# AwarenessFund — Competing Investment Hypotheses **Author**: research-lead (A-C-Gee) **Date**: 2026-05-18 **Audience**: Russell Korus, Apex/Jordana (Pyonair), Corey, capital-markets-lead, sister AiCIV insiders **Frame**: *Analytical substrate for thesis discussion — not investment advice. Each hypothesis is structured for falsification, not advocacy. Confidence levels are PRELIMINARY pending the 24-month back-test.* --- ## Executive Summary The fund's core thesis is **AI-supply-chain capex shift across mining → energy → chip-fabs → AI-endpoint over a 5-year horizon**. Below are **six competing hypotheses** that could each independently explain (or undermine) the thesis. The back-test design must discriminate between them — not just confirm the consensus view. Confidence levels are PRELIMINARY (HIGH/MEDIUM/LOW pending evidence). Each hypothesis carries explicit falsification conditions. --- ## Hypothesis Matrix | # | Hypothesis (short) | Direction | Preliminary Confidence | Time-Horizon | |---|-------------------|-----------|------------------------|--------------| | **H1** | Picks-and-shovels outperforms concentrated hyperscaler bets | Pro-thesis | MEDIUM-HIGH | 2-5 yr | | **H2** | Energy bottleneck is THE moat — utility/grid operators dominate | Pro-thesis | MEDIUM-HIGH | 3-5 yr | | **H3** | Mining is the late-cycle play (2027-2030 inflection) | Pro-thesis (delayed) | LOW-MEDIUM | 3-7 yr | | **H4** | Chip-fab oligopoly concentrates (NVDA/TSMC pure-play beats equipment-makers) | Counter-thesis (within stack) | MEDIUM | 1-3 yr | | **H5** | AI-endpoint data-center capex peaks within 18mo (S-curve top) | Counter-thesis | LOW-MEDIUM | 1-2 yr | | **H6** | The thesis is correct but the S&P 500 already prices it (no fund-alpha) | Anti-thesis (market efficiency) | MEDIUM | continuous | --- ## H1 — Picks-and-shovels outperforms concentrated hyperscaler bets **Claim**: A diversified basket across mining + energy + chip-fab equipment + data-center infra outperforms concentrating in hyperscalers (NVDA/MSFT/GOOG/META) over the 24-month back-test window. **Mechanism**: Hyperscaler margins compress as competition intensifies (open-source models, custom silicon); supply-chain vendors capture the margin because demand exceeds supply across raw inputs. **Evidence FOR (current)**: - ASML/AMAT/LRCX margins expanding 2023-2025 even as NVDA gross margins flatten - TSM capex acceleration outpacing customer revenue growth - Energy capex (utilities, transformers) seeing first-time AI-attributable demand - Historical analog: Cisco-era infra plays (RBOCs, Juniper) outperformed concentration in single names **Evidence AGAINST**: - 2023-2024 NVDA delivered 200%+ return — concentration crushed diversification - Hyperscaler margin compression hypothesis has been wrong for 5+ years - Picks-and-shovels diversification dilutes the alpha **Falsification test**: Run portfolio of equal-weighted top-5 names per vertical (20 names) vs equal-weighted top-5 hyperscalers, 24mo back-test. H1 confirmed if vertical basket Sharpe > hyperscaler basket Sharpe. **Confidence**: MEDIUM-HIGH — historical precedent is strong but recent 2-yr period has been brutal for the hypothesis. --- ## H2 — Energy bottleneck is THE moat **Claim**: Power generation/transmission/grid-interconnect is the binding constraint on AI scaling. Utilities, gas-peaker operators, transformer OEMs (Hitachi, Siemens Energy), and behind-the-meter providers capture disproportionate alpha. **Mechanism**: US grid needs +1 TW capacity over 3-5 yr to support AI compute. Interconnect queue is 3-7 yr deep. Capacity ⇒ pricing power ⇒ utility re-rating. **Evidence FOR**: - Hyperscaler PPAs at premium pricing (Microsoft signed Three Mile Island restart, etc.) - Vistra, NRG, Constellation 2024-2025 returns (200-400%+) on AI-energy narrative - FERC interconnect queue data: 2.6 TW in queue, ~25% completion rate - Natural gas turbine (Mitsubishi, GE Vernova) backlog at multi-year highs **Evidence AGAINST**: - Utility valuations already pricing this — entry now is late - Regulatory drag (state PUC approvals slow re-rate) - Solar+battery costs may flip the storyline before bottleneck binds - Hyperscaler self-build (Stargate, Meta nuclear) bypasses public utilities **Falsification test**: Back-test energy basket vs broad AI basket over 24mo. Confirm if energy-basket Sharpe > broad-AI Sharpe AND grid-interconnect-queue depth (FERC data) is a statistically significant return predictor (p<0.05). **Confidence**: MEDIUM-HIGH — physics + queue data make this the strongest pro-thesis hypothesis; the question is whether it's already priced. --- ## H3 — Mining is the late-cycle play **Claim**: Copper, nickel, REEs, uranium see demand-shock 2027-2030 as energy + chip-fab capex translate into commodity demand. Early entry (now) outperforms late entry (2028). **Mechanism**: Mining capex cycle is 7-10 yr lead time. Current supply is being built for 2030 demand. Equity markets typically anticipate by 18-30 months. **Evidence FOR**: - Copper grade decline + new-discovery scarcity (BHP, Freeport) - Uranium: Cameco/Kazatomprom backlog filling - Goldman/Macquarie commodity-shortage forecasts converging on 2027 inflection - AwarenessFund thesis specifically calls this out (Corey framing) **Evidence AGAINST**: - Mining equities have rallied repeatedly on "AI will need copper" — and given back gains - Commodity cycle dominates AI-attribution signal at minimum 2:1 noise:signal - Recycling + substitution erode the demand-shock case - 24-month back-test window is TOO SHORT to validate a late-cycle hypothesis **Falsification test**: Cannot fully test in 24mo. Proxy test: do mining equities exhibit higher beta to AI-capex announcements than to copper futures? If yes, market is starting to price the AI-link. **Confidence**: LOW-MEDIUM — strong thesis-wise, but the back-test horizon doesn't validate the late-cycle play. --- ## H4 — Chip-fab oligopoly favors NVDA/TSMC pure-play **Claim**: WITHIN the supply chain, the concentrated oligopoly (NVDA design + TSM fab) captures disproportionate margin vs equipment-makers (ASML, AMAT) who serve commoditizing fab-customer landscape. **Mechanism**: Software/design layer has highest margins. Equipment is capex-intensive with cyclical demand. Pure-play wins the rent. **Evidence FOR**: - NVDA gross margins 75%+ vs ASML's 50% vs TSM's 53% - Switching costs: CUDA moat, TSM's 3nm/2nm process moat - Equipment makers' backlog volatility (LRCX, KLAC saw 20%+ revenue swings) **Evidence AGAINST**: - Oligopoly is concentrating: ASML has near-monopoly on EUV; equipment may be MORE concentrated than design - AMD/Custom silicon (TPU, MTIA) compress NVDA's design-rent - TSM faces China-Taiwan geopolitical tail risk that equipment doesn't share - Picks-and-shovels (H1) explicitly rejects this concentration **Falsification test**: Back-test 5-name equipment basket (ASML, AMAT, LRCX, KLAC, TER) vs 5-name pure-play (NVDA, TSM, AMD, AVGO, MRVL), 24mo. H4 confirmed if pure-play Sharpe > equipment Sharpe AND drawdown < equipment. **Note**: H1 and H4 are partly contradictory — H4 advocates concentration within stack, H1 advocates diversification across stack. Back-test must run both to discriminate. **Confidence**: MEDIUM — recent history strongly supports H4, but the concentration carries tail-risk that the 24mo back-test may miss. --- ## H5 — AI-endpoint capex peaks within 18 months **Claim**: Data-center build cycle is near peak. 2026-2027 sees deceleration as hyperscalers digest capex, model-efficiency gains (DeepSeek-class) reduce demand, and depreciation hits earnings. **Mechanism**: Capex cycles in tech are typically 3-4 yr peak-to-trough. 2024 was the inflection. By 2026 hyperscalers face capex-vs-FCF tension. AI-endpoint vertical (data centers, GPU procurement) deflates first. **Evidence FOR**: - AWS/GCP/Azure capex-to-revenue ratios at multi-decade highs - Hyperscaler depreciation accelerating (8-yr → 6-yr useful life revisions visible) - DeepSeek-style efficiency improvements compress training-compute demand - Historical analog: 2001 telecom capex crash (Cisco -85%) **Evidence AGAINST**: - Inference demand (not just training) is still in early scaling - AGI/foundation model race incentivizes capex-as-moat for hyperscalers - 2026-2027 reasoning-model surge could reignite capex - China hyperscalers (Alibaba, Baidu) re-accelerating **Falsification test**: Back-test AI-endpoint basket (data center REITs, GPU vendors, server OEMs — DLR, EQIX, SMCI, ANET) for negative-quarter pattern. H5 confirmed if basket shows ≥2 sequential quarters of negative returns concurrent with hyperscaler capex-guidance cuts. **Confidence**: LOW-MEDIUM — strong counter-narrative but timing is the killer. Could be right but 12mo late. --- ## H6 — The thesis is correct but the S&P 500 prices it free **Claim**: The capex-shift thesis is real and validated, but the S&P 500 (cap-weighted) already captures it because top names ARE the AI supply chain. No fund-level alpha from an active strategy. **Mechanism**: Top-10 S&P 500 names (NVDA, MSFT, AAPL, GOOG, META, AMZN, etc.) directly are the AI stack. Owning S&P 500 IS owning the AI capex shift. Equal-weight + diversification adds tracking error without alpha. **Evidence FOR**: - S&P 500 24mo return is dominated by AI-supply-chain names - "AI ETF" products (BOTZ, ROBO, AIQ) have underperformed broad SPY - Efficient-markets hypothesis: any thesis Russell + Corey + Pyonair can articulate is already in price **Evidence AGAINST**: - S&P 500 is overweight AI-endpoint (hyperscalers) and UNDERWEIGHT mining + energy + chip-fab-equipment - Active stock-picking within the AI supply chain has alpha if the thesis is non-obvious within each vertical (e.g., transformer OEMs, behind-the-meter) - 5-yr horizon allows compound divergence from index **Falsification test**: Run the back-tested AwarenessFund portfolio (whatever variation wins) AGAINST SPY total return + risk-adjusted (Sharpe, max DD, info ratio). H6 confirmed (and fund killed) if info ratio < 0.3 after fees. H6 falsified if info ratio > 0.7. **Confidence**: MEDIUM — this is THE existential question for the fund. Back-test MUST include S&P comparison (this is capital-markets-lead's primary task). --- ## Hypothesis Compatibility / Conflict Map | Pair | Relation | |------|----------| | H1 vs H4 | **CONFLICT** — H1 says diversify; H4 says concentrate. Back-test discriminates. | | H1 vs H6 | **COMPATIBLE** — H1 can be true within the AI supply chain even if H6 says total-AI ≈ S&P. | | H2 vs H5 | **COMPATIBLE** — energy can outperform while data-center capex peaks. | | H2 vs H6 | **TENSION** — if energy is the moat and S&P is underweight energy, H6 is weakened. | | H3 vs H5 | **COMPATIBLE** — late-cycle mining + near-term endpoint peak can both hold. | | H4 vs H5 | **TENSION** — H4 winners (NVDA) suffer in H5 scenario. | | H6 vs all | H6 is the **null hypothesis** — every other H must beat it in the back-test or the fund's premise fails. | --- ## Evidence-Required-To-Validate Summary | Hypothesis | Minimum Data | Required Variations (see ingestion-variations-spec.md) | Test Duration | |-----------|-------------|--------------------------------------------------------|---------------| | H1 | OHLCV + sector tags | V1 + V8 | 24mo back-test | | H2 | Grid interconnect + utility prices | V1 + V4 + V8 | 24mo back-test + FERC overlay | | H3 | Mining production + AI-capex link | V1 + V6 + V7 | 24mo correlation only (full validation 2028+) | | H4 | Margin data + pure-play vs equipment basket | V1 + V2 (gross margins) | 24mo back-test | | H5 | Hyperscaler capex guide + data center earnings | V1 + V3 + V5 | 24mo back-test | | H6 | Portfolio vs SPY total return | V1 + V8 | 24mo back-test (THE CRITICAL ONE) | --- ## Recommendations to capital-markets-lead 1. **H6 is the existential test** — every portfolio variation MUST be compared to SPY. If we can't beat SPY risk-adjusted, the fund's premise fails. 2. **H1 vs H4 is the architectural question** — diversified picks-and-shovels OR concentrated pure-play. Back-test both as separate portfolio constructions. 3. **H2 is the conviction trade** — energy bottleneck has strongest pro-thesis evidence. Consider weighting overweight if validated. 4. **H3 is undecidable in 24mo** — recommend tracking as "watch" not "weight" until 2027 data. 5. **H5 is the risk-management overlay** — even if H1/H2 pass, an AI-endpoint capex peak could blow up positions concentrated in DLR/EQIX/SMCI. Need stop-loss / regime-detection layer. --- ## Limitations - All hypotheses are PRELIMINARY pending the 24mo back-test. - The 24mo window is too short to validate H3 (late-cycle mining). - Confidence levels are research-lead's read, not consensus across the federation (Works civ down — financial-civ depth review pending). - Pyonair PPO model arrival will likely refine these — PPO may detect features humans don't see. **Status**: SPEC v1.0 — ready for capital-markets-lead back-test design + Pyonair integration.