Eight axes for the LLM supply chain (interactive v2)
Why v2 exists (v1 retro)
v1 combined three numbers: 3-year total return, NVDA beta, current AI-revenue share. Z-score two, subtract, call the result a “gap,” sort 22 verticals into priced-in / fair / lagging. Three critique passes broke it: the beta term added nothing, tercile cutoffs flipped on n=22, and 11 of 22 labels swapped sides when the AI-share prior shifted 10 points.
v1 also had no long-run pool. It could not separate “already monetised” from “yet to be monetised” because it never anchored the pie. v2 pins a fixed economic premise, scores each vertical on 8 orthogonal axes, and treats the composite as an aggregate, not an oracle.
AI’s economic uplift will be 3x regular computing
The slider's m-multiplier is applied with per-vertical elasticity e_i (sourced from each vertical's D8 Jevons score: model-labs and hyperscalers ~2.0, commodities ~0.5). Capture rate c is blended 50/50 with a layer-specific c_layer_i (D4-sourced: software/EDA ~35%, commodities ~5%). Drag the sliders — picks actually shift, not just the dollars.
BEA pins the US digital economy at $2.6T. For a global supply chain that's a US-only proxy — global ICT-driven frontier productivity, ex EM catch-up, lands in roughly the same $2-4T/yr range (Jorgenson-Stiroh ICT-TFP attribution; OECD frontier productivity). Apply 3×, 20.0% vendor capture, and the implied annual vendor pool at maturity (~2035) is $1.56T, split across 22 verticals.
How we got there
Restated for the math. Every $1 of cyber-physical TAM implies $3 of AI-stack TAM over the next 5-10 years. The headline axiom above is the same claim, framed as the article’s load-bearing assumption.
Baseline. BEA Digital Economy Satellite Account: US digital economy at 10.0% of GDP, $2.6T in 2022. I pick BEA over McKinsey/Goldman because it’s a government current-state number, not a consultancy projection. Stacking 3x on top of a projection would double-count.
BEA Digital Economy is a stock (current sector size), not a flow (counterfactual uplift). ICT-TFP flow estimates (Oliner-Sichel, Jorgenson-Stiroh) land in the same ~$2.5T/yr range. GDP-B welfare-inclusive measures (Brynjolfsson NBER w25695) suggest ~$4.5T. I pick BEA for traceability, not for it being the highest-fidelity measure. (Critique A.)
Multiplier and capture. 3 × $2.6T = $7.8T annual AI GDP uplift at maturity (~2035). Vendor capture in historical IT runs 15-25% (Bain $990B / McKinsey $2.6-4.4T ≈ 28%; global software vs BEA digital ≈ 26%). The slider’s c sets the flat anchor for a per-vertical heterogeneous blend; default 20.0% reproduces the article’s headline pool of $1.56T when paired with article defaults for m and B.
The 3x is the consensus-of-consultancies upper-tercile, not a median forecast. Defensible 90% confidence range is roughly 0.5x (Acemoglu, OECD pessimistic case) to 5x (PwC, McKinsey value-creation framings). Picking 3x isn’t conservative. At 1x the pool shrinks to ~$500B; at 5x, ~$2.6T. The composite ranking is approximately scale-invariant – the picks don’t change, but the absolute headroom claims do. (Critique B.)
e_i = 0.5 + 1.5 × (d8_score_i / 10) # m-elasticity per vertical c_layer_i = 0.05 + 0.30 × (d4_score_i / 10) # layer-specific capture (D4) c_blend_i = 0.5 × c + 0.5 × c_layer_i # flat-vs-layer 50/50 blend raw_i = share_i × B × c_blend_i × m^e_i # raw per-vertical implied_i = raw_i / Σraw × (B × m × c) # renormalize to headline pool
Top 3 by implied 2035 AI revenue:
| Rank | Vertical | Implied 2035 AI revenue |
|---|---|---|
| 1 | hyperscalers-cloud | $609.8B |
| 2 | ai-accelerators | $224.7B |
| 3 | model-labs-software | $192.5B |
Biggest pool is not best bet. “Pool size” and “remaining upside” are different questions.
- 3× upper-tercile (range 0.5× to 5×; currently 3×). The slider's m is applied with per-vertical elasticity e_i ∈ [0.5, 2.0], sourced from each vertical's D8 score. Most academic economists sit well below 3×.
- Capture rate is heterogeneous: c_i = 0.5 × c + 0.5 × c_layer_i, where c_layer_i ∈ [0.05, 0.35] is derived from each vertical's D4 score (EDA/software ~0.35, commodities ~0.05). The slider's c (currently 20.0%) sets the flat anchor; the blend produces 5–50× layer variation as observed in industry.
- $2.6T US-only is a proxy for global ICT-frontier uplift. A naive global digital-economy baseline (~$16T) overstates because it includes emerging-market industrialization, which came from labor reallocation, not computers. The right baseline (developed-world ICT-TFP frontier ex EM catch-up) is roughly $2–4T/yr (currently $2.6T), putting our $1.56T pool inside the right order of magnitude.
- Annual maturity at 2035 vs cumulative NPV -- this is a rate, not a present value. No discount rate stated. The current annual pool is $1.56T.
- Reference P/S = 3.0 (S&P median). Vertical-specific multiples (6x software, 1.5x utility) would change the gap. At 3× uplift on $2.6T, gross AI revenue is $7.8T.
Eight axes the v1 ranking ignored
Each vertical gets eight 0-10 scores, one per dimension. The strongest pair (D3 supply elasticity vs D8 Jevons demand) hits |r| = 0.642, below the 0.7 merge threshold. All eight survive.
- D1 Already-rallied penalty
- 5-year total return, inverted. 10 = least rallied. A 200% rally on the same story leaves less juice.
- Top: nuclear-SMR (priced for the rally) | Bottom: lithography (post-EUV run)
- D2 Premise-implied TAM headroom
- Premise-implied 2035 AI revenue (§1) minus today, log-scaled and rank-percentiled. Only forward-revenue axis.
- Top: copper-rare-earth (premise_gap_log = 1.36) | Bottom: ai-accelerators (already monetised)
- D3 Supply elasticity / bottleneck severity
- Lead times to add a unit. Long = inelastic = pricing power. EUV ~18 mo. CoWoS 12-24. Copper 5-15 yr. SMRs 5-10. Transformers 18-48 mo.
- Top: nuclear-SMR 9.51 / copper-rare-earth | Bottom: model-labs-software (instant scale)
- D4 Value-capture intensity
- Gross margin × moat width. Where the dollar lands inside the value chain.
- Top: eda-ip 9.48 (80%+ GM duopoly) | Bottom: nuclear-SMR 0 (capex-heavy, fragmented)
- D5 Substitution risk
- Probability the dominant solution gets displaced in 10 years. CPO over pluggable; ASICs over GPUs; SMR over gas peakers; liquid over air.
- Top: nuclear-SMR 10 (no baseload substitute) | Bottom: ai-accelerators 1.25 (rack-level ASIC risk)
- D6 Capex × cycle position
- Capex / forward revenue × cycle stage. Early-cycle premium, late-cycle penalised. D3 correlation r = 0.37.
- Top: datacenter-reits 10 | Bottom: eda-ip 0
- D7 Geopolitical exposure
- Revenue / supply / customer exposure to friendly jurisdictions. Higher = safer.
- Top: copper-rare-earth 10 (CHIPS-funded caveat) | Bottom: ic-substrates 0 (JP/TW/KR + PRC)
- D8 Jevons elasticity
- 10x inference-cost drop -- demand elastic (high) or flat (low)? Demand mirror of D3.
- Top: hyperscalers-cloud, model-labs-software 10 | Bottom: copper-rare-earth 0
How we got there
| Eight axes, 0-10, higher = more remaining upside. Phase 3 orthogonality at | Spearman r | > 0.7; strongest pair (d3_supply_elasticity × d8_jevons) r = -0.642, below the bar. |
The eight carve the question: price (D1), forward revenue (D2), supply (D3), profitability (D4), tech trajectory (D5), capex (D6), jurisdiction (D7), demand elasticity (D8).
Equal-weight rank, top to bottom
Unweighted mean of the eight scores. Top of the table: infrastructure picks-and-shovels (hyperscalers-cloud 6.97, copper-rare-earth 6.81, industrial-gases-water 6.72). Bottom: already-priced AI silicon plus speculative optics. Only power-semis-vrm gains rank under both premise-tilt and contrarian-tilt -- the cleanest asymmetric pick in the matrix.
Top-5, equal weight
Bottom-5
How we got there + tilt-shift sanity checks
Equal-weight composite: unweighted mean of the eight scores. Top: infrastructure picks-and-shovels (power, water, copper, hyperscale platforms). Bottom: already-priced AI silicon plus speculative optics. Headline: under an explicit premise and 8-axis scoring, the unsexy physical-input layers outrank the silicon layers everyone bought for 36 months.
Premise-tilt (D2 ×2). Rewards supply-constrained names with unpriced forward demand.
Contrarian-tilt (D1 ×2). Surfaces names that haven’t moved.
The asymmetric pick: power-semis-vrm gains under BOTH tilts. Supply constrained, unpriced, stocks haven’t run. Rare combination. Board-power / VRM (MPS, Vicor, Infineon PMIC) is the cleanest setup.
| vertical | composite equal |
rank equal |
composite premise |
rank premise |
composite contrarian |
rank contrarian |
Δ premise vs equal |
Δ contrarian vs equal |
|---|
Nuclear/SMR jumped 16 places vs v1
v1 saw a 4.2x rally and called it "priced_in". v2 reads four more axes -- supply, substitution, geopolitics, headroom -- all pointing the same way.
How we got there
v1 ranked nuclear-smr-uranium #21 of 22, “priced_in“: 4.2x 3-year return, 4% AI-share-today, no substrate. v2 keeps the rally penalty but adds four axes that all favour SMR:
| Axis | Score | What it measures |
|---|---|---|
| D2 headroom | 9.29 | Premise-implied 2035 TAM minus today |
| D3 supply | 9.51 | 5-10 yr permitting, most inelastic |
| D5 substitution | 10 | No baseload alternative inside 10 yrs |
| D7 geopolitics | 10 | US/Allied uranium plus DOE LPO |
Under contrarian-tilt (D1 doubled) the rank drops to #9. Asymmetric, not unconditional – a forward-supply-curve bet, not a momentum bet.
Where this is still wrong
Eight axes beats three numbers, but it isn't the truth. Premise-level weaknesses listed in the five caveats above (3× upper-tercile, 20.0% blended capture, $2.6T US-only baseline, annual-vs-NPV horizon, flat P/S). Method-level weaknesses below: fragility on top-decile names, a judgement-call allocation split, substitution risk scored from notes not markets, and n=22.
datacenter-reits swings 8 ranks under leave-one-out: drop D2 and the thesis collapses. power-transformers-grid ranges 7-18. lithography 7-18. model-labs-software 3-15. Not diversified theses. Position-size accordingly.
Robust longs (smallest leave-one-out ranges)
copper-rare-earth (3), industrial-gases-water (3), utilities-merchant-power (3), wfe-deposition-etch (4), hyperscalers-cloud (4). Three are top-5.
The matrix is a sketch of where picks-and-shovels sit when the rally is held to an explicit premise. More honest than v1. Not a buy list.
(Leave-one-out ranges are computed at default premise. Live LOO redraws above; ranges shown here are the v1 defaults.)
References & further reading
Predecessor
- Earlier framing: 3-input gap-metric ranking (v1) — v1 ranking this post rebuilds: 3-yr return, NVDA beta, AI-revenue share, z-scored into a “gap”.
Contrarian critiques (internal reviews)
Each premise input was challenged by a dedicated internal review; findings folded into the caveat callout above. Notes are in the repo, not published.
- Critique A — baseline. BEA = stock not flow; ICT-TFP ~$2.5T/yr; GDP-B ~$4.5T. Current slider B = $2.6T.
- Critique B — multiplier. 90% CI [0.5x, 5x], median ~1.5x. 3x sits at the 80th percentile. Current slider m = 3×.
- Critique C — capture rate. Range [10%, 35%]; EDA / hyperscalers ~35%, mid-stack 15-20%, commodities 5-8%. Current slider c = 20.0%.
- Critique D — US vs global. Hybrid: $2.6T US for US vendors, ~$16T global for foreign. Global pool ~$9.6T. Current implied pool: $1.56T.
- Critique E — time horizon. Annual maturity, not NPV. No discount rate stated. Current uplift: $7.8T.
Premise (the 3x baseline)
- BEA Digital Economy Satellite Account, SCB Dec 2023 — US digital economy 10.0% of GDP, $2.6T in 2022. The baseline. (infographic PDF)
- McKinsey — “The economic potential of generative AI” — $2.6-4.4T annual gen-AI value, 63 use cases. Capture-rate endpoint #1. (Slider c = 20.0%.)
- Bain Global Technology Report (2024) — $990B AI products and services by 2027. Capture-rate endpoint #2. (Implied pool: $1.56T.)
- Goldman Sachs — Briggs / Kodnani, “Generative AI could raise global GDP by 7%” — 7% / ~$7T global GDP uplift over 10 years. 3× cross-check.
- Stanford AI Index 2025 — economy chapter — $252.3B 2024 corporate AI investment; $33.9B private gen-AI. (Total at-maturity uplift: $7.8T.)
- WIPO — global software industry $675B in 2024 — software vs BEA digital ratio ~26% historical capture. (Slider c = 20.0%.)
Per-dimension data sources
- D1: per-vertical 5-yr total returns via yfinance / Yahoo Finance, equal-weight per
data/verticals/*.json. - D2: derived from
phase1_allocations.csv(external inputs under “Premise” above). - D3: Silicon Analysts — TSMC CoWoS Q1 2026; MobileWorldLive — ASML EUV €39B backlog; IEA — Future Transmission Grid; eepower — transformer supply; Utility Dive — GE Vernova 80GW gas-turbine backlog; SMR Intel — NRC tracker; Mining Visuals — 17.9-yr copper discovery-to-production; Introl — HBM 2026 supercycle; Carbon Direct — interconnection queues.
- D4: SEC EDGAR 10-K/Q gross margins via yfinance (US); Yahoo per-ticker for non-US.
- D5: SemiAnalysis cluster reads, IDTechEx CPO 2026-2036, Menlo Ventures LLM survey, Stordis UEC 1.0.
- D6: MSFT 10-K; TSM 10-K. Capex/sales + cycle stage.
- D7: US Commerce CHIPS preliminary terms (Intel, TSMC, Samsung, Micron, GF); BIS Oct-2022 / Oct-2023 export controls; IEA Electricity 2024; company IR (Linde, Coherent, Arista, Vertiv).
- D8: a16z — Economic Case for Generative AI; SemiAnalysis “AI Datacenter Energy Dilemma” / “Power Crisis”; RAND — AI’s Power Requirements.
