Own-the-Bottleneck — Own-the-Bottleneck

Theme thesis — tracked as a cross-cutting view, no single ticker.

Investment Read as of 2026-06-06

The Read

Own-the-Bottleneck remains a SOLID investment. The increasing scale of global grid buildout and hyperscaler activity confirm the core thesis that power and infrastructure constraints are the primary limiting factor for AI growth. (P=0.81)

Bull case

• AI compute demands massive power, making grid buildout a primary constraint. Global grid capex is projected to exceed $650 billion in 2026 source.
• Major players are building out capacity directly. Grupo Fotones launching a €1.15bn data center campus shows concrete, large-scale capital deployment source.
• The underlying compute demand remains high, evidenced by ongoing discussions about running advanced models locally, suggesting efficiency gains are not yet solving the power crunch source.

Bear case / what breaks it

• A sudden, major efficiency breakthrough in inference (e.g., a new chip architecture) that drastically lowers the power needed per computation could flatten the TWh demand curve.
• If multiple bottlenecks—like transformers, HALEU, and copper—see simultaneous, unexpected supply increases, the pricing power premium shrinks.
• If geopolitical stability improves significantly, allowing for rapid, non-constrained deployment of critical inputs (e.g., rare earths, specialized components), the "chokehold" premium fades.

What the latest signal says

The signals confirm intense, localized capital deployment and regulatory focus on power. The Rystad estimate of $650B+ in grid capex and the specific data center announcements (Grupo Fotones, Claro Argentina) validate the thesis that massive, tangible investment is flowing into the physical infrastructure layer.

Posterior history

Thesis detail

Core thesis

AI compute is a demand shock cascading down a power supply chain that can't keep

up. The durable money isn't the headline names — it's whoever controls the

scarce inputs and conversion steps everyone depends on. Own the bottleneck,

not the company behind it. Processors beat miners.

Pillars (with priors)

1. Power is the binding constraint on AI (DC load 183 TWh '24 → 325–580 '28) · P = 0.85

2. Value concentrates at the conversion step (HALEU not uranium; refining not mining) · P = 0.80

3. 6 of 11 top bottlenecks are single-country-controlled → supply security = pricing power · P = 0.75

4. Energy *addition*, not transition — every source scales at once · P = 0.65

Expected news (the prior)

• Steady stream of hyperscaler power deals + rising data-center load forecasts
• Tight bottleneck data (transformer lead times, HALEU output, Dy/copper)
• Recurring export-control / tariff actions on critical inputs

Thesis-breaking triggers (→ set relevant P near 0)

• ☐ Durable inference-efficiency step-change flattens AI power demand
• ☐ Multiple bottlenecks ease at once (new supply across transformers/HALEU/REE)
• ☐ Alliance supply (FORGE/AUKUS/DOD) matures early, breaking single-country chokeholds

Leading vs lagging indicators

• Leading: capex reallocation, regulatory drafts, cost-curve crossings, order pipelines
• Lagging: commodity spot prices, company revenue/margins

Key metrics

• DC TWh trajectory · transformer lead times (128–144 wk) · HALEU output · Dy price · hyperscaler capex path

Cross-arena sensors

B1 (demand/efficiency), B2 (the constraint), B7 (supply security), B4 (magnets).

The biggest signals are *cross-bucket connections*.

Component theses

CLF (GOES) · LEU (HALEU) · CEG (nuclear) · MP (rare earths) · FCX/SCCO (copper)

Posterior log

• 2026-05-26 · P 0.76→0.81 ↑ · UP · · https://www.bloomberg.com/news/articles/2026-05-27/sqm-boosts-lithium-guidance-as-earnings-top-estimates
• {{date}} · created from AI-Energy research · — · AI-Energy-Thesis-Scaffold