Gobi X: Creating more energy for AI, not taking it from society

The article argues AI’s binding constraint has shifted from chips to power—US datacenter demand is projected to rise from 31GW in 2025 to 66GW by 2027, while the IEA estimates datacenters’ electricity use at ~1.5% of world power in 2024, climbing to ~3% by 2030. Envision’s Mission Gobi targets 5GW of “green AI” computing in deserts by 2030 by building dedicated renewable energy systems where electricity is abundant, aiming to avoid drawing down supply meant for households and public services. Overall, the piece is constructive on long-term infrastructure solutions, but highlights near-term system stress from rising rack power densities (up from ~5kW toward ~200kW) and grid/transformer constraints.

Analysis

The market is still pricing AI as a chip scarcity story, but the more durable monetization bottleneck is power delivery. That shifts alpha from semiconductor beta toward grid-adjacent names with backlog visibility and pricing power: electrical equipment, transformers, switchgear, EPCs, and utilities with scarce interconnect rights. The second-order loser set is the “build it and they will lease it” data-center complex: if megawatts lag racks, revenue recognition slips even when demand is real, which can compress near-term multiples despite strong long-dated demand.

The key timing issue is that the bottleneck is not solved in one earnings season. In the next 1-3 months, watch utility capex guides, interconnection queue reform, and transformer/switchgear lead times; those are the catalysts that can re-rate winners. Over 6-18 months, the bigger winner is likely energy-native compute clustered where firm power is cheapest, while projects that depend on retrofitting urban grids face longer delays and higher financing costs.

Contrarian view: consensus is overly fixated on “more renewables” and underappreciates that the scarce asset is firm megawatts, not just clean megawatts. If AI load growth forces regulators to protect household rates, the upside to utilities may be capped, but the real scarcity rent should accrue to equipment suppliers and owners of interconnect rights. Thesis is falsified if grid delivery improves faster than expected or if hyperscalers can show materially higher on-time capacity conversion than the current pipeline implies.

Key Decisions for Investors

• Overweight electrical infrastructure beneficiaries over AI hardware for the next 3-6 months: long ETN/GEV or PWR versus short SMH on the thesis that power bottlenecks delay GPU monetization more than they delay equipment orders.
• Add a relative-value long in data-center power/thermal infrastructure names on dips after earnings. Target 6-12 months; risk/reward improves if backlog conversion and lead-time commentary stay tight.
• Use a pair trade: long grid-capex winners, short data-center REIT exposure (e.g., DLR/EQIX) into any rally, because megawatt scarcity can push out lease commencements and cap near-term NOI growth.