POWER
AI compute demand structurally outruns power supply.
78/100 conviction▲ +4 strengtheningUpdated 8 Jul 2026
For two years the constraint on AI was compute — whoever had the GPUs won. That era is ending. Capital is now colliding with physics: you can order the chips, you can’t order the power. Hyperscaler capex is scaling faster than grids can deliver interconnection, and the binding constraint is migrating off silicon onto the unglamorous hardware that energises a data centre — transformers, switchgear, substations.
What’s driving it
- Microsoft’s commercial remaining performance obligations (contracted backlog) roughly doubled to ~$625B, with management saying demand “continues to exceed available supply” and framing Azure’s growth in energized megawatts and interconnection/equipment lead times (Microsoft FY26 Q2).
- Average US large-power-transformer lead times sit near 128 weeks — up from 24–30 months pre-2020, with some specialised units quoted out to four years (Wood Mackenzie, Q2 2025 survey).
- The four largest hyperscalers are guiding to on the order of ~$725B combined capex in 2026 (analyst aggregation of guidance) — money that cannot buy grid capacity that does not yet exist.
The bear case
The same capex is cash-funded by trillion-dollar balance sheets, and supply
is responding — capacity expansions and Chinese transformer imports as easing vectors. Hyperscalers can also sidestep the grid queue via on-site generation (bring-your-own-power).
What would prove us wrong: transformer lead times contracting below ~80 weeks within 12 months; a material cut to hyperscaler capex guidance; or the power-constraint story failing to reappear in a second independent quarter.
Where the exposure sits
Companies described by what they do in the affected layer — not securities we’re selecting, and nothing here is a buy or sell.
- GE Vernova — large power transformers and grid equipment
- Eaton — electrical switchgear and power management
- Siemens Energy — grid technology and transformers; the order backlog is the tell
Sources: Microsoft FY26 Q2 (Investor Relations); Wood Mackenzie Q2 2025 transformer lead-time survey (pv-magazine, POWER Magazine); 2026 hyperscaler capex aggregation (Tom’s Hardware, CNBC).