A 5-vertical, picks-and-shovels equity strategy targeting the infrastructure of the AI capex super-cycle.
Trillions of dollars of capex are migrating from human knowledge work into AI infrastructure today, and we believe a second wave will migrate from AI infrastructure into robotic physical labor on a roughly 5-year horizon. We do not bet on model labs, application-layer AI, or hyperscaler concentration. We bet on the infrastructure stack the model-lab AI race requires to continue: mining, energy, chip-fab equipment, AI-endpoint physical buildout, and SPACE-infra optionality. Confidence: HIGH on direction; MEDIUM-LOW on point estimates and timing.
A short audio walkthrough of where the thesis sits, what it is and isn't, and why the substrate-discipline layer matters as much as the picks-and-shovels selection.
The 5 verticals are sequentially gated: chip-fab equipment must exist before chips exist; chips before AI endpoints are useful; AI endpoints must be powered and made of materials; SPACE-infra is the deferred-but-inevitable safety valve for terrestrial physical limits.
| Vertical | The bottleneck | Cycle | Why-now |
|---|---|---|---|
| Mining | Discovery-to-pour lead time 8–12 yr; copper, uranium, rare earths supply inelastic at AI-demand timescales | 8–12 yr (late-cycle) | Capex at multi-year highs but well below 2012 super-cycle peak — room to run. USGS Annual Reports; MP Materials 10-K FY2024. |
| Energy | Grid build-out cannot match AI compute demand; transformer lead times stretched ~50 weeks (2019) → 120+ weeks (2024) | 5–10 yr | Microsoft signed PPA for Three Mile Island restart (Sep 2024 — first US nuclear-restart-for-hyperscaler). Dominion 2024 IRP forecasts 85% Virginia load growth by 2039. |
| Chip-fabs | EUV monopoly (ASML), 4-firm WFE oligopoly (~85% market share), CoWoS advanced-packaging sold out through 2026 | 2–3 yr | ASML EUR 28.3B revenue 2024 with ~18-month tool lead times. CHIPS Act $27.6B finalized in 2024. |
| AI-Endpoint | Data-center rack power-density jumped from ~10 kW (2020) to 80+ kW (2024 AI racks); liquid cooling is the new bottleneck | <1 yr | Vertiv FY2024 record year on liquid cooling. Equinix $8.7B FY2024 revenue. Arista 35% revenue concentration in two hyperscalers. |
| SPACE-Infra | Terrestrial grid + cooling + latency budgets push R&D toward orbital compute; SpaceX drove launch cost from ~$10K/kg to ~$2–3K/kg | 5–10 yr (with 20+ yr Dyson-swarm optionality) | SpaceX 134 successful Falcon launches in 2024 vs <40 industry-total a decade prior. Lonestar Holdings + Ramon Space hold publicly-announced DoD/NASA orbital-compute contracts. |
If model labs commoditize AI margins (price wars, open-weight catch-up, DeepSeek-class efficiency gains), picks-and-shovels actually benefits — more inference deployments = more endpoints = more power = more materials. We bet on AI deployment, not AI margin capture.
It is entirely possible that the S&P 500 already prices the AI capex shift and that no fund-level alpha exists. This is hypothesis H6, and it is the existential test for the fund. The back-test is designed to disprove it, not to confirm what we already believe. Every back-test variation is benchmarked against a 4-tier ladder: S&P 500 TR · equal-weight S&P · NASDAQ-100 · custom sector basket. A variation is "interesting" only if it beats all four on risk-adjusted basis.
Below are the most-defensible representative tickers per vertical, with the explicit counter-argument that could falsify each vertical's thesis. The full 73-name candidate universe lives in ticker-universe-spec.md (download below).
AI hardware and grid build-out is materials-intensive — copper for interconnect and grid, nickel for batteries, uranium for clean baseload, rare earths for motors and magnets, lithium for storage, silver for PV. Mining capex cycles are 8–12 years from discovery to first pour, so supply elasticity is near-zero on the timescales AI demand needs.
US grid needs ~+1 terawatt of new generation over a 10-year horizon to support AI data-center demand. Picks-and-shovels = utility owners of generation/transmission near hyperscaler sites + independent power producers with nuclear/gas baseload + grid-equipment makers (transformers, switchgear, HVDC).
The picks-and-shovels of AI compute are NOT the model labs and NOT even the chip designers — they are the equipment that makes the chips (lithography, deposition, etch, metrology) and the specialty materials (photoresists, gases, wafers, advanced packaging substrates). A 4-company oligopoly at the high end (ASML, AMAT, LRCX, KLAC) plus a Japanese duopoly in materials.
AI-endpoint = the physical buildings + cooling + networking + power distribution where AI compute happens. Explicitly excludes model labs (no Anthropic-proxy, no OpenAI-IPO speculation, no application-layer AI) and hyperscalers (MSFT/GOOGL/AMZN/META are mixed-margin businesses).
Terrestrial AI-endpoints are running into hard physical ceilings — grid power, water for cooling, fiber latency budgets, real-estate adjacency to demand. Picks-and-shovels here is commercial launch + reusability, satellite manufacturers + comms-backhaul, defense primes holding DoD/SDA/NASA space-systems contracts (mixed-beta caveat), rad-hardened space-grade compute, and space-solar / Dyson-swarm frontier (20+ year optionality only).
This is the section that explains why ACG's involvement should matter to a sophisticated LP. Most fund-management organizations cannot tell you, in detail, why they chose the data sources they chose, what hypotheses they ruled out, what their false-positive rate is on LLM-extracted numbers, how they audit their own work, or what would make them publicly retract a claim.
legacy_pre_amendment flag. Structural, not aspirational. v1.1 schema shipped 2026-05-14.The LP-readable claim. The same vendor-substrate-discipline-scorecard ACG publishes publicly is the rubric we have been asked to apply to ourselves on this fund's back-test work. We will publish that self-assessment when the back-test results ship. If it is ugly, we will publish that too. The discipline of publishing the ugly self-grade is, itself, the alpha.
We name these openly because diligence-grade LPs will discover them anyway. We would rather be the source than be caught.
| Gap | Substance | Mitigation / path |
|---|---|---|
| ACG has no native back-test engine | yfinance-class CLI ingester, generic ingestion plumbing, strong LLM-extraction discipline — but zero equity portfolio back-test capability in the repo today (audited 2026-05-18). | Pyonair owns this. The division of labor IS the design. If Pyonair PPO code does not arrive in a reasonable timeframe, ACG (mind-lead) writes a minimal vectorized harness as fallback. |
| No transaction-cost model beyond a simple linear-impact stub | Critical for honest Sharpe / info ratio numbers at fund-AUM scale. | A 5/15/+2/+10 bps cost model is specified; calibrate from realized spreads in Polygon paid data if upgraded. |
| No survivorship-bias-corrected ticker universe | All current scraping is "what's listed today." Historical delistings are missing. | Manual delisting enumeration for the 24-month window. CRSP-like dataset ($500+/mo academic tier) is v2 upgrade path. |
| Works (federation's financial-civ) is currently DOWN | Works (Kimi K2.6) is the sister-civ with deepest financial-domain depth. Factor-construction critique is unavailable until restart. | Hengshi (Qwen) is healthy and can serve as cross-grading peer for research outputs. |
| TG / blog instrumentation lag | Subscriber open/click instrumentation for ai-civ.com blog posts is not yet wired. Cannot measure LP-funnel response to publicly-shipped substrate-discipline IP today. | Slot-4 instrumentation target carried from 2026-05-17; unmet as of this document's date. |
| Pyonair PPO code has not yet arrived | All RL/PPO research on the ACG side is preparatory; integration will happen when code lands. | Asynchronous timeline. Independent ACG work (ingestion + protocol + universe) is ungated and proceeds. |
| Time-zone alignment with Pyonair team unknown | Async coordination assumed. | TGIM substrate (cross-civ task platform) is the standing wire if cadence becomes a friction point. |
| No live brokerage / execution layer | Back-test analytics only. Live execution is fund-back-end work outside ACG's scope. | Russell + Pyonair territory. |
| No risk-management / compliance / regulatory framework | When fund formally launches, this is a hard external dependency. | Counsel + compliance vendor to be retained at fund-formation stage. Out of scope for this thesis-overview. |
| 24-month back-test window is too short to validate H3 (late-cycle mining) | Mining capex cycle is 8–12 years; equity markets typically anticipate by 18-30 months. The thesis we believe most strongly in (H3 supports it) is the one the back-test can least validate. | Proxy test in v1 (mining-equity-beta-to-AI-capex-announcements). Full validation requires 2028+ data. |
| Statistical power at 24 monthly observations is honestly low | t-stats and bootstrap CIs are reported as substrate for partner discussion, NOT as inference-grade evidence. | Treat as descriptive, not inferential. Honest framing in every report. Longer back-test window (5y+) is Polygon paid upgrade path. |
Raw source documents. Every claim in the thesis traces back to one of these. Markdown format.
Read in full. Then read again.
This document is a DRAFT THESIS OVERVIEW for pre-launch partner and prospective-LP discussion only. It is NOT a solicitation to invest, NOT a private placement memorandum, NOT investment advice, and NOT an offer of securities. No fund vehicle yet exists. No price targets, return projections, or performance forecasts appear in this document by design.
Past performance does not predict future results. Back-tests are hypothetical, derived from historical price data, and do not reflect actual trading, real-world transaction costs, taxes, or fund-management fees. Survivorship bias, selection bias, and look-ahead bias may be material despite the controls described. Forward-looking statements derived from third-party sources carry their authors' biases and should be independently verified before any investment decision.
Investing in equities (including any future Awareness Fund vehicle) involves risk, including risk of total loss. Concentrated thematic strategies carry higher volatility than diversified passive index strategies. The capex-shift thesis described in this document may be wrong. The information in this document is current as of 2026-05-18 and is subject to revision without notice.
All quantitative claims are confidence-tagged HIGH/MEDIUM/LOW per the rubric in the Thesis Substantiation document. LPs and partners must conduct their own due diligence and consult independent legal, tax, and investment counsel before making any commitment.