Themes

Priority sectors for research and capital. Themes organize watchlists and risk concentration — they are not automatic buy lists.

The four core themes are not four independent bets. They are linked by a massive increase in the economic value of computation, electricity, and autonomous machines:

AI requires compute → compute requires power → AI + sensors + cheap compute enables robotics → robotics + AI transform defence → defence, reshoring, and data centers require more industrial capacity and electricity → electricity constraints drive grid, nuclear, gas, and power-equipment investment.

That is a capital-investment supercycle, not merely an “AI bull market.” Leadership migrates down the supply chain. A 10-year structural runway does not mean a 10-year bull market in today’s winners. Theme caps still do not diversify the shared AI-capex factor (mandate rule 10).

The standing research question: where is the next bottleneck forming that the market is not yet obsessing over?

Clocks (working map)

Revisit with the quarterly theme pass. These are duration maps, not entry signals — IEA, NATO, and hyperscaler capex headlines are consensus inputs (see mandate: citing them is not evidence of earliness).

Overlapping waves rather than a relay: compute scarcity (now) → physical infrastructure bottlenecks (power, transformers, cooling, optics, grid) → inference and monetization → physical AI → defence rebuild and electricity buildout through the mid-2030s.

Expect 10–15% corrections often, several 20–30% thematic drawdowns, 30–50% hits in quality growth names, and 50–80% collapses in speculative ones. Those episodes are part of the strategy if research is done in advance; some will be real fundamental deterioration. The dossier’s warning / reduce / invalidation framework is how we tell the difference.

Core themes

1. AI infrastructure

Why: Hyperscaler and enterprise AI demand drives compute, networking, data centers, and the supply chain behind them.

Look for: CapEx cycles, GPU/accelerator supply, data-center development, optical/networking, cooling, semiconductor equipment and packaging — evidence of spend and bottleneck relief before multiples fully expand.

AI memory & storage (sub-focus, not a separate theme): HBM, server DRAM, enterprise NAND/SSD, and emerging tiers (e.g. HBF) sit inside this theme because they are binding AI compute bottlenecks. Tag primary theme as AI infrastructure. Size with extra cycle discipline (see mandate) — memory often looks cheapest near peak EPS.

Avoid as default: Pure narrative names with no path to earnings or cash flow, chasing vertical moves after consensus is saturated.

Exit signals (theme level): hyperscaler capex guidance revisions turning negative; supplier orders / book-to-bill dropping below 1 (e.g. VRT, NVT); GPU-cloud utilization and pricing weakening (neocloud disclosures); HBM contract pricing rolling over. Any two deteriorating together means the buildout phase is ending — picks-and-shovels suppliers were the epicenter of the 2000–02 unwind (Cisco, Corning, JDSU fell 80–95%), not the shelter.

2. Energy

Why: Power is a binding constraint on AI and electrification. Energy is both a theme and a cross-cutting input to AI infra.

Look for: Generation and capacity additions, grid equipment, interconnection progress, nuclear and related fuel cycles, natural gas where it bridges demand, commodities tightly linked to the buildout (e.g. copper — also in Adjacent).

Avoid as default: Treating every energy ticker as an “AI power” proxy without asset-level or contract-level evidence. Note: the fundamental buildout being early does not make the pricing early — data-center power is a heavily discovered trade; apply the crowding checklist (mandate rule 6) before every entry.

Exit signals (theme level): transformer and switchgear lead times normalizing; grid-equipment order intake rolling over; data-center interconnection queues shrinking; merchant power / PPA pricing softening.

3. Robotics and AI

Why: Automation and embodied AI are a multi-year adoption curve; humanoid and industrial robotics may reprice as pilots become orders.

Look for: OEM adoption, factory CapEx, unit economics, backlog, real deployments vs demos.

Avoid as default: Pricing science-fiction timelines as near-term revenue; full size into hype spikes without commercial proof. The component-layer shortcut ("30 failed OEMs still need sensors") is not automatic: lidar volumes exploded while Velodyne, Quanergy, and Luminar destroyed capital because component ASPs deflated faster than volumes grew (same pattern as solar cells and LEDs). Prefer qualified, profitable design-in suppliers over pre-profit optionality; size pre-profit names as an explicitly speculative sleeve.

4. Defence

Why: Structural budget and capability shifts (drones, autonomy, munitions, C4ISR, air defence) can create multi-year demand independent of consumer cycles.

Look for: Contract awards, program ramps, export approvals, budget line visibility, capacity expansion.

Avoid as default: Ignoring political/program cancellation risk; illiquid names that cannot be exited if narratives reverse. Budgets are not revenue: NATO/EU headline commitments increasingly route to European suppliers (Rheinmetall, Thales, Saab, BAE) via procurement nationalism — do not book European budget headlines as tailwinds for US small-caps. Underwrite US names on their own bookings, backlog conversion, and margin trajectory.

Bottleneck search

Do not hunt small caps because they are small. Hunt companies that control a bottleneck whose economics are changing faster than investor perception.

The pattern that can produce a 3× (rather than a market-like 25%) looks like: obscure ~$2–10B company → critical component in a rapidly scaling system → customer qualification creates switching costs → orders inflect → fixed-cost leverage → estimates rise repeatedly → coverage expands → the multiple rerates.

The search is not limited to names labeled “AI.” Hunting grounds include power conversion and grid hardware, advanced cooling, connectors and cabling, optical components and test, specialty semiconductor IP, packaging and testing, precision motion, sensors, robotics components, drone propulsion, defence electronics, industrial automation, and specialty materials. Liquidity (mandate rule 7) still binds; illiquid microcaps need an explicit size cap.

Adjacent watch (not core allocation by default)

Promote to core only with repeated evidence and mandate fit:

Exploration

Other sectors are allowed when:

1. The opportunity is asymmetric on evidence, not vibes.
2. It does not blow theme concentration without conscious choice.
3. It is logged under theme Other (or a new named theme if it becomes recurring).

Theme hygiene

• Revisit theme definitions quarterly.
• Tag every instrument and signal with a primary theme (and optional secondary).
• Track performance and error patterns by theme in reviews — some themes will deserve more or less capital over time.
• Each quarterly pass should name the next under-obsessed bottleneck, not only “which AI stock to buy.”
• Themes are labels, not diversification. Most of the core universe loads on one factor — hyperscaler capex sentiment — so theme caps alone do not control risk. Measure cross-holding correlation and treat the combined AI-capex complex as one risk (mandate rule 10). The shared industrial transformation above explains the correlation; it does not excuse ignoring it.