AI capex + strategic materials — the 2026 cluster

Cross-cutting synthesis tying four materials to the single demand driver that's rewriting multiple markets in 2025-2026: hyperscaler AI + data-centre capex. This is the clearest cross-material cluster in the atlas right now.

The cluster

Four materials whose 2025-2026 price action has been pulled upward by the same underlying force:

And a 5th material adjacent but not in the atlas:

• Specialty steel / alloys for data-centre chassis,

cooling systems, transformer cores — not individually covered but benefits directly

The demand chain

AI compute → data centres → power infrastructure + semis → 5 atlas-covered materials.

`` hyperscaler AI capex (MSFT/AMZN/GOOG/META, ~$200B/yr+) │ ├── new data centres │ ├── copper (power wiring, cooling) │ ├── uranium (SMR power offtake) │ ├── GaN power semis (fast PSU, switching) │ └── aluminum, steel, specialty alloys │ └── new silicon compute (Nvidia, AMD, TSMC) ├── silicon wafers (semi-grade) ├── germanium (fiber optics, SiGe transistors) ├── gallium (RF, power semis) └── rare earths (magnets in storage, motors in HVAC) ``

Why this cluster matters for the framework

In the atlas framework, these four materials sit in different cells despite sharing one demand driver:

• Copper: structural bull (cost-curve discipline). Chinese

concentration in refining is distributed enough that policy isn't the dominant variable.

• Uranium: structural bull (supply inelasticity from mining

lead times + Russian-enrichment sanction dynamic). Demand is real but deployment is 2028-2032 for the SMR piece.

• Silicon: bear at the polysilicon solar level

(Chinese oversupply), but semi-grade subset is effectively premium-priced long-term-contract market. AI capex benefits semi-grade silicon, not polysilicon.

• Gallium (GaN): policy-premium bull (Chinese export

controls Aug 2023 + GaN demand from data centre power semis).

Same upstream demand force, different supply-side regimes = different framework cells. The AI-capex demand force amplifies prices in 3 of the 4 cells but not the "competitive oversupply" bear cell (solar polysilicon). Semi-grade silicon, however, is premium-priced because it's a different supply regime than solar polysilicon despite being the same element.

Where the AI capex money actually lands (by material)

Copper — ~$100-300 per kW of data centre

• ~$25-50M/data centre in copper alone (for 100-500 MW

facilities)

• Grid upgrades to feed data centres = multiples more
• Copper demand from data centres globally could be

3-6% of total demand by 2028 (from ~1% today)

Uranium — ~$30-50M per SMR-year

• Each SMR-300 class reactor produces ~300 MW, needs

~30 tU/year of enriched fuel

• Microsoft + X-energy deal (2024): ~400 MW capacity
• Meta pursuing 1-4 GW nuclear capacity
• Amazon + Dominion SMR announcement (2024)
• First deployments 2028-2032 — demand is forward-looking

Silicon (semi-grade wafers) — ~$100k per wafer fab

• Each 300mm wafer = ~$200 raw material + massively higher

value-add

• Nvidia H100/B100 volumes driving wafer demand
• TSMC, Samsung, Intel fabs consuming growing polysilicon
• Wafer supply is concentrated (Shin-Etsu, SUMCO, GW,

Siltronic) — ultra-tight when AI capex accelerates

Gallium (GaN power semis) — ~$200-500 per kW of HPC power

• Data centre PSUs transitioning to GaN (higher efficiency,

smaller size, lower cooling needs)

• Fast-charging for AI rack infrastructure increasingly GaN
• Wolfspeed + Infineon are primary beneficiaries
• Silicon carbide (SiC) partially competes but GaN wins

in the 100V-600V switching space

Investment implications

For country ETFs (cross-ref to MacroLens work):

• US (SPY) — hyperscaler headquarters + semi ecosystem;

the demand source itself + Nvidia/AMD/Intel beneficiaries

• Taiwan (EWT) — TSMC foundry; every AI chip passes

through here

• Korea (EWY) — Samsung memory + foundry; Hynix HBM
• Japan (EWJ) — Shin-Etsu + SUMCO wafers; Tokyo Electron

+ Advantest equipment

• Netherlands (EWN) — ASML EUV (pure play on AI compute

demand)

• Canada (EWC) — Cameco uranium; copper miners
• Australia (EWA) — Lynas REE, copper majors (BHP Rio),

lithium (less AI-direct)

• Chile (ECH) — copper exposure for the grid build-out

Notable non-exposures: Germany (EWG) and France (EWQ) are relatively peripheral to AI capex because the hyperscaler demand is dominantly US-centred.

What could break the cluster thesis

Every thesis has kill criteria. For this one:

1. Hyperscaler capex cut — if MSFT/AMZN/GOOG/META announce 20%+ capex reductions (because AI ROI doesn't materialize, or because macro recession hits), the entire cluster weakens simultaneously. Quarterly earnings (late April / early May 2026) are the data. 2. Chip inventory build — if AI chip purchases slow but fabs keep producing, semi-grade silicon demand signal weakens first. 3. SMR deployment delays — uranium narrative is most forward-looking; delays hurt narrative more than current prices. 4. Grid build-out slowing — if US+EU fail to accelerate transmission, the copper-data-centre thesis bottlenecks. 5. GaN substitution by SiC or Si superjunction — GaN's advantage is real but narrowing; if SiC drops in price faster, gallium demand plateaus.

How atlas covers vs doesn't cover the cluster

Covered well:

• Copper demand drivers (see copper.md)
• Uranium SMR + data centre story (see uranium.md)
• Gallium GaN path (see germanium-gallium.md)
• Silicon semi-grade vs polysilicon split (see

silicon.md)

Under-covered (future additions):

• Indium + tellurium — thin-film solar / optics for

data-centre fiber

• Aluminum — structural + electrical infrastructure

for data centres

• Steel / specialty alloys — beyond atlas scope but

significant materials

Synthesis

The AI-capex cluster is the clearest thematic cross- material story in the 2026 atlas. Four materials benefit from the same underlying demand force through different paths. The framework predicts they should behave differently (different cells) and they do — but all four price directions have been up in 2025-2026, which is the thematic signal.

For a country-ETF investor, the practical takeaway is: exposure to US + Taiwan + Korea + Netherlands + Japan captures the chip-side; exposure to Canada + Australia + Chile captures the materials-side. Together that's a "long the AI physical supply chain" portfolio — already a meaningful share of the non-Chinese investable universe.

Cross-references

• docs/minerals/materials/copper.md
• docs/minerals/materials/uranium.md
• docs/minerals/materials/silicon.md
• docs/minerals/materials/germanium-gallium.md
• docs/minerals/ANALYSES/chinese-export-control-playbook.md

— the OTHER major cross-cutting theme (export controls interact with the AI capex theme via Ga/Ge + indium etc.)

• docs/thinking/2026-04-22-tech-ai-value-chain.md —

country-ETF angle on the same AI capex theme

• docs/minerals/FRAMEWORK.md — the regime framework