This study asks a single question: across 22 verticals in the LLM-inference supply chain, how much of the AI demand wave is already priced into the equities, and where is the market still behind? After a critique pass, the cleanest under-priced AI exposures are lithography (70% AI share, only +115% over 3y; ASML's 2024 DUV/China shock masked the tailwind) and datacenter REITs (45% AI share, +104% over 3y; rate-driven derating swamped surging AI bookings) — both verticals where the rally has lagged the underlying AI revenue share. The verticals most priced for perfection are silicon-photonics-optics (+1351% 3y, the cleanest AI rally on the board) and power-transformers-grid (+775% 3y on only 25% AI share, lifted heavily by the GEV spin and Siemens Energy turnaround). Big caveats: the model's ai-accelerators = lagging verdict is largely an artifact — its 85% AI-share input dominates a single −1.0 z-score coefficient, and a more honest 45–50% basket-weighted AI share flips NVDA's group to fair/priced-in; separately, the gap formula ignores the TAM-uplift term (a combined ~$1.27T runway if AI share climbs to 80% across the cohort), which would pull copper-rare-earth and industrial-gases-water into lagging despite their modest rallies. See WHY below for the per-vertical reasoning.
The chart below defaults to market-detrended view — what the supply-chain rallied after subtracting the S&P 500's own rally over the same window.
Lithography lagging is plausible. ASML and the small-cap Japanese names carry 70% AI share but the basket only returned +115% over 3y. The 2024 ASML earnings shock (DUV-to-China export controls, slower 2025 guidance) compressed the multiple at exactly the moment EUV-for-HBM and EUV-for-Blackwell were ramping. Yen-weak Japanese tickers also drag the equal-weighted basket; ASML alone returned +122%. The AI tailwind is real and not yet in the tape.
Datacenter-REITs lagging is plausible. EQIX/DLR carry ~45–60% AI-attributable bookings today but the 2022–23 rate-shock derating pulled the basket to +104% over 3y — below the cohort median. Long-duration REIT cash flows got hit by the discount rate even as AI hyperscaler leases were filling the 2025–27 pipeline. This is the classic "fundamentals ahead of multiple" setup.
ai-accelerators = lagging is largely a model artifact. NVDA alone returned +591% over 3y; the equal-weighted basket returned +376–421% — top-3 in the study. The label is forced by the formula's +2.20σ z_ai input (85% AI share) combined with the −1.0 coefficient on z_ai. Recompute the basket's AI share on a revenue-weighted basis (GOOGL's $350B+ revenue base swamps the average), and the share is ~31%, market-cap-weighted is ~60%. At a defensible 45–50%, the label flips to fair/priced-in. A +591% NVDA print is the AI repricing.
Silicon-photonics-optics is the cleanest AI rally in the dataset. +1351% 3y, gap of +3.66σ. LITE/COHR/FN/CIEN/AAOI rode the 800G pluggable cycle: LightCounting puts AI optics at $5B (2024) → $10B+ (2026), LITE cloud/AI is >60% of revenue climbing to 87% by 2027, 800G units 24M (2025) → 63M (2026). About 75%+ of the rally is AI-cloud back-end fabric.
Copper-rare-earth priced-in is mostly not AI. The 4% AI share is right. The rally is ~30% energy-transition / EV / electrification, ~25% supply disruption (Cobre Panama shutdown, Grasberg force majeure, Niger coup-adjacent supply), ~20% China stimulus + Trump Section 232 tariff arbitrage, ~15% rare-earth geopolitics (China export controls, MP-Pentagon deal). AI is ~10% — the label is technically correct but the reason the model cites (AI overpricing) is the wrong story.
The gap formula's tercile labels in the middle are mostly noise. 11 of 22 labels flip on a ±10pp move in a single estimated AI-share input. hyperscalers-cloud (lagging) and networking-switching (fair) differ by 0.13 in gap — categorically opposite labels separated by less than one rounding error. Trust the extremes, hover the middle.
Practical translation, no fabricated price targets: if AI infrastructure spend keeps grinding higher, the conceptually cleanest pair-trade frame from this dataset is long the genuinely-lagging AI exposures (lithography, datacenter-REITs, industrial-gases-water with its 8.9× TAM uplift) versus short the priced-for-perfection narrative trades (silicon-photonics-optics on the 800G cycle peak, power-transformers-grid where ~65% of the rally is non-AI structural). The middle of the ranking is watchlist material, not signal.
%%{init: {'themeVariables': {'fontSize': '18px', 'fontFamily': 'system-ui'}, 'flowchart': {'curve': 'basis', 'nodeSpacing': 40, 'rankSpacing': 50}}}%%
flowchart LR
classDef lag fill:#dff5e2,stroke:#1f7a36,color:#0a3b18,font-weight:600;
classDef pin fill:#fbe2e2,stroke:#a02323,color:#5b1414,font-weight:600;
classDef fair fill:#eeeeee,stroke:#555,color:#222,font-weight:600;
subgraph row_inputs["inputs"]
direction TB
copper_rare_earth["Copper & Rare Earths"]
industrial_gases_water["Industrial Gases & Water"]
end
subgraph row_ip-tools["ip-tools"]
direction TB
eda_ip["EDA & Silicon IP"]
lithography["Lithography"]
wfe_deposition_etch["WFE: Deposition, Etch, Impl…"]
end
subgraph row_silicon["silicon"]
direction TB
foundry_logic["Foundry — Logic"]
hbm_dram["HBM & DRAM"]
ic_substrates["IC Substrates"]
end
subgraph row_packaging["packaging"]
direction TB
advanced_packaging["Advanced Packaging"]
ai_accelerators["AI Accelerators"]
end
subgraph row_interconnect["interconnect"]
direction TB
silicon_photonics_optics["Silicon Photonics & Datacom…"]
networking_switching["Networking — Switching, Ret…"]
power_semis_vrm["Power Semiconductors — VRM …"]
end
subgraph row_facility["facility"]
direction TB
datacenter_cooling_thermal["Datacenter Cooling — Therma…"]
datacenter_reits["Datacenter REITs"]
electrical_equipment["Electrical Equipment"]
end
subgraph row_power["power"]
direction TB
power_transformers_grid["Power Transformers & Grid"]
gas_turbines["Gas Turbines"]
nuclear_smr_uranium["Nuclear — SMR & Uranium"]
utilities_merchant_power["Utilities & Merchant Power"]
end
subgraph row_consumers["consumers"]
direction TB
hyperscalers_cloud["Hyperscalers & Cloud"]
model_labs_software["Inference-Consuming Softwar…"]
end
copper_rare_earth --> eda_ip
eda_ip --> foundry_logic
foundry_logic --> advanced_packaging
advanced_packaging --> silicon_photonics_optics
silicon_photonics_optics --> datacenter_cooling_thermal
datacenter_cooling_thermal --> power_transformers_grid
power_transformers_grid --> hyperscalers_cloud
class lithography,datacenter_reits,ai_accelerators,foundry_logic,wfe_deposition_etch,industrial_gases_water,datacenter_cooling_thermal,hyperscalers_cloud lag;
class networking_switching,ic_substrates,gas_turbines,eda_ip,model_labs_software,power_semis_vrm,advanced_packaging fair;
class copper_rare_earth,utilities_merchant_power,electrical_equipment,hbm_dram,power_transformers_grid,nuclear_smr_uranium,silicon_photonics_optics pin;We have 22 equal-weight vertical indices indexed to 100 at 2021-05-31. Some rose spectacularly (silicon-photonics-optics +1351% 3y); some are flat. The honest question for an AI supply-chain study is: how much of each move is just "the market went up" versus AI-specific demand?
This note picks a detrending method, a benchmark, and writes the worked example.
The companion script compute_detrended.py produces returns_vertical_detrended.csv.
| # | Name | Formula | Pros | Cons |
|---|---|---|---|---|
| 1 | Arithmetic excess return | r_excess(t) = r_v(t) − r_m(t) (both cumulative from base) |
Easy to read in pp. | Asymmetric in compounding; ignores beta; large r_v makes pp gap deceptive. |
| 2 | Level ratio (multiplicative excess) | L_d(t) = 100 · L_v(t) / L_m(t) (both indexed to 100 at vertical's baseline) |
Equivalent to a dollar-neutral long-v / short-m pair-trade. Symmetric in log space. No estimation. Chart remains read as "cumulative pair-trade return". | Implicit β = 1 assumption; doesn't strip systematic risk for high-/low-β verticals. |
| 3 | Beta-adjusted (CAPM α-cumulation) | Estimate β over 36-month rolling daily log returns; α_cum(t) = log(L_v(t)) − β·log(L_m(t)) |
Strips market-risk loading honestly. | Requires β estimation (window choice, regime breaks); short-history verticals fail; β itself shifted post-2023 (AI verticals' β to S&P ↑ because they ARE the index now); explanation cost high; chart no longer reads as a tradeable return. |
| 4 | Non-AI benchmark (method 2 with synthetic basket) | Same as #2 but L_m = equal-weight of XLP/XLV/XLU/XLY (consumer staples, healthcare, utilities, consumer disc.) — i.e., S&P sectors with minimal AI exposure. |
Avoids "benchmark contaminated by AI" critique. | Synthetic basket has different fundamental drivers (rates, defensive flows); ETF baskets only go back so far; introduces sector-rotation noise that isn't really "market trend". |
The honest options:
Pick: ^GSPC with disclosed caveat. The S&P is the asset 99% of readers internalize as "the market". The bias from Mag7 share growth is real but small relative to the >1000% moves we're explaining — the lithography vertical isn't +115% over 3y because Mag7 dragged the S&P up; it's because of EUV/wafer demand that is downstream-but-distinct from the Mag7 capex line. We acknowledge the bias in the chart caption and move on.
Rationale:
1. Readability: detrended_level / 100 − 1 reads as "cumulative return of a
long-vertical / short-S&P, equal-notional, monthly-rebalanced pair trade
from 2021-05-31". That is a thing a portfolio manager would actually run; it
has dollar meaning.
2. No estimation risk: no β window, no regime issues, no degrees of freedom.
3. Same axis as raw chart: the existing chart Y-axis is "% return from
baseline". Toggling detrending keeps the same axis — just swaps one series
for another. Method 3 would need a log-axis or "α" axis, breaking UX.
4. Symmetric in log space: a vertical that doubled while S&P doubled shows
exactly 0% detrended, which matches intuition. Method 1's arithmetic excess
would show "+100 pp − 100 pp = 0 pp" coincidentally correctly here but would
misbehave when both moves are large.
5. Method 3 is the right answer for a true alpha study, but this is a
demand-side / theme-attribution study, not a portfolio-construction study.
We choose readability + honesty about the limitation.
"Cumulative return of a dollar-neutral pair trade: long the equal-weight vertical basket, short the S&P 500, both rebalanced monthly, from 2021-05-31. A line at +0% means the vertical exactly tracked the market; +100% means it doubled relative to the market."
Vertical baseline = 2021-05-31. S&P 500 adj_close at 2021-05-28 (snapped month-end) ≈ 4204.11.
| date | L_v (raw) | GSPC | L_m (=100·GSPC/GSPC_base) | L_d = 100·L_v/L_m | detrended r |
|---|---|---|---|---|---|
| 2021-05-31 | 100.00 | 4204.11 | 100.00 | 100.00 | 0.0% |
| 2021-11-30 | 117.14 | 4567.00 | 108.63 | 107.83 | +7.8% |
| 2022-05-31 | 116.22 | 4132.15 | 98.29 | 118.24 | +18.2% |
| 2022-11-30 | 109.52 | 4080.11 | 97.05 | 112.85 | +12.8% |
| 2023-05-31 | 130.57 | 4179.83 | 99.42 | 131.33 | +31.3% |
| 2023-11-30 | 157.96 | 4567.80 | 108.65 | 145.38 | +45.4% |
| 2024-05-31 | 206.62 | 5277.51 | 125.53 | 164.59 | +64.6% |
| 2024-11-30 | 155.53 | 6032.38 | 143.49 | 108.39 | +8.4% |
| 2025-05-31 | 152.01 | 5911.69 | 140.62 | 108.11 | +8.1% |
| 2025-11-30 | 199.89 | 6849.09 | 162.91 | 122.69 | +22.7% |
| 2026-05-31 | 281.22 | 7473.47 | 177.77 | 158.20 | +58.2% |
Numbers produced by compute_detrended.py.
Raw 5y lithography return: +181.2%. Detrended 5y: +58.2%. Roughly 2/3 of the vertical's headline rally was "the market went up", 1/3 was lithography-specific. The 3y detrended (2023-05 to 2026-05) is only +20.5% — most of the lithography outperformance happened in the 2021-23 window, then collapsed to roughly market in 2024-2025 before a 2026 reacceleration. The priced-in story holds: EUV demand IS real, but the public-market multiple expansion did most of the work.
When the "Detrend" toggle is ON, each vertical's line shows what you'd have made running a long-vertical / short-S&P pair trade from May 2021, with the S&P benchmark serving as a stand-in for "what generic equity exposure delivered". A flat line means the vertical did exactly what the market did; an upward-sloping line means the vertical outperformed; a downward-sloping line means it underperformed.
| Benchmark | 1y | 3y | 5y (since 2021-05) |
|---|---|---|---|
| ^GSPC | +26.4% | +78.8% | +77.8% |
| ^NDX | +38.1% | +106.8% | +115.4% |
| XLK | +57.1% | +124.2% | +170.8% |
| SOXX | +163.9% | +245.3% | +288.0% |
| NVDA | +59.4% | +469.7% | +1228.9% |
This note describes what the broad market did during the window of our LLM-inference supply-chain study, so that readers know what the "subtract market trend" toggle in the chart is actually netting out. All numbers are computed live from the price CSVs in data/prices/ against a 2021-05-31 baseline.
2021-05 → end of 2021: the tail of a speculative bull. The window opens in May 2021, near the end of the post-COVID liquidity boom. Meme stocks, SPACs, unprofitable growth, and crypto were all close to their highs. The S&P 500 kept grinding up and made an all-time high on 2022-01-03 at ~4,797. Tech and semis were stretched on valuation, and the Fed was still on emergency settings.
2022: the rate-hike bear market. As the Fed began the fastest hiking cycle in 40 years, long-duration assets sold off hard. From the 2021-05-31 baseline, the S&P 500 was at -1.7% by 2022-05 and bottomed at -14.7% on 2022-09-30 (month-end basis). The Nasdaq-100 was deeper: -16.7% at the October 2022 mark. Semis fell further still — SOXX was at -23.4% by Oct 2022.
Oct–Nov 2022 — ChatGPT launches near the trough. The intra-window closing low for the S&P 500 was 2022-10-12 (3,577), and on a month-end basis the trough lands at 2022-09-30 (-14.7% vs baseline). ChatGPT was released publicly on 2022-11-30, about six weeks after that low. The launch did not single-handedly trigger the rally — rate-hike expectations were already peaking — but it created the narrative engine that absorbed every subsequent dollar of dovish surprise.
2023 → 2024: the Mag7-led AI rally. From the trough, large-cap tech led a narrow but ferocious rally. Nvidia's Q1-FY24 datacenter print in May 2023 was the inflection. By 2024-05 NVDA was at +576.2% vs 2021-05; SOXX at +66.8%; the Nasdaq-100 at +35.4%. The S&P 500 recovered to baseline around 2023-06-30 and kept going. Breadth was thin: most of the index gain came from a handful of AI-adjacent names.
2024-25: rate cuts begin, breadth widens, capex super-cycle. The Fed pivoted; the rally broadened beyond Mag7. Hyperscaler AI capex guidance crossed $300B/yr aggregate. Nvidia's revenue ran north of $130B. Power, gas turbines, transformers, and rare-earth magnets started showing up as binding constraints — themes this study is built around.
2025-26: the AI buildout matures. By 2026-05-31 the S&P 500 sat at a 5-year total return of +77.8%, the Nasdaq-100 at +115.4%, XLK at +170.8%, SOXX at +288.0%, and NVDA at +1228.9%.
| Benchmark | 1y | 3y | 5y (since 2021-05-31) |
|---|---|---|---|
| ^GSPC | +26.4% | +78.8% | +77.8% |
| ^NDX | +38.1% | +106.8% | +115.4% |
| XLK | +57.1% | +124.2% | +170.8% |
| SOXX | +163.9% | +245.3% | +288.0% |
| NVDA | +59.4% | +469.7% | +1228.9% |
S&P 500 in-window trough: 2022-09-30 at -14.7% vs baseline. Recovery to baseline: 2023-06-30.
Almost everything in the supply chain we cover is, to some degree, a tech or growth stock. When the Nasdaq-100 rises +115.4% over five years, a name that did +115.4% did not outperform anything — it just rode market beta. The "subtract market trend" toggle in the HTML divides each ticker's return path by the benchmark's return path, leaving only the excess return — the part not explained by being long the broad market. Without that step, a chart of winners reads more like a chart of beta than a chart of AI exposure.
By 2025-26 the S&P 500 is itself heavily AI-weighted. The Magnificent Seven (NVDA, MSFT, AAPL, GOOGL, AMZN, META, TSLA) collectively account for roughly 30% of the index's market cap, and several of them are the largest customers and suppliers in the AI buildout. Detrending against the S&P partly nets AI out of itself. A purer benchmark would be the equal-weight S&P 500 (RSP) or a non-tech basket — both of which underperformed the cap-weighted index materially over this window. Treat the detrended view as conservative attribution: it understates the true AI premium, because the thing we are subtracting already contains the trade.
gap (most lagging first); click any column to re-sort| vertical | 3y total return | beta NVDA | AI share % | revenue 2025 $bn | AI rev today $bn | AI rev @ 80% $bn | uplift $bn | uplift × | label |
|---|---|---|---|---|---|---|---|---|---|
| Lithography | 115.3% | 0.17 | 70.0% | 28.4 | 19.9 | 22.7 | 2.8 | 1.14× | lagging |
| Datacenter REITs (Colocation + Wholesale) | 103.7% | 0.18 | 45.0% | 119.0 | 53.5 | 95.2 | 41.6 | 1.78× | lagging |
| AI Accelerators (GPUs/ASICs/TPUs) | 376.4% | 0.53 | 85.0% | 200.0 | 170.0 | 160.0 | -10.0 | 0.94× | lagging |
| Foundry — Logic | 254.2% | 0.37 | 58.0% | 169.5 | 98.3 | 135.6 | 37.3 | 1.38× | lagging |
| WFE: Deposition, Etch, Implant, Metrology | 145.5% | 0.49 | 55.0% | 115.7 | 63.6 | 92.6 | 28.9 | 1.45× | lagging |
| Industrial Gases & Water (fab inputs + DC cooling/humidification) | 22.6% | 0.03 | 9.0% | 120.0 | 10.8 | 96.0 | 85.2 | 8.89× | lagging |
| Datacenter Cooling — Thermal Management | 259.7% | 0.47 | 55.0% | 11.5 | 6.3 | 9.2 | 2.9 | 1.45× | lagging |
| Hyperscalers & Cloud | 107.1% | 0.36 | 35.0% | 419.0 | 146.7 | 335.2 | 188.6 | 2.29× | lagging |
| Networking — Switching, Retimers, DPUs | 225.7% | 0.57 | 55.0% | 33.0 | 18.1 | 26.4 | 8.2 | 1.45× | fair |
| IC Substrates (ABF / FC-BGA / BT) | 423.5% | 0.09 | 35.0% | 20.0 | 7.0 | 16.0 | 9.0 | 2.29× | fair |
| Gas Turbines | 550.3% | 0.26 | 55.0% | 30.2 | 16.6 | 24.2 | 7.5 | 1.45× | fair |
| EDA & Silicon IP | 185.4% | 0.52 | 45.0% | 21.2 | 9.5 | 17.0 | 7.4 | 1.78× | fair |
| Inference-Consuming Software / App Layer | 265.5% | 0.39 | 35.0% | 170.0 | 59.5 | 136.0 | 76.5 | 2.29× | fair |
| Power Semiconductors — VRM / Vertical Power Delivery | 120.7% | 0.40 | 22.0% | 73.7 | 16.2 | 59.0 | 42.7 | 3.64× | fair |
| Advanced Packaging (OSAT, substrates, FOPLP, backend test) | 412.0% | 0.29 | 35.0% | 50.0 | 17.5 | 40.0 | 22.5 | 2.29× | fair |
| Copper & Rare Earths | 121.7% | 0.24 | 4.0% | 280.0 | 11.2 | 224.0 | 212.8 | 20.00× | priced_in |
| Utilities & Merchant Power | 234.2% | 0.20 | 8.0% | 420.0 | 33.6 | 336.0 | 302.4 | 10.00× | priced_in |
| Electrical Equipment (Datacenter Power Distribution) | 375.8% | 0.29 | 18.0% | 24.0 | 4.3 | 19.2 | 14.9 | 4.44× | priced_in |
| HBM & DRAM | 631.1% | 0.21 | 30.0% | 38.0 | 11.4 | 30.4 | 19.0 | 2.67× | priced_in |
| Power Transformers & Grid | 775.1% | 0.28 | 25.0% | 28.0 | 7.0 | 22.4 | 15.4 | 3.20× | priced_in |
| Nuclear — SMR & Uranium | 418.3% | 0.47 | 8.0% | 200.0 | 16.0 | 160.0 | 144.0 | 10.00× | priced_in |
| Silicon Photonics & Datacom Optics | 1351.1% | 0.64 | 55.0% | 31.5 | 17.3 | 25.2 | 7.9 | 1.45× | priced_in |
ai-accelerators really lagging?The ranking model places ai-accelerators in the bottom tercile (gap = −1.485). Common sense pushes back: NVDA is the AI bellwether. This note stress-tests each input.
The vertical is an equal-weighted month-end index of 7 names (see _notes_returns.md lines 48–62). All seven names had strong 3y returns when computed from the underlying price files, anchored at 2023-05-22:
| ticker | 3y total return | 5y total return |
|---|---|---|
| NVDA | 591.3% | 1282.7% |
| 2454.TW (MediaTek) | 645.8% | 537.7% |
| AVGO | 535.3% | 896.9% |
| AMD | 332.9% | 503.7% |
| MRVL | 327.0% | 320.4% |
| INTC | 303.8% | 129.9% |
| GOOGL | 208.8% | 227.1% |
(Computed live from data/prices/*.csv.)
Critical finding: the premise that "INTC and MediaTek dragged it down" is false. MediaTek (+646%) actually beat NVDA on 3y, and INTC tripled. The real drag is GOOGL (+209%), because GOOGL is a mega-cap whose denominator is Alphabet's entire enterprise value, not its TPU silicon.
| Method | 3y total return |
|---|---|
| NVDA only | 591% |
| Equal-weighted 7-ticker basket (raw prices) | 421% |
| Equal-weighted 7-ticker basket (vertical CSV, 2023-05-31→2026-05-31 month-end) | 376% |
| Market-cap-weighted basket (rough caps) | 483% |
| Pure-merchant subset NVDA/AMD/AVGO/MRVL (cap-wt) | 568% |
Most fair interpretation: the equal-weighted basket return (376–421%) is not dragged down by losers — every constituent more than 3x'd. It just happens to be lower than NVDA-solo because no other name matched NVDA's 6.9x. But it is still the 3rd highest 3y return of any vertical in the study (only silicon-photonics-optics at 1351% and power-transformers-grid at 775% beat it). So the basket is not weak; the basket is loud — just not as loud as silicon-photonics' 13.5x rip.
The vertical JSON justifies 85% by NVDA-weighting (NVDA ≈70% of vertical revenue, and 88% of NVDA is data-center). That is internally coherent if you weight by AI-accelerator revenue. But the model treats the basket's z_ret as an equal-weighted equity index — so the symmetric input would be an equal-weighted AI share of the 7 names.
My rough revenue-weighted recompute (NVDA 88% AI / AVGO 35% / GOOGL 15% / AMD 25% / MRVL 50% / INTC 5% / MediaTek 5%, weighted by FY25 revenue with GOOGL ≈$350B dominating):
The 85% number is honest if you read the vertical as "AI-accelerator dollars sold" and ignore GOOGL/MediaTek/INTC's non-accelerator revenue. It's overstated if you want symmetry with the equity-return basket. There is no clean answer; both views are defensible.
The cross-vertical mean of ai_share_today_pct is ≈25.7% with sd ≈16.9. So:
Re-running the gap with corrected AI share (everything else equal):
| ai_share | gap | rank | label |
|---|---|---|---|
| 85% (current) | −1.485 | 3/22 | lagging |
| 70% | −0.911 | 5/22 | lagging |
| 60% | −0.469 | 9/22 | fair |
| 50% | +0.011 | 13/22 | fair |
| 45% | +0.261 | 15/22 | priced_in |
The label flips on a single subjective input. That is the most damning finding. The gap formula in this model has one binary input (which AI-share definition you adopt) that swings the verdict by 1.65σ.
Replacing total_return_3y with total_return_5y from cagr.csv:
Even on a 5y horizon, the +85% ai_share input dominates, so the label doesn't change.
| weights | gap | rank | label |
|---|---|---|---|
| (1.0, 0.5, −1.0) baseline | −1.485 | 3/22 | lagging |
| (0.5, 0.5, −1.0) | −1.549 | 2/22 | lagging |
| (1.0, 0.0, −1.0) | −2.076 | 2/22 | lagging |
| (1.0, 1.0, −0.5) | +0.208 | 15/22 | priced_in |
The label is invariant to discounting return or removing beta, but flips to priced_in when you halve the AI-share penalty and double the beta reward. The −1.0 coefficient on z_ai is doing essentially all the work.
Why ai-accelerators COULD still be lagging (genuine bull case): - Custom-ASIC outsourcing is just starting. Google TPU v8, Meta MTIA v2, AWS Trainium3, Microsoft Maia 2 all ramp 2026–2028 — AVGO and MRVL revenue from those is largely unbooked. - NVDA Blackwell B200/B300 gross margins (70%+) are higher than priced; consensus has them compressing. - AI infrastructure spend trajectory ($300B → $500B+ annual capex) is barely 30% complete; the durable winners (NVDA/AVGO) take fixed revenue share through 2030.
Why ai-accelerators is likely already priced in (genuine bear case): - NVDA at $5.2T market cap is ~6% of S&P. The "AI boom" is largely defined as NVDA's run. A +591% 3y return on the world's most-followed stock is the textbook definition of fully discounted. - The 85% ai_share is the entire thesis the market spent 2023–25 pricing in. It would be circular to label a vertical "underpriced because AI" when "AI" is precisely why it ran 6x. - Lithography, datacenter-reits, and gas-turbines are more genuinely "AI exposure not yet priced" stories — they have similar AI exposure but flat returns. ai-accelerators looks "lagging" only by an artifact of its outlier z_ai.
Within the model as specified, ai-accelerators is lagging because its z_ai (+2.20σ) is the largest in the universe and a single −1.0 coefficient on that z-score mechanically forces a −1.5σ gap regardless of what the price did. The 591% NVDA-solo return and 376% basket return are not small — they are top-3 in the study — but no realistic equity return can offset a +2σ AI input in this formula.
Outside the model, a reasonable observer would more likely call it priced-in — or at best "fair." The 376–591% return range is the AI repricing. Halving the AI-share input to a defensible 45–50% (revenue-weighted across the equal-weighted basket) flips the label to fair/priced_in, and that re-anchoring is more honest than the current 85%.
The current ranking labels seven verticals as priced-in or rallied with positive gaps to the AI exposure score. The implicit narrative: AI demand is what drove these stocks. But AI share-of-revenue is 4-30% in each, so most of the rally must logically come from something else. This note interrogates that claim and attempts an attribution split per vertical with cited sources.
The vertical's 3y total return of 1.22 looks unimpressive in absolute terms — it is below the ai-accelerators basket (3.76) — but the model flags it as priced-in because the AI exposure score is so low. So the question is narrower: did the modest rally that did occur come from AI, or somewhere else?
Alternative drivers: - Supply disruption stack: First Quantum's Cobre Panama shutdown Nov 2023 (~1.5% of global supply removed); 2025 accidents at Ivanhoe's Kamoa-Kakula (DRC) and Freeport's Grasberg (Indonesia, force majeure declared, 2026 guidance slashed). INN 2025 Year-End Review - Capex underinvestment 2015-2020: Global ore grades fell from 1-2% to <0.7%; project capex was for maintenance, not new capacity. ICSG projects 150kt deficit by 2026 driven by inability to deliver supply, not demand volatility. Crux Investor - China stimulus 2024-25: Beijing's "more proactive" fiscal policy + "moderately loose" monetary policy for 2026, with grid/renewable/data-center spending all copper-intensive. INN - Trump Section 232 tariff arbitrage (Jul 2025): 50% tariff on semi-finished copper; pre-announcement, COMEX-LME spread blew out to $2,637/t, premium then collapsed when refined cathode was excluded. Big speculative noise. White & Case - Energy transition baseline: IEA expects clean-energy techs to lift refined copper use to 33 Mt by 2035, 37 Mt by 2050 (vs ~27 Mt in 2024). EVs + grid hardening + heat pumps dwarf AI-DC tonnage. - Rare earth catalysts unrelated to AI: China export controls Apr 2025 + Oct 2025 expansion (paused Nov 2025-Nov 2026); MP Materials Pentagon $400M preferred + $110/kg NdPr floor (Jul 2025) — a price-floor backstop more than an AI thesis. CNBC, CSIS
Best-guess attribution (judgment): 30% energy transition / EV / electrification, 25% supply disruption + capex underinvestment, 20% China stimulus + tariff arbitrage, 15% rare-earth-specific geopolitics (DoD, China controls), 10% AI/DC. AI is not the load-bearing narrative here; it sits atop a deeper structural copper deficit story.
Alternative drivers: - PJM capacity auction repricing: Clearing prices jumped from $28.92/MW-day (2024/25) to $269.92 (2025/26) to $329.17 (2026/27) — a 10x repricing in two auctions. PJM's own forecast attributes 94-97% of the 32 GW load growth 2024-2030 to data centers. Utility Dive, IEEFA - Coal retirement cliff: ~6 GW PJM fossil already retired pre-2024 auction; 15 GW more coal planned by 2029. Tightening reserve margins drive auction clears even without AI. - ERCOT heat-wave seasonality: 2024 set 85,559 MW peak record Aug 20; 2025 spring saw 1,600% intraday spikes. VST/NRG unhedged capacity can earn $2-10M/hour at peak. Fortune - Microsoft-CEG TMI deal (Sep 2024): drove CEG +25% in one day, sparked re-rating across nuclear IPPs. AI-PPA is the single most cited catalyst. CNBC
Best-guess attribution: 55% AI/DC capacity scarcity (PJM data shows this is overwhelmingly the marginal load), 20% coal-retirement-driven scarcity, 15% weather/seasonal merchant cash flows, 10% other (rate-base capex). AI is the cleanest explanation here — PJM's own forecasting attributes 94-97% of load growth to data centers, which is much higher than the 8% revenue-share number suggests because the marginal MWh prices the whole stack.
Alternative drivers: - Megaproject + reshoring backlog: ETN Q4'25 disclosed 54% YTD megaproject announcements were data centers, the rest US reshoring. $3T NA megaproject backlog. Eaton mega-project revenue +30% 2024→2025. Eaton Q4 2025 transcript - IRA + CHIPS Act: ongoing fab/factory build (TSMC AZ, Samsung TX, Intel OH, GF NY) demands MV switchgear and dry transformers — no AI exposure but same vendor list. - Grid hardening post-Uri (Texas Feb 2021): state PUCs raised resilience capex; T&D contractors (PWR, MYRG, PRIM) saw multi-year backlog growth predating AI hyperscaler narrative. - Electrification baseline: EV charging, heat-pump conversions, building electrification.
Best-guess attribution: ~50% data-center demand (most explicit company disclosure), 25% non-DC reshoring/megaprojects, 15% grid hardening + utility-side capex, 10% electrification baseline. AI is real, but ~54% of Eaton's megaproject sources are explicitly DC — that includes both AI-training and conventional cloud/colo, so AI-attributable is closer to 35-40%, not the 18% in the JSON. The JSON likely under-counts AI here.
Alternative drivers (memory bear case): - Classic DRAM cycle recovery: 2022-23 trough was a deep oversupply crash. 2024-25 was always going to recover on inventory bleed and capex discipline — the cycle is endogenous. - Wafer reallocation, not demand: Up to 40% of advanced wafer capacity has been redirected from commodity DRAM to HBM, which shrinks commodity DRAM supply 3-4x per HBM chip equivalent. The price spike on commodity DDR5 is a supply-side artefact, not new demand. TrendForce - Cyclical caution by suppliers: Samsung/SK Hynix are choosing not to expand aggressively, sustaining prices regardless of AI.
Counter-evidence (AI bull case): - HBM TAM going from $17B (2024) to $36-38B (2025) to ~$58B (2026) — pure AI demand. TrendForce - 2026 HBM supply fully booked. - The wafer-reallocation mechanism itself is caused by AI demand pulling capacity into HBM, so it isn't really separable.
Best-guess attribution: 60% AI-HBM demand (the pull source), 25% supply discipline / wafer-mix repricing (which is downstream of AI anyway), 15% normal cyclical recovery from the 2022-23 trough. AI is the dominant driver, full stop. The 30% AI share number in the JSON understates this because it's measured by HBM revenue/total DRAM revenue — but HBM scarcity sets the marginal price for the entire DRAM stack.
Alternative drivers: - Aging fleet: 40-year average asset life, 70% of US transmission lines nearing end of 50-80 year lifecycle. Replacement cycle would be tight regardless of AI. Industrial Sage - Interconnection queue: ~2,300 GW renewables + storage in queue end of 2024; every project needs LPTs. Lead times moved from 30-60 weeks pre-pandemic to 4 years. Demand for GSU transformers +274% 2019-2025 — pre-dates the AI capex bubble. - Renewables build (IRA-driven 2022-onward) and EV grid upgrades. - GEV spin-off (Apr 2, 2024) re-rating: GEV stock has quadrupled since spin. Investor base rotation + cleaner pure-play disclosure mechanically drove multiple expansion separate from AI. GE Vernova press - Siemens Energy turnaround: ENR.DE +300%+ from 2024 lows largely driven by resolution of the Gamesa wind quality crisis (€1B+ provisions in 2023) plus grid order book. Without the wind crisis lift-off, baseline would be lower. TS2.tech - Korean exporters (Hyundai Electric, HD): weak KRW + tariff arbitrage + US plant ramps. Stock +900%+ off lows over 3 years, mostly from US grid orders for both AI and non-AI utility customers.
Best-guess attribution: 35% AI/DC (acknowledged by all OEMs), 25% renewables interconnection + IRA-driven grid build, 20% replacement / aging-fleet cycle, 10% GEV spin re-rating + Siemens Gamesa-turnaround idiosyncratic mechanical lift, 10% reshoring industrial loads. AI is real but not majority — the transformer shortage was visibly building 2019-2023 before AI capex spiked, and IEA/queue data show renewables + EVs are larger MW additions than DCs even today.
Alternative drivers: - Russia uranium import ban (Prohibiting Russian Uranium Imports Act, May 2024): forced Western utilities to buy alternatives; Russia historically supplied ~35% of US enriched uranium. This is the single largest non-AI catalyst for U3O8 spot. Cameco market commentary cited in Oregon Group - Niger coup July 2023: removed ~5% of global uranium supply, drove spot from $50/lb to >$100/lb Jan 2024 — well before any hyperscaler nuclear PPA. Bloomberg - Cameco supply cuts: 2023 production guidance cuts at Cigar Lake (18→16.3 Mlb) and McArthur River (15→14 Mlb). - COP28 triple-nuclear pledge (Dec 2023): 22-31 countries committed to triple nuclear capacity by 2050. Japan restarts, France ramp, China builds — none of this is AI. WNA - CEG-Microsoft TMI deal (Sep 2024): this is the unambiguous AI catalyst that re-rated the IPP cohort.
Best-guess attribution: 35% uranium-supply shock (Russia ban + Niger + Cameco), 30% global nuclear renaissance pledges + utility re-contracting (non-AI), 25% AI hyperscaler PPAs (TMI, AWS-Susquehanna, Comanche Peak, Meta-Oklo, Google-Kairos), 10% SMR speculation / Sam Altman halo on OKLO">OKLO. AI is a real, important catalyst but not the dominant one — uranium spot rallied to $100+ in Jan 2024 nine months before MSFT-CEG, on supply factors alone.
This is the most extreme gap on the board (+3.66) and the 3y return of 13.5x dominates.
Alternative drivers (small): - Telecom 5G last-mile cycle: Lumentum telecom revenue was 54% of optical sales mid-2024 with 24% YoY growth from 5G ZR transceivers. But this is a secondary cycle and shrinking relative to cloud. - Coherent DCI / ZR pluggables (DC-to-DC interconnect): a real telecom-adjacent cycle but the buyers are still hyperscalers. - M&A repricing: COHR (II-VI acquisition), CIEN Nubis acquisition (Sep 2025), INFN-Nokia close (Feb 2025) — corporate actions distort comps but don't account for 13.5x.
Counter-evidence (AI bull case): - LITE cloud/AI >60% of revenue, climbing to 87% by 2027. - 800G transceiver units: 24M shipped in 2025, 63M projected 2026 (2.6x). LightCounting puts AI-specific optics at $5B (2024) → $10B+ (2026). - FN datacom $88M → $305M/qtr on AI 800G. - LITE +166% YTD, COHR +97% YTD.
Best-guess attribution: 75% AI-cloud back-end fabric demand (800G/1.6T), 15% optical transport + DCI hyperscaler buildout (still ultimately AI-driven), 5% telecom 5G + carrier, 5% M&A/idiosyncratic. This is the cleanest AI-driven rally on the entire list. The 55% AI share in the JSON is probably understated; the marginal revenue is overwhelmingly AI.
This is an adversarial review of the gap = z(3y_total_return) + 0.5·z(beta_NVDA) − z(ai_share_today_pct) ranking and the tercile labels (lagging / fair / priced_in) it produces. All numerical claims are recomputed from analysis/cagr.csv, analysis/risk.csv, analysis/tam.csv, and the 22 data/verticals/*.json files (see analysis/_critique_compute.py).
The metric is, by construction, return minus AI-share. Spearman rank-correlations of gap against its inputs:
gap vs. total_return_3y: +0.61gap vs. ai_share_today_pct: −0.63gap vs. beta_nvda: +0.11Beta is decorative. The "priced-in" score is roughly return rank minus AI-share rank. That mechanically penalises verticals whose stocks have rallied because they are the AI demand source. ai-accelerators (3y total return 376%, 85% AI share, label lagging) and lithography (115% return, 70% AI share, label lagging) get flagged as under-priced relative to their AI exposure — but those are exactly the verticals where AI demand was first capitalised. Calling NVDA-and-ASML "lagging" while flagging silicon-photonics-optics (1351% return, 55% AI share) as "priced in" is consistent with the formula but the semantic label is misleading. A more defensible reading: lagging means "share-of-AI-revenue has not yet shown up in the multiple" — but for verticals whose entire valuation is AI demand, this case is structurally impossible to detect with cross-sectional z-scores.
A reframed metric — "rally is on FUTURE expectation" — would label as "priced in" verticals with high return AND low AI share (e.g. power-transformers-grid, nuclear-smr-uranium, silicon-photonics-optics). The current formula already does this for those three, which is its main correct insight. But it incorrectly extends the same logic to ai-accelerators, calling them "lagging" because the formula has no way to distinguish "already monetised" from "yet to be monetised."
Tercile cuts: q1=−0.582, q2=+0.266. Bordering verticals are within ±0.05 of a cut but get categorically different labels:
| vertical | gap | label | distance to cut |
|---|---|---|---|
| industrial-gases-water | −0.695 | lagging | −0.11 below q1 |
| datacenter-cooling-thermal | −0.640 | lagging | −0.06 below q1 |
| hyperscalers-cloud | −0.582 | lagging | 0.00 exactly on q1 |
| networking-switching | −0.449 | fair | +0.13 above q1 |
| power-semis-vrm | +0.196 | fair | −0.07 below q2 |
| advanced-packaging | +0.266 | fair | exactly on q2 |
hyperscalers-cloud and networking-switching differ by 0.13 in gap but get categorically opposite labels. Recommend: drop the tercile labels for HTML rendering and use a continuous diverging colour scale anchored at gap=0 (e.g. red→white→green), with the numerical gap as the hover/tooltip value. Keep the three categorical bins only in _notes_ranking.md for narrative.
ai_share_today_pct is partly triangulated (advanced-packaging 35%, datacenter-cooling 55%, hyperscalers-cloud 35%, etc.). With weight −1.0·z(ai_share), a 10pp shift to a single vertical's share — well within the uncertainty band on a non-measured estimate — flips its label for 11 of 22 verticals (50% of the sample):
| vertical | shift | from | to |
|---|---|---|---|
| advanced-packaging | −10pp | fair | priced_in |
| copper-rare-earth | +10pp | priced_in | fair |
| datacenter-cooling-thermal | −10pp | lagging | fair |
| eda-ip | +10pp | fair | lagging |
| gas-turbines | +10pp | fair | lagging |
| hyperscalers-cloud | −10pp | lagging | fair |
| ic-substrates | +10pp | fair | lagging |
| industrial-gases-water | −10pp | lagging | fair |
| networking-switching | +10pp | fair | lagging |
| power-semis-vrm | −10pp | fair | priced_in |
| utilities-merchant-power | +10pp | priced_in | fair |
The verticals whose labels are not flipped by ±10pp on their own ai_share are the extremes (lithography, ai-accelerators, silicon-photonics-optics, hbm-dram, etc.). Every middle-tercile assignment is, in practice, an artefact of an estimated parameter. The label is more sensitive to the analyst's prior on AI revenue mix than to anything measured from prices.
Spearman correlation of gap vs. beta_nvda is only +0.11; the 0.5·z(beta) term shifts a few orderings but never changes a top/bottom-tercile assignment alone. The conceptual problem is worse: beta_nvda is partly mechanical — semiconductor-adjacent names will correlate with NVDA simply because they share factor exposure (cyclicals, growth, dollar-yen). Utilities and copper miners have low NVDA beta not because they lack AI exposure but because their fundamentals are driven by power demand cycles and commodity prices. Including beta tags those verticals as "less priced in" even when their rallies (utilities-merchant-power +234%, copper-rare-earth +122% over 3y) are AI-narrative-driven. Recommendation: drop the beta term entirely, or replace it with a "share of post-Nov-2022 return attributable to the AI factor" estimate from a two-factor regression (NVDA + market). Half a z-score of an ambiguous variable is not earning its keep.
Ticker counts per vertical range from 3 to 8. The thin baskets are statistically fragile:
eda-ip — 3 tickers (SNPS, CDNS, ARM">ARM). ARM">ARM IPO'd 2023-09-14, so the basket's 3y CAGR is effectively an SNPS-CDNS average plus a fragment.foundry-logic — 4 tickers (TSM, INTC, GFS, UMC), three of them in non-US listings; GFS itself IPO'd 2021-10 and just barely clears the 3y window.datacenter-reits — 6 tickers but four are foreign (GDS, VNET, AJBU.SI, plus two US). Currency moves are baked into local-currency returns; no FX adjustment.lithography — 6 tickers, 5 of which are small Japanese names (7735.T, 8035.T, 6920.T, 7731.T, 7751.T) whose adj_close in JPY has been hammered by yen weakness. The basket return (1.15× over 3y) is partly an FX artefact, not an AI-pricing signal. ASML alone returned 1.22×; the JPY tickers drag the equal-weight mean.Equal-weighting an English-listed mega-cap with a small Tokyo Section 1 name in a 3-ticker basket is not a defensible cross-sectional ranking input.
Re-running with total_return_5y:
| vertical | 3y label | 5y label |
|---|---|---|
| datacenter-cooling-thermal | lagging | fair |
| gas-turbines | fair | lagging |
Most labels are stable. The 5y window covers the pre-ChatGPT cool-off; gas-turbines look worse (the COVID-recovery base was richer), datacenter-cooling looks better (5y captures the VRT/MOD multi-year run). This suggests the methodology is not especially window-sensitive at the extremes, but the borderline cases are completely unstable under modest specification changes — same conclusion as §3.
Recomputing total return from 2022-10-31 (ChatGPT launch month-end) to today changes 6 of 22 labels:
| vertical | 3y-baseline label | ChatGPT-baseline label |
|---|---|---|
| datacenter-cooling-thermal | lagging | fair |
| eda-ip | fair | lagging |
| gas-turbines | fair | lagging |
| hyperscalers-cloud | lagging | fair |
| model-labs-software | fair | priced_in |
| utilities-merchant-power | priced_in | fair |
Two are interesting:
utilities-merchant-power drops out of priced-in. Its 3y total return is heavily weighted by the 2024 utilities rally; from 2022-10 onward the total return is only 182%, which is mid-pack. The "priced in" call is partly a window choice.model-labs-software becomes priced-in. APP, PLTR drove a 747% return from ChatGPT launch — clearly an AI-narrative trade. The 3y window dilutes it with 2022 drawdowns.There is genuine time-frame snooping risk here. The 2021-05 baseline encompasses the COVID/zero-rate peak, the 2022 rate-shock drawdown, then the AI rally. A baseline that starts at the event the user is studying (ChatGPT) is more honest. Recommend running both, presenting both, and flagging verticals whose labels disagree as "window-sensitive."
tam_uplift_multiple is in the data and nowhere in the gap. Yet it is the single most direct measure of how much further AI demand can push this vertical's revenue:
copper-rare-earth: 20.0× uplift if AI hits 80% shareutilities-merchant-power: 10.0× upliftnuclear-smr-uranium: 10.0× upliftindustrial-gases-water: 8.9× upliftelectrical-equipment: 4.4× upliftCompare to lithography (1.14×) or ai-accelerators (0.94× — already over the 80% threshold). A vertical with a tiny AI share TODAY but a huge multiple if AI demand expands is not "priced in"; it is structurally early. The current formula calls copper-rare-earth "priced in" because returns outpaced its 4% AI share — but the TAM math says only 5% of its potential AI exposure is realised. Adding −0.5·z(log(tam_uplift_multiple)) to the gap (so high-uplift verticals are pulled toward lagging) reclassifies:
copper-rare-earth: priced_in → laggingutilities-merchant-power: priced_in → fairdatacenter-cooling-thermal: lagging → fairmodel-labs-software: fair → priced_inadvanced-packaging: fair → priced_inThis is the single most consequential change. The ranking is currently a cross-section of price action vs. a static revenue mix; adding TAM uplift converts it to price action vs. forward AI exposure, which is what an investor actually wants.
Revised formula:
gap_revised = 1.0·z(total_return_3y) + 1.0·z(total_return_chatgpt)/2
− 1.0·z(ai_share_today_pct)
− 0.5·z(log(tam_uplift_multiple))
Rationale: (a) drop the beta term — confounded and barely contributes; (b) average two return windows (3y and ChatGPT-launch) to defuse window-snooping; (c) add log-TAM-uplift with a negative sign so verticals with large forward AI revenue runway are pulled toward lagging; (d) report a continuous gap value with a diverging colour scale instead of terciles.
Under just the TAM-augmented revision (3y window only, dropping beta is left for follow-up), the top-3 lists become:
Top-3 lagging (revised):
lithography (gap −2.29) — same as before; the high-AI-share, low-return anomalyindustrial-gases-water (gap −1.46) — moved up the lagging list because of its 8.9× TAM upliftdatacenter-reits (gap −1.38) — unchanged in positionTop-3 priced-in (revised):
silicon-photonics-optics (gap +4.04) — extreme rally + modest TAM ceiling = clearly extendedpower-transformers-grid (gap +1.83) — 11× return + modest TAM upliftnuclear-smr-uranium (gap +1.29) — high return relative to both today's AI share and forward TAMMost important reclassification: copper-rare-earth exits priced_in, becomes lagging. A 122% three-year return looks like a rally until you realise current AI share is only 4% and the addressable AI revenue if utilisation hits 80% is 20× today's level. The current methodology punishes copper for going up at all; the revised methodology recognises there is 19× of upside still on the table. That single re-label is the cleanest demonstration that the current gap formula is incomplete.
The methodology is internally consistent but semantically slippery. It reliably identifies two extremes — old-economy verticals that have rallied on AI narrative with little AI revenue (correct "priced in") and high-AI-share semiconductor verticals (mislabelled "lagging" because the cross-section can't see what's already monetised). The middle is noise: 11 of 22 labels flip on a 10pp wiggle of an estimated AI share. Adding TAM uplift and dropping beta is the highest-leverage change.
After HBM, CoWoS-L / SoIC / FOPLP capacity is the hard physical bottleneck for AI accelerator output: every NVIDIA Blackwell/Rubin and AMD MI3xx die needs interposer + HBM stacking + substrate before it ships. TSMC's CoWoS output is scaling from ~35k wpm (late 2024) toward ~75k wpm (end 2025) and ~130k wpm targeted by end 2026, with NVIDIA reportedly booking >50% of 2026 CoWoS capacity. OSATs (ASE, Amkor, KYEC, PTI, Chipbond) and ABF substrate makers (Unimicron) capture the spillover — flip-chip, fan-out, FOPLP, and back-end test — and are raising 2026 capex sharply (ASE +$1B, Amkor to $2.5-3B vs $0.9B 2025).
AI-share sources: yolegroup.com · trendforce.com
AI accelerators are the literal substrate of LLM inference — every token generated runs on a GPU, TPU, or custom ASIC in this vertical. NVIDIA's Blackwell ramp plus the hyperscaler custom-silicon wave (Broadcom-designed Google TPU and OpenAI ASIC, Marvell-designed AWS Trainium and Microsoft Maia) is the single biggest dollar bucket in the entire AI capex stack and the tightest binding constraint on inference cost and capacity. Whoever owns the accelerator socket captures the highest gross margin in the chain, which is why merchant GPU economics (NVDA 70%+ GM) and custom-ASIC partnerships (AVGO/MRVL) both compound from here.
Copper for AI datacenter busways, transformers, cabling and substation transmission is the single largest commodity-tonnage input enabling US/EU AI infrastructure scale-out — every 100 MW campus consumes ~2,700-3,300 t Cu. The S&P Global 'Copper in the Age of AI' (Jan 2026) report confirms data center copper is just ~1.1 Mt of 28 Mt global refined copper in 2025 (~4%), so incremental AI demand is a small share of total mined copper but matters acutely at the margin because the market is already in deficit (Wood Mackenzie: 304 kt shortfall 2025) and 30-year supply elasticity is low (declining ore grades, 15+ year mine lead times). Rare earths (NdPr for permanent magnets in HVAC fans/pumps and high-efficiency motors) are a much smaller dollar TAM (~$4-8 Bn REE oxide market) but geopolitically critical: MP/LYC are the only non-China at-scale producers and US export-control reciprocity has made them strategic call-options.
AI-share sources: spglobal.com · press.spglobal.com
Rack power densities are jumping from ~10–20 kW (CPU era) to 100–250 kW (Blackwell NVL72 and beyond), and air cooling physically cannot remove that heat at the chip — so direct-liquid cooling, rear-door heat exchangers, and (for some workloads) immersion become mandatory. That forces every new AI hall to add CDUs, manifolds, cold plates, and higher-temperature chillers on top of (not instead of) traditional CRAH/chiller spend, so the cooling vertical compounds with both AI capex and AI rack density. Dell'Oro's Direct Liquid Cooling segment grew 156% Y/Y in 2Q25 and is the single fastest-growing line in the entire AI supply chain.
AI-share sources: delloro.com · delloro.com
Datacenter REITs are the physical real estate layer of LLM inference — every token served runs inside one of these facilities or a hyperscaler-owned equivalent. AI demand has flipped the cycle from oversupply to severe undersupply: CBRE/JLL/Cushman all show colocation preleasing at 70-80%+, hyperscale wholesale rents up double-digits YoY, and bookings concentrated in AI tenants (50-60%+ of new leases at EQIX/DLR Q3-Q4 2025). The investability question for this vertical is whether power, grid interconnects, and zoning let the REITs deliver against backlogs at promised cap rates, or whether private capital (Blackstone QTS, KKR-GIP CyrusOne, DigitalBridge, Compass) ends up capturing most of the AI growth offshore from public-equity investors.
AI-share sources: fool.com · finance.yahoo.com
Every AI accelerator on earth is designed with tools from Synopsys or Cadence and licenses IP from Arm, Synopsys DesignWare, or Cadence — making the EDA/IP layer the highest-margin, lowest-capex chokepoint in the LLM-inference stack. The Big-3 EDA vendors hold >85% combined share with effectively no credible new entrants (Synopsys' Ansys close in Jul 2025 further consolidated multi-physics), and EDA has outgrown semiconductor R&D by ~6pp/yr since 2018 as hyperscaler custom-silicon programs multiply the number of design starts and verification compute hours. Owning SNPS+CDNS+ARM">ARM gives picks-and-shovels exposure to every AI silicon shop simultaneously, with subscription revenue models that smooth the wafer cycle.
AI-share sources: newsletter.semianalysis.com · sec.gov
Hyperscale AI halls are 100-300 MW each and require an integrated stack of medium-voltage switchgear, transformers, busways, UPS, PDUs, and (increasingly) coolant distribution units — exactly the catalog of Eaton, Schneider, ABB, and the rack-level specialists nVent and Vertiv. Lead times have stretched from ~12 weeks pre-2023 to 50-80+ weeks for MV gear and large UPS modules, so backlogs (Eaton $19.6B, Powell $1.6B) are effectively pre-booked revenue through 2027. The vertical is the most direct industrial-economy chokepoint between an AI accelerator order and an actual running cluster, which is why every name here re-rated 3-5x off 2022 lows even as their non-DC end markets stayed flat.
AI-share sources: utilitydive.com · futurumgroup.com
AI accelerators are produced almost exclusively on TSMC's N5/N4/N3 (and soon N2) lines, making the leading-edge logic foundry the single largest physical chokepoint in the LLM-inference supply chain. Demand is supply-constrained, not demand-constrained: capacity utilization at advanced nodes is sold out through 2026 and capex is being pulled forward. The only credible competitive risk to TSMC's near-monopoly is Intel Foundry's 18A/14A ramp, which is why INTC is in the basket as an asymmetric option even though most logic-foundry AI revenue today still lands at TSM.
AI-share sources: futurumgroup.com · ainvest.com
Behind-the-meter and near-meter gas peakers/CCGTs are the fastest-deployable answer to the multi-year grid-interconnect queue that is the binding constraint on AI campus buildouts; this has driven gas-turbine new-unit orders to ~70 GW in 2025 with backlogs at the big-three OEMs stretching to 2029-2030 and prices up 195% by 2027 (Wood Mackenzie). The vertical is structurally supply-constrained — global mfg capacity ~60-70 GW vs. ~110 GW order book — which gives GE Vernova, Siemens Energy and MHI durable pricing power and visibility unmatched anywhere else in the energy stack. Gensets (CAT/CMI/GNRC) and fuel cells (BE) ride the same demand wave at the smaller / faster-deploy tier and are included as a barbell to the three large OEMs.
AI-share sources: power-eng.com · utilitydive.com
HBM3E (and ramping HBM4) is the gating bottleneck for AI inference throughput: every Nvidia H100/H200/B200 and AMD MI300/MI355 carries 6-8 HBM stacks, and SK Hynix/Samsung/Micron have sold out 2025 and largely 2026 capacity. Unlike commodity DRAM, HBM is sold on multi-quarter LTAs at 3-5x ASP premium, so the AI-cycle revenue mix inside DRAM-makers is shifting rapidly upward. The vertical compounds two signals: (1) commodity DRAM cycle, (2) AI-specific HBM scarcity rent — only SK Hynix is currently capturing the latter at scale.
AI-share sources: trendforce.com · trendforce.com
Hyperscalers are the demand pull of the entire LLM inference supply chain — every dollar of NVDA/AVGO/TSMC/HBM revenue ultimately traces back to a capex line on one of these eight balance sheets. Their stocks therefore encode the market's belief about AI return-on-invested-capital rather than AI inference volumes per se: when MSFT/GOOGL/AMZN/META capex compounds at 30-50% YoY but cloud AI revenue grows slower than capex, multiples compress (the 2025 Q3 'where's the ROI?' selloff). Critical cause-vs-effect caveat: a falling hyperscaler stock can mean either (a) AI inference is structurally less profitable than hoped, OR (b) capital intensity is reaccelerating ahead of monetization — only (a) is bearish for the rest of the supply chain, and (b) is in fact bullish for accelerators, power, and DC REITs. So this basket is the AI-ROI sentiment proxy, not the AI-inference-volume proxy.
AI-share sources: news.microsoft.com · sec.gov
IC substrates — specifically high-layer-count ABF/FC-BGA substrates — are the chokepoint that sits between the silicon die and the PCB on every AI accelerator package: CoWoS interposers ride on these substrates, and the substrate sub-segment is currently sold out with multi-quarter lead times. Capacity is concentrated in a tight oligopoly (Ibiden, Unimicron, Samsung Electro-Mechanics, AT&S, Kinsus, Nan Ya PCB) where greenfield lines take 2-3 years and require Ajinomoto's proprietary ABF film as a sole-sourced input. As AI server unit volumes scale and package sizes grow (more substrate area per accelerator), this vertical's AI mix should keep climbing through 2027-2028 even before any TAM expansion from co-packaged optics and glass-core substrates.
AI-share sources: digitimes.com · investing.com
Industrial gases are an under-appreciated picks-and-shovels exposure: every leading-edge fab consumes huge volumes of high-purity N2, Ar, He, H2, and specialty gases on long-dated on-site contracts, and Linde/APD/Air Liquide are an oligopoly with rate-base-like economics. Water is the second-derivative bet — AI data centers and the new US/EU fab buildouts are the marginal load that strains local water tables, but the pure water-utility names (AWK, WTRG) get paid on regulated tariffs that lag and barely move on AI demand, so the cleaner AI exposure inside this vertical sits in the gas majors plus water-tech (XYL, PNR) rather than in the rate-based utilities.
AI-share sources: assets.linde.com · airproducts.com
Lithography — and specifically EUV — is the narrowest bottleneck in the LLM inference stack: ASML has a 100% monopoly on EUV scanners required for 5nm/3nm/2nm logic and HBM4 DRAM, while Lasertec has a near-monopoly on the actinic mask-inspection that gates every EUV mask. Without these tools no AI accelerator (NVIDIA Blackwell/Rubin, AMD MI400, Google TPU v7) and no HBM3E/HBM4 stack can be manufactured at volume, making this the single most concentrated supply-chain choke point. The vertical is also among the most-priced-in: ASML 2025 backlog hit €38.8B and EUV revenue grew 39% YoY, but China export controls and lumpy High-NA ramp create asymmetric risk versus the broader AI capex bull case.
AI-share sources: asml.com · tomshardware.com
This is the demand side of LLM inference — the application-layer companies whose product economics depend on calling models in production at scale. The cleanest exposures are AI-native ad-tech (APP) and pure 'sell AI to the enterprise' platforms (PLTR), backed by software incumbents monetizing inference through AI add-on SKUs (NOW Now Assist) and inference-adjacent picks-and-shovels (DDOG observability, SNOW Cortex, MDB vector DB). Most pure model labs (OpenAI, Anthropic, xAI, Mistral, Cohere) are private; public exposure is via these consumption-layer proxies where genuine AI-driven revenue is being disclosed in earnings.
GPU clusters scale only as fast as their fabric: every Hopper/Blackwell/MI300 rack needs ~$0.20-0.30 of switching, retiming and optics per $1 of compute, and a meaningful fraction of inference latency lives in the network. Dell'Oro pegs cumulative AI back-end switch spend at >$100B over 2025-2029, with Ethernet rapidly displacing InfiniBand (>2/3 of back-end ports in 3Q25). The basket is built to capture both the OEM layer (ANET, CSCO, HPE) that monetizes shipped boxes and the silicon/connectivity layer (MRVL, CRDO, ALAB) that monetizes per-bit serdes growth from 100G -> 200G -> 400G/lane.
AI-share sources: delloro.com · delloro.com
Hyperscalers have moved from 'we need green PPAs' to 'we need 24/7 firm power and we will pay double wholesale for it' — and there is no other large carbon-free baseload source. That re-rates existing US nuclear operators (CEG, VST, TLN) on long-dated PPAs at premium prices, opens a venture-style optionality on the SMR designers (OKLO">OKLO, SMR), and tightens the fuel cycle (CCJ uranium, LEU/BWXT for HALEU and components). The actual AI-tied revenue today is single-digit percent of the vertical, but the marginal buyer of every incremental MW and pound of U3O8 is now AI infrastructure, which is what the equity prices already partially discount.
AI-share sources: utilitydive.com · constellationenergy.com
Each next-gen AI accelerator socket is pulling 600-1000A today and >1500A on the 2026-27 roadmaps, so the only way to keep IR-drop and conversion losses tolerable is to move the final DC-DC stage from the motherboard edge to directly under the die — Vertical Power Delivery. That structural shift, combined with the industry-wide migration to 800VDC rack architectures (NVIDIA AI Factory), is forcing a near-complete redesign of the power tree and pulls in more silicon content per rack (Infineon guides $12-15k of power-semi BOM per 130kW rack vs. low-single-digit-thousands for traditional server racks). The basket combines the incumbent VRM specialists (MPWR, VICR) with the broad-line analog/power players (IFX, ADI, TXN, ON) and the wide-bandgap optionality (POWI, NVTS) where GaN/SiC content scales fastest.
AI-share sources: simplywall.st · sec.gov
Large power transformers (3-5 year lead times, 77% price increases since 2019, Wood Mackenzie projecting a 30% global supply deficit for power transformers in 2025) are the silent physical chokepoint of the AI build-out — every gigawatt of hyperscale capacity needs MV/HV step-down hardware that only a handful of OEMs can produce, and U.S. demand is up 119% since 2019. The basket pairs the four global transformer oligopolists (GEV/Prolec, Hitachi Energy via 6501.T, Siemens Energy, HD Hyundai Electric) with the three largest U.S. T&D contractors who must physically install grid upgrades for AI campuses (PWR, MYRG, PRIM), plus Hammond Power as a pure-play dry-type / specialty exposure. Capacity expansions (GEV Prolec, Hitachi $457M VA plant, Hyundai Alabama, Siemens $1B U.S. capex) won't materially relieve the shortage until 2028+, so the supply-constrained pricing regime should persist through the 2026-2027 inference-capex peak.
AI-share sources: spglobal.com · marketscreener.com
Optical interconnects are the rate-limiter for scale-out AI clusters once GPU and HBM ship: each XPU now needs multiple 800G or 1.6T ports for back-end fabric, and roadmaps to 3.2T plus co-packaged optics (CPO) collapse pluggable margins toward the laser/EML and packaging layers. Lumentum and Coherent control the scarce 200 Gbps-per-lane EML supply that gates the 1.6T ramp; Fabrinet captures the packaging-as-a-service rent for Nvidia/Cisco/Ciena. The vertical is at ~55% AI exposure today and is the highest-conviction leveraged bet on cluster bandwidth growth, but is also where CPO disruption (Nvidia/Broadcom/TSMC) could compress the pluggable TAM after 2027.
AI-share sources: lightcounting.com · cignal.ai
Utilities and IPPs are the physical bottleneck for AI inference: training/inference racks need 24/7 firm power within transmission-feasible distance of fiber backbones, and new generation + interconnect takes 4-7 years to bring online. The merchant IPP cohort (VST, TLN, CEG) is being rerated as a quasi-infrastructure asset class because long-dated PPAs with investment-grade hyperscaler counterparties convert previously-cyclical merchant cash flows into bond-like contracted revenue, while regulated utilities in data-center alleys (D in Virginia, AEP in Ohio, PEG in NJ, DUK in the Carolinas) capture growth via rate-based capex on transmission upgrades, gas peakers, and large-load tariffs. The risk asymmetry favors owners of existing dispatchable capacity (nuclear, CCGT, coal-to-gas conversions) over greenfield developers.
AI-share sources: spglobal.com · eia.gov
Every wafer that ends up in an LLM accelerator or HBM stack passes through deposition, etch, implant, clean, CMP, and metrology tools — making this the broadest, least-bypassable layer of the AI silicon supply chain after litho. The Big Five (AMAT, LRCX, KLAC, TEL, ASML) are an oligopoly with high switching costs, and AI is steepening the WFE intensity per wafer (more layers, more EUV-adjacent steps, more 3D stacking, tighter process-control budgets). Advanced packaging metrology (KLAC, ONTO) and HBM-specific deposition/etch backlogs (LRCX, AMAT) are the highest-beta sub-segments to AI capex.
%%{init: {'themeVariables': {'fontSize':'14px','fontFamily':'system-ui'}, 'flowchart':{'curve':'basis'}}}%%
flowchart TD
P[Plan: vertical taxonomy
plan/MASTER_PLAN.md] --> A[22 parallel vertical-fact agents
tickers + AI share + thesis]
A --> V1{Verify JSON schema}
V1 --> F[Price fetch: yfinance
141 tickers · 2021-05 → 2026-05]
F --> V2{Verify ≥36mo history}
V2 --> R1[returns_vertical.csv
equal-weight indices]
V2 --> R2[risk.csv
vol, beta_NVDA, max DD]
V2 --> R3[ranking.csv
z-scores + gap]
V2 --> R4[tam.csv
uplift if AI hits 80%]
R1 --> V3{Verify CSVs · cross-check}
R2 --> V3
R3 --> V3
R4 --> V3
V3 --> H1[HTML v1 build]
H1 --> FB[User feedback]
FB --> C1[Critique: ai-accelerators]
FB --> C2[Critique: priced-in drivers]
FB --> C3[Critique: methodology]
FB --> CL[Chart lib research → uPlot]
FB --> GL[Glossary builder: 387 entries]
C1 --> ES[Exec-summary draft]
C2 --> ES
C3 --> ES
CL --> H2[HTML v2 build]
GL --> H2
ES --> H2
H2 --> V4{Verify v2}
V4 --> Z[Open in Zen browser]
Source: data/verticals/*.json (22 hand-curated fact sheets); analysis/*.csv (computed from yfinance Adj Close, 2021-05 → 2026-05); critiques in analysis/critique_*.md; glossary 387 entries (137 tickers · 100 concepts · 126 companies · 22 verticals · 2 indices). See plan/MASTER_PLAN.md for full pipeline. Run 2026-05-26.