power/Infra | -53.3% | -24.8% | 56% |
| FIX | Power/Infra | -49.7% | -6.5% | 48% |
| AVGO | Semis | -48.3% | -19.8% | 51% |
| SOXX | ETF | -45.8% | -8.6% | 49% |
| PWR | Power/Infra | -45.5% | -0.8% | 41% |
| APH | Networking | -37.6% | -18.2% | 37% |
| SPY | Benchmark | -33.7% | -7.1% | 38% |
VRT is at its all-time high (0% drawdown). PWR (-0.8%), TSM (-3.4%), and FIX (-6.5%) are also near highs, meaning these names have NOT pulled back meaningfully yet.
Deep value pullback zone: ALAB (-51.4%), LEU (-55.1%), WULF (-61.9%), CLSK (-87.4%) are in deep drawdowns. The question is whether these are damaged goods or opportunity -- the regime analysis (Section 4.6) helps answer this.
Ideal pullback zone (15-25% from highs): AVGO (-19.8%), APH (-18.2%), LITE (-18.2%), MU (-15.4%), GLW (-23.0%), VST (-24.8%). These have proven quality (high Sharpe) and are experiencing moderate drawdowns that create entry opportunities.
4.3 Correlation Structure
Intra-subsector average correlations:
| Sub-sector | Avg Correlation | Min | Max | Pairs |
|---|---|---|---|---|
| Miners | 0.593 | 0.209 | 0.860 | 10 |
| Semis | 0.517 | 0.317 | 0.731 | 45 |
| Power/Infra | 0.376 | 0.221 | 0.641 | 15 |
| Networking | 0.646 | 0.646 | 0.646 | 1 |
Cross-subsector average correlations:
| Pair | Avg Correlation |
|---|---|
| Miners vs Semis | 0.260 |
| Miners vs Power/Infra | 0.255 |
| Miners vs Networking | 0.260 |
| Semis vs Power/Infra | 0.382 |
| Semis vs Networking | 0.542 |
| Power/Infra vs Networking | 0.445 |
Highly correlated pairs (>0.70):
| Pair | Correlation | Implication |
|---|---|---|
| CLSK vs WGMI | 0.860 | Redundant miner exposure |
| CIFR vs WGMI | 0.751 | Redundant miner exposure |
| IREN vs WGMI | 0.760 | Redundant miner exposure |
| MU vs WDC | 0.731 | Redundant memory exposure |
Key findings:
- Miners are highly correlated internally (avg 0.59) but weakly correlated with everything else (avg 0.26). This means miners are a distinct risk factor -- they diversify a semi/infra portfolio but adding multiple miners provides diminishing diversification benefit.
- Semis and Networking are moderately correlated (0.54), consistent with both being driven by AI capex cycles.
- Power/Infra is the most diversifying sub-sector -- low correlation with miners (0.26) and moderate with semis (0.38).
- WGMI, CLSK, CIFR, IREN are nearly interchangeable from a portfolio perspective. Owning more than one miner adds little diversification.
ETF correlations with company universe:
| ETF | Avg Corr | Max Corr | Best Captures |
|---|---|---|---|
| SOXX | 0.548 | 0.812 | Broad semi exposure |
| SOXL | 0.548 | 0.810 | Same as SOXX (3x leveraged) |
| MNRS | 0.559 | 0.984 | Miners (near-perfect with WGMI) |
| PRYMY | 0.359 | 0.497 | Moderate broad coverage |
| AAOI | 0.308 | 0.426 | Idiosyncratic (not an ETF) |
| COMM | 0.269 | 0.376 | Weak -- doesn't capture AI infra well |
SOXX is the best ETF proxy for broad AI infra exposure. MNRS is essentially a miner basket (0.98 correlation with WGMI). COMM provides poor sector coverage.
4.4 Volatility Profiles
| Ticker | Sub-sector | Current 20d Vol | Median 20d Vol | 95th Pctl | Vol of Vol |
|---|---|---|---|---|---|
| IREN | Miners | 68.1% | 108.7% | 216.3% | 43.2% |
| CIFR | Miners | 55.5% | 106.7% | 170.4% | 36.0% |
| CLSK | Miners | 44.1% | 100.4% | 203.1% | 90.0% |
| WULF | Miners | 80.2% | 98.9% | 171.2% | 36.4% |
| SNDK* | Semis | 72.2% | 90.8% | 145.7% | 31.1% |
| SOXL | ETF (3x) | 48.8% | 90.0% | 127.9% | 24.1% |
| ALAB* | Semis | 58.7% | 84.1% | 168.2% | 36.2% |
| WGMI | Miners | 38.7% | 79.8% | 127.7% | 28.3% |
| AAOI | Single* | 53.3% | 77.5% | 145.3% | 34.3% |
| BE | Power/Infra | 67.6% | 76.7% | 144.3% | 31.1% |
| LEU | Power/Infra | 78.1% | 76.6% | 152.5% | 38.7% |
| MNRS* | ETF | 39.2% | 68.8% | -- | -- |
| COMM | ETF | 12.3% | 59.1% | 102.1% | 38.0% |
| LITE | Semis | 43.5% | 43.5% | 82.2% | 17.9% |
| MU | Semis | 34.4% | 43.7% | 74.3% | 16.2% |
| COHR | Semis | 48.3% | 48.6% | 73.0% | 12.6% |
| VRT | Power/Infra | 51.1% | 42.2% | 64.7% | 16.1% |
| WDC | Semis | 39.5% | 42.6% | 72.4% | 15.2% |
| NVDA | Semis | 39.1% | 41.7% | 70.1% | 15.2% |
| VST | Power/Infra | 43.7% | 31.1% | 54.1% | 14.3% |
| AVGO | Semis | 28.3% | 33.8% | 55.2% | 11.3% |
| FIX | Power/Infra | 23.9% | 32.9% | 56.6% | 12.3% |
| JBL | Networking | 29.1% | 32.5% | 59.5% | 12.9% |
| PWR | Power/Infra | 26.3% | 29.6% | 49.7% | 10.3% |
| TSM | Semis | 39.7% | 31.4% | 53.9% | 12.3% |
| PRYMY | ETF | 22.7% | 31.5% | 62.7% | 18.2% |
| SOXX | ETF | 27.0% | 30.3% | 48.1% | 9.3% |
| GLW | Semis | 32.2% | 26.8% | 46.1% | 11.3% |
| APH | Networking | 28.0% | 22.5% | 38.6% | 9.3% |
| QQQ | Benchmark | 20.7% | 21.0% | 37.8% | 9.1% |
| SPY | Benchmark | 19.2% | 16.3% | 30.9% | 8.2% |
The volatility gap between sub-sectors is enormous.
- Miners: Median 20d vol of 80-109%. These stocks routinely move with annualized vol near 100%.
- Semis: Median 34-49%. Manageable for active traders.
- Power/Infra: Median 29-42% (excluding BE and LEU outliers). PWR at 29.6% is the lowest-vol AI infra name.
- Networking: Median 22-33%. APH at 22.5% is nearly as stable as QQQ.
Vol-of-vol matters for position sizing. CLSK's vol-of-vol is 90% -- meaning its volatility is itself wildly unpredictable. APH's vol-of-vol is 9.3%, meaning you can reliably size positions. The practical difference: a position size calculated for CLSK could be wrong by 2-3x within a month.
4.5 Regime-Conditional Performance
| Ticker | Sub-sector | Bull Sharpe | Bear Sharpe | Resilience |
|---|---|---|---|---|
| PWR | Power/Infra | 1.77 | 0.55 | STRONG |
| ALAB* | Semis | 1.32 | 0.50 | STRONG |
| FIX | Power/Infra | 1.85 | 0.02 | STRONG |
| AVGO | Semis | 1.13 | 0.23 | STRONG |
| LEU | Power/Infra | 0.87 | 0.23 | STRONG |
| VST | Power/Infra | 0.83 | 0.13 | STRONG |
| CLSK | Miners | 0.87 | 0.09 | STRONG |
| APH | Networking | 0.85 | -0.04 | MODERATE |
| JBL | Networking | 1.48 | -0.10 | MODERATE |
| TSM | Semis | 1.63 | -0.22 | MODERATE |
| NVDA | Semis | 1.52 | -0.28 | MODERATE |
| VRT | Power/Infra | 1.79 | -0.36 | MODERATE |
| COHR | Semis | 1.01 | -0.38 | MODERATE |
| CIFR | Miners | 0.83 | -0.46 | MODERATE |
| SOXX | ETF | 1.55 | -0.54 | WEAK |
| GLW | Semis | 0.98 | -0.80 | WEAK |
| MU | Semis | 1.27 | -0.80 | WEAK |
| SPY | Benchmark | 2.09 | -0.87 | Baseline |
| BE | Power/Infra | 1.26 | -0.95 | WEAK |
| WULF | Miners | 1.19 | -1.22 | WEAK |
| WDC | Semis | 1.60 | -1.26 | WEAK |
| LITE | Semis | 1.73 | -1.34 | WEAK |
| AAOI | Single* | 1.27 | -1.40 | WEAK |
| SNDK* | Semis | 4.42 | -1.42 | WEAK |
| WGMI | Miners | 1.23 | -1.57 | WEAK |
| IREN | Miners | 1.47 | -1.69 | WEAK |
Critical insight: Power/Infrastructure names dominate in bear resilience. PWR, FIX, VST, and LEU all maintain positive Sharpe in bear markets. This makes them the best "buy during pullback" candidates because they hold value when the market is selling.
NVDA and AVGO show moderate-to-strong bear resilience -- these mega-cap semis are less cyclical than smaller names.
Miners are deeply regime-dependent. IREN (-1.69), WGMI (-1.57), WULF (-1.22) have the worst bear-market Sharpes. Buying these during a pullback means betting on a regime change -- if the bear persists, they accelerate losses.
4.6 Tail Risk
| Ticker | Sub-sector | VaR (95%) | CVaR (95%) | Worst Day | Days <-5% |
|---|---|---|---|---|---|
| CLSK | Miners | -10.15% | -15.35% | -52.19% | 20.4% |
| SOXL | ETF (3x) | -10.72% | -15.22% | -38.59% | 16.9% |
| WULF | Miners | -9.58% | -14.36% | -34.95% | 17.0% |
| IREN | Miners | -10.32% | -13.39% | -36.45% | 21.6% |
| CIFR | Miners | -9.78% | -13.22% | -46.75% | 19.6% |
| AAOI | Single* | -8.45% | -12.81% | -32.18% | 13.0% |
| SNDK* | Semis | -6.65% | -12.61% | -21.30% | 9.8% |
| ALAB* | Semis | -8.24% | -11.88% | -28.03% | 11.2% |
| BE | Power/Infra | -7.58% | -11.68% | -42.50% | 12.6% |
| LEU | Power/Infra | -7.51% | -11.41% | -31.42% | 11.9% |
| WGMI | Miners | -7.63% | -9.88% | -20.75% | 15.4% |
| NVDA | Semis | -4.80% | -7.11% | -18.76% | 4.8% |
| VRT | Power/Infra | -4.82% | -7.64% | -36.74% | 4.6% |
| VST | Power/Infra | -3.86% | -6.44% | -28.27% | 2.7% |
| FIX | Power/Infra | -3.44% | -6.17% | -25.71% | 2.3% |
| AVGO | Semis | -3.75% | -5.63% | -19.91% | 2.1% |
| TSM | Semis | -3.40% | -4.79% | -14.03% | 1.5% |
| APH | Networking | -2.70% | -4.45% | -13.88% | 1.1% |
| SOXX | ETF | -3.58% | -5.11% | -15.23% | 1.8% |
| PWR | Power/Infra | -3.21% | -4.96% | -18.32% | 1.8% |
| GLW | Semis | -2.92% | -4.85% | -16.40% | 1.6% |
| JBL | Networking | -3.36% | -5.68% | -16.49% | 2.1% |
| QQQ | Benchmark | -2.45% | -3.58% | -11.98% | 0.5% |
| SPY | Benchmark | -1.78% | -2.97% | -10.94% | 0.3% |
IREN drops >5% on 21.6% of all trading days -- more than one day per week. CLSK at 20.4% is similar. Compare with APH at 1.1% or PWR at 1.8%.
CLSK's -52.19% worst single day is catastrophic for any position. Even at half-size allocation, that is a -26% portfolio impact in one session.
4.7 Relative Strength (Current)
3-Month excess return vs SPY (most relevant for pullback timing):
| Ticker | Sub-sector | 3M Return | Excess vs SPY |
|---|---|---|---|
| AAOI | Single* | +164.5% | +165.8% |
| SNDK* | Semis | +115.2% | +116.5% |
| LITE | Semis | +79.2% | +80.5% |
| WDC | Semis | +54.0% | +55.2% |
| MU | Semis | +52.2% | +53.4% |
| GLW | Semis | +44.1% | +45.4% |
| VRT | Power/Infra | +43.4% | +44.7% |
| BE | Power/Infra | +42.7% | +43.9% |
| COHR | Semis | +42.3% | +43.5% |
| FIX | Power/Infra | +35.0% | +36.3% |
| SOXL | ETF (3x) | +30.2% | +31.5% |
| TSM | Semis | +29.2% | +30.4% |
| PWR | Power/Infra | +22.0% | +23.3% |
| PRYMY | ETF | +18.7% | +20.0% |
| JBL | Networking | +14.6% | +15.9% |
| SOXX | ETF | +11.0% | +12.3% |
| WULF | Miners | +8.0% | +9.3% |
| NVDA | Semis | +0.6% | +1.8% |
| APH | Networking | +1.5% | +2.8% |
| AVGO | Semis | -8.9% | -7.7% |
| ALAB | Semis | -7.1% | -5.8% |
| WGMI | Miners | -11.6% | -10.4% |
| IREN | Miners | +0.1% | -1.4% |
| CIFR | Miners | -16.8% | -15.6% |
| CLSK | Miners | -29.7% | -28.5% |
| LEU | Power/Infra | -20.1% | -18.9% |
NVDA is notable for weak recent momentum (+1.8% excess over 3M) despite strong fundamentals. AVGO is actually negative over 3 months (-7.7% excess). Both mega-cap semis are lagging -- this is the pullback the user is watching.
Memory/photonics names (LITE, WDC, MU, GLW, COHR) are showing strong momentum -- this is the AI infrastructure buildout trade rotating into less-obvious supply chain names.
4.8 Liquidity Analysis
| Ticker | Sub-sector | Avg 20d $ Vol | HL Spread |
|---|---|---|---|
| NVDA | Semis | $36,372M | 3.16% |
| MU | Semis | $13,357M | 3.95% |
| SNDK* | Semis | $12,059M | 6.35% |
| AVGO | Semis | $8,293M | 2.77% |
| SOXL | ETF (3x) | $5,135M | 3.69% |
| TSM | Semis | $4,507M | 2.66% |
| LITE | Semis | $3,669M | 5.06% |
| SOXX | ETF | $2,717M | 1.68% |
| WDC | Semis | $2,639M | 4.47% |
| GLW | Semis | $1,944M | 3.87% |
| VRT | Power/Infra | $1,895M | 4.36% |
| COHR | Semis | $1,799M | 4.30% |
| BE | Power/Infra | $1,706M | 6.57% |
| IREN | Miners | $1,549M | 8.00% |
| APH | Networking | $1,310M | 3.07% |
| ALAB | Semis | $821M | 6.00% |
| VST | Power/Infra | $875M | 3.34% |
| PWR | Power/Infra | $632M | 3.91% |
| WULF | Miners | $573M | 9.29% |
| FIX | Power/Infra | $695M | 3.88% |
| AAOI | Single* | $655M | 11.63% |
| CIFR | Miners | $461M | 9.34% |
| JBL | Networking | $292M | 4.13% |
| LEU | Power/Infra | $213M | 8.24% |
| CLSK | Miners | $235M | 7.85% |
| COMM | ETF | $72M | 4.81% |
| WGMI | Miners | $21M | 6.66% |
| PRYMY | ETF | $5M | 2.92% |
| MNRS | ETF | $0.3M | 4.05% |
PRYMY ($5M/day) and MNRS ($0.3M/day) are effectively illiquid. These are not viable trading vehicles for any meaningful position size.
WGMI at $21M/day is borderline -- acceptable for small positions only.
Miners have wide spreads (7.8-9.3%) vs. networking (3.1-4.1%) and large-cap semis (2.7-4.0%). This spread cost compounds on active trading.
4.9 Stop-Loss Frequency (15% Trailing)
| Ticker | Sub-sector | 15% Stops/Year |
|---|---|---|
| SNDK* | Semis | 2.8 |
| LEU | Power/Infra | 2.1 |
| SOXL | ETF (3x) | 2.0 |
| ALAB* | Semis | 2.0 |
| NVDA | Semis | 1.6 |
| PWR | Power/Infra | 1.6 |
| AVGO | Semis | 1.5 |
| MU | Semis | 1.5 |
| WGMI | Miners | 1.5 |
| VRT | Power/Infra | 1.3 |
| JBL | Networking | 1.3 |
| LITE | Semis | 1.2 |
| VST | Power/Infra | 1.2 |
| GLW | Semis | 1.2 |
| APH | Networking | 1.1 |
| COHR | Semis | 1.1 |
| FIX | Power/Infra | 1.0 |
| TSM | Semis | 1.0 |
| SOXX | ETF | 0.9 |
| PRYMY | ETF | 0.9 |
| BE | Power/Infra | 0.8 |
| WULF | Miners | 0.7 |
| CIFR | Miners | 0.7 |
| IREN | Miners | 0.5 |
| AAOI | Single* | 0.4 |
| CLSK | Miners | 0.4 |
| WDC | Semis | 0.4 |
| COMM | ETF | 0.1 |
Miners show paradoxically low stop frequency (0.4-0.7/yr) because once they drop through the stop, they stay down -- the trailing stop fires once and the stock never recovers enough to reset.
The most "stop-loss friendly" names (manageable frequency with good recovery) are FIX (1.0/yr), TSM (1.0/yr), APH (1.1/yr), COHR (1.1/yr), and SOXX (0.9/yr). These oscillate enough to trigger stops but then recover, allowing re-entry.
4.10 Sub-Sector Clustering
Hierarchical clustering on return correlations identified 4 natural groups:
| Cluster | Members | Our Labels | Coherent? |
|---|---|---|---|
| 1 | WULF, CLSK, CIFR, IREN, WGMI | Miners | YES |
| 2 | MU, SNDK, WDC, BE | Semis + Power/Infra | Mixed |
| 3 | NVDA, AVGO, TSM, ALAB, COHR, GLW, LITE, VST, PWR, FIX, VRT, APH, JBL | Semis + Power/Infra + Networking | Large mixed cluster |
| 4 | LEU | Power/Infra | Singleton |
Miners cluster coherently -- our sub-sector label matches the statistical grouping perfectly. These names truly trade as a group.
Semis and Power/Infra are fragmented. The memory names (MU, WDC) cluster with BE rather than with mega-cap semis. The large Cluster 3 mixes semis, power/infra, and networking into one mega-cluster -- statistically, these all respond to the same "AI capex" factor.
LEU stands alone -- nuclear/uranium has its own idiosyncratic drivers (regulatory, geopolitical) that decouple it from the rest of the AI infra theme.
5. Signal Quality Metrics
This research is a vehicle comparison and pullback opportunity assessment, not a directional signal. The standard IC/hit-rate framework is adapted to measure the reliability and persistence of vehicle characteristics.
5.1 Volatility Persistence (Position Sizing Reliability)
| Bucket | Tickers | Vol-of-Vol Range | Sizing Reliability |
|---|---|---|---|
| Highly reliable | APH, SOXX, QQQ, SPY | 8-9% | Excellent |
| Reliable | PWR, GLW, AVGO, TSM, FIX, JBL, COHR | 10-13% | Good |
| Moderate | NVDA, MU, WDC, VRT, VST, LITE | 14-18% | Acceptable |
| Unreliable | WGMI, SOXL, BE, LEU, SNDK | 24-39% | Poor |
| Chaotic | IREN, CIFR, WULF, CLSK | 36-90% | Not usable |
5.2 Regime Consistency (All-Weather Quality)
Names that maintain positive Sharpe across both bull and bear regimes:
| Ticker | Bull Sharpe | Bear Sharpe | Verdict |
|---|---|---|---|
| PWR | 1.77 | +0.55 | All-weather |
| FIX | 1.85 | +0.02 | All-weather (marginal bear) |
| AVGO | 1.13 | +0.23 | All-weather |
| VST | 0.83 | +0.13 | All-weather |
| LEU | 0.87 | +0.23 | All-weather |
| CLSK | 0.87 | +0.09 | All-weather (marginal bear) |
Only 6 names out of 29 are positive in both regimes. This is the quality filter.
5.3 Composite Information: Pullback Opportunity Scorecard
The composite score weights: Sharpe quality (25%), drawdown depth opportunity (20%), bear resilience (15%), volatility manageability (15%), liquidity (15%), 3M relative strength (10%).
Top 10 pullback buy candidates:
| Rank | Ticker | Sub-sector | Score | Why |
|---|---|---|---|---|
| 1 | AVGO | Semis | 6.5 | High Sharpe, ideal DD depth (-20%), bear-positive, liquid |
| 2 | APH | Networking | 6.5 | Excellent Sharpe, ideal DD (-18%), lowest vol, moderate bear |
| 3 | VST | Power/Infra | 6.4 | Strong Sharpe, good DD depth (-25%), bear-positive |
| 4 | SNDK* | Semis | 6.1 | Extreme Sharpe (short history), significant pullback |
| 5 | FIX | Power/Infra | 5.8 | Highest Sharpe, all-weather, near highs though |
| 6 | NVDA | Semis | 5.6 | Quality name, moderate pullback (-14%), high liquidity |
| 7 | GLW | Semis | 5.5 | Decent Sharpe, good DD zone (-23%), moderate vol |
| 8 | MU | Semis | 5.4 | Good Sharpe, moderate DD (-15%), high momentum |
| 9 | PWR | Power/Infra | 5.3 | Best all-weather Sharpe, but near highs (-0.8%) |
| 10 | LITE | Semis | 5.3 | Strong returns, ideal DD depth (-18%), high momentum |
Bottom 5 -- Avoid or require special conviction:
| Rank | Ticker | Sub-sector | Score | Why |
|---|---|---|---|---|
| 26 | WGMI | Miners | 3.5 | Deep DD, terrible bear performance, low liquidity |
| 27 | CLSK | Miners | 2.9 | -87% drawdown, chaotic vol, catastrophic tail risk |
| 28 | IREN | Miners | 2.8 | -52% DD, worst bear Sharpe (-1.69), extreme vol |
| 29 | WULF | Miners | 2.6 | -62% DD, terrible bear Sharpe, 98.9% median vol |
6. Recommended Architecture
6.1 Tiered Watchlist Framework
Based on the evidence, the AI Infra watchlist should be organized into three tiers:
Tier 1 -- Core Holdings (Buy on 15-25% pullbacks):
These names have Sharpe >0.90, bear-regime resilience, manageable volatility, and sufficient liquidity. They are the "buy the dip" names during market weakness.
| Ticker | Sub-sector | Key Metric |
|---|---|---|
| AVGO | Semis | Sharpe 1.02, bear +0.23, vol 33.8% |
| APH | Networking | Sharpe 0.95, lowest vol (22.5%), -18% DD now |
| FIX | Power/Infra | Sharpe 1.20, all-weather, near highs |
| PWR | Power/Infra | Sharpe 1.12, best bear Sharpe (+0.55), near highs |
| NVDA | Semis | Sharpe 1.12, mega-liquid, -14% DD now |
| TSM | Semis | Sharpe 1.00, moderate bear, near highs |
| VST | Power/Infra | Sharpe 0.88, bear-positive, -25% DD now |
Tier 2 -- Opportunistic (Buy on 20-30% pullbacks with tighter stops):
Higher volatility but compensated returns. Require more active management.
| Ticker | Sub-sector | Key Metric |
|---|---|---|
| VRT | Power/Infra | Sharpe 1.07, at ATH, wait for pullback |
| JBL | Networking | Sharpe 0.92, moderate vol, -11% DD |
| MU | Semis | Sharpe 0.78, higher vol, strong momentum |
| GLW | Semis | Sharpe 0.73, moderate vol, -23% DD |
| LITE | Semis | Sharpe 0.84, higher vol, strong momentum |
| COHR | Semis | Sharpe 0.64, photonics growth, -16% DD |
| LEU | Power/Infra | Sharpe 0.98, but -55% DD and extreme vol |
Tier 3 -- Speculative / Avoid:
Poor risk-adjusted returns, extreme volatility, or insufficient diversification benefit.
| Ticker | Sub-sector | Reason |
|---|---|---|
| WULF | Miners | 99% median vol, -62% DD, terrible bear Sharpe |
| CLSK | Miners | 100% median vol, -87% DD, 505 kurtosis |
| IREN | Miners | 109% median vol, worst bear Sharpe (-1.69) |
| CIFR | Miners | 107% median vol, -42% DD |
| WGMI | Miners | 80% median vol, low liquidity ($21M/day) |
| BE | Power/Infra | 77% median vol, -92.5% historical max DD |
| SOXL | ETF (3x) | Volatility decay, -91% max DD, use SOXX instead |
| AAOI | Single* | 78% vol, -97% max DD, worst bear Sharpe |
| COMM | ETF | Sharpe 0.26, poor sector capture |
| PRYMY | ETF | $5M/day liquidity -- not tradeable |
| MNRS | ETF | $0.3M/day liquidity -- not tradeable |
| ALAB* | Semis | Only 2 years of data, -51% DD, needs more history |
ETF Selection: If seeking broad AI infra exposure via ETF, SOXX is the clear winner -- best liquidity among sector ETFs ($2.7B/day), 0.82 Sharpe, manageable drawdown (-46%). Avoid SOXL (leveraged), COMM (poor fit), PRYMY/MNRS (illiquid).
6.2 Position Sizing by Sub-sector
For a target portfolio volatility of 20%:
| Sub-sector | Representative Vol | Max Position Size |
|---|---|---|
| Networking (APH, JBL) | 22-33% | 60-90% |
| Power/Infra (PWR, FIX, VST) | 30-43% | 47-67% |
| Semis (AVGO, TSM, NVDA) | 34-51% | 39-59% |
| Miners (any) | 80-110% | 18-25% |
6.3 Current Opportunity Assessment (March 2026)
Best buy opportunities right now (quality name + meaningful pullback):
- AVGO (-19.8% DD) -- Top Sharpe, bear-positive, pullback into the 15-20% sweet spot.
- APH (-18.2% DD) -- Lowest vol, strong Sharpe, rarely this cheap relative to peak.
- VST (-24.8% DD) -- Power/infra with bear resilience, deeper pullback than most quality names.
- GLW (-23.0% DD) -- Moderate vol, decent Sharpe, AI-driven fiber/glass demand.
- MU (-15.4% DD) -- Memory cycle play, strong 3M momentum (+53% excess), moderate pullback.
Wait for deeper pullback:
- FIX (-6.5%), PWR (-0.8%), VRT (0.0%), TSM (-3.4%) -- These are near highs. Quality names but no entry discount currently.
Requires special conviction (high risk):
- ALAB (-51.4%) -- If you believe in the networking chip thesis, this is a deep discount on a young name.
- LEU (-55.1%) -- Nuclear/uranium idiosyncratic play, not correlated with AI infra.
6.4 Proposed Monitoring Signal
For ongoing watchlist management:
- SOXX relative strength vs QQQ (20-day rolling) -- measures AI infra sector momentum vs. broad tech
- Miner-to-semi ratio (equal-weight WULF+CLSK+CIFR vs. NVDA+AVGO+TSM) -- measures speculative vs. quality AI infra sentiment
- VIX regime filter -- when VIX >25, restrict to Tier 1 names only; when VIX <20, Tier 2 becomes acceptable
- Drawdown monitoring -- alert when Tier 1 names enter 15-25% DD zone (buy opportunity) or >30% DD zone (thesis review needed)
7. Limitations
-
Short history for several names. ALAB (2 years), SNDK (1 year), MNRS (1 year), WGMI (4 years) have insufficient data for statistically robust conclusions. SNDK's 3.05 Sharpe is almost certainly not sustainable -- it reflects a short momentum run.
-
Survivorship bias. We only analyze names that currently exist on the watchlist. Failed AI infra companies that delisted are not represented, creating upward bias in return statistics.
-
Regime sample size. The bear regime sample contains only ~290 days concentrated in 2022. This is one event (Fed hiking cycle), not multiple independent bear markets. Bear-regime Sharpe estimates have wide confidence intervals.
-
Miner dual exposure. WULF, CLSK, CIFR, IREN are simultaneously exposed to Bitcoin price and AI compute demand. Their behavior during the overlap period is dominated by Bitcoin cycles. A pure AI compute demand assessment is not possible from price data alone.
-
AAOI misclassification. The watchlist classifies AAOI as an ETF but it is a single stock (Applied Optoelectronics). Its extreme volatility and drawdown profile are consistent with a micro-cap single name, not an ETF.
-
Correlation instability. Return correlations computed over multi-year windows may not reflect current regime correlations. The AI infrastructure theme is relatively new (2023 onward), and correlations during the AI buildout phase may differ from the broader 2018-2026 sample.
-
No fundamental analysis. This research examines price behavior only. Whether ALAB's -51% drawdown represents fundamental deterioration or a buying opportunity requires company-specific analysis outside this scope.
-
Composite score subjectivity. The pullback opportunity scorecard uses hand-tuned weights (25% Sharpe, 20% DD depth, etc.). Different weights would produce different rankings. The score is a starting point, not a definitive ranking.
-
Liquidity proxy limitations. The high-low spread proxy overstates actual bid-ask spreads. True execution costs require Level 2 data or broker fill analysis.
-
No tax or fee considerations. The stop-loss frequency analysis does not account for tax consequences of frequent stop-outs, which materially affect net returns for taxable accounts.