Article published Jul 6, 2026. Prices below use latest available snapshots.
no live data (1) — unresolved, delisted, or non-US symbols
Question: For the AI compute buildout in power-rich regions (Canadian hydro, Australian land/renewables), who wins — and what upstream physically gates it? Verdict: open (thesis is real; entry timing and operator-lease confirmation are the open variables)
Data as of ~Jul 1–2 close, sourced through the EDGAR-vs-Massive reconciliation gate. Where a number couldn't be verified, it's marked — nothing here is fabricated.
What we're asking
The binding constraint on AI compute is not chips — it's speed-to-power. You can buy GPUs; you cannot conjure a gigawatt with a grid interconnection (US queues run years). So the edge belongs to whoever already holds large power + land + cooling + a substation. Which companies have that, and what do they physically depend on?
What we found — two layers, two risk profiles
Layer 1 — Operators (own the power): the miner→neocloud pivot
Bitcoin miners spent years acquiring exactly what a GPU datacenter needs — cheap power contracts, land, substations, cooling. Converting a MW from hashing to AI compute is a ~10–40× revenue upgrade. This is the hidden operator layer behind the buildout.
| Ticker | Angle | Rev growth | Honest gate output |
|---|---|---|---|
| IREN | Canada (BC hydro) + Australia roots | +168% (EDGAR-only) | net margin disputed — not trusted; FCF unavailable |
| HUT (Hut 8) | Canadian, real HPC arm | +45% | — |
| BITF (Bitfarms) | Quebec hydro | — | +89%/mo, RSI 85 — parabolic; needs a "what happened" |
| APLD | US "build shell, lease to neocloud" | −13% (transition dip) | — |
| CORZ | CoreWeave's partner (CRWV tried to buy it) | −37% (transition) | — |
| CRWV | US pure-play neocloud (reference) | +168% | −23% net margin — burning |
| NBIS | Europe (sovereign compute) | +53% | ~breakeven; $216 needs verify (thin data) |
Every operator returned FCF unavailable — the gate refuses to fabricate it, and structurally they're all deeply FCF-negative buildout plays. High-beta, dilution-prone, running a business they've never operated. The bull case rests entirely on locking multi-year hyperscaler leases that justify the capex.
Layer 2 — Arms-dealers (sell the guts): what the blind trace found
Asked "what physically bottlenecks the buildout," a blind, prices-off derive trace converged — independently — on the grid/electrical layer:
| Ticker | What | Fundamentals | Caveat |
|---|---|---|---|
| GEV (GE Vernova) | switchgear + power transformers | rev $9.34B, +16% | 3/3 conviction but RSI 74, +21%/mo, 1.9% op margin — extended, poor entry |
| ETN (Eaton) | switchgear + datacenter PDUs | rev $7.45B, +17% | margins compressing; calmer entry |
| NEE (NextEra) | controls renewable interconnection queues | rev $6.10B, +2% | slow-growth utility |
| ENS (EnerSys) | DC power / batteries / UPS for dense racks | — | the non-obvious one |
| WNC (Wabash Natl) | modular datacenter containers | — | not in our data yet — needs a fetch |
These are established, real-margin, durable — the trend-hold-appropriate expression of the same buildout.
Verdict + reasoning (open)
- Durable expression = the arms-dealers, not the operators. But GEV is a great thesis at a bad price (extended, thin margins — the trace flagged the deceleration itself). ETN and ENS are the calmer entries.
- The genuinely fresh finds: ENS (batteries/UPS) and modular containers (WNC). Nobody frames AI datacenters as a battery story; the physics does. WNC isn't in our data — worth adding.
- The operators (IREN/HUT/BITF/…) are the high-beta optionality — own only if you want the leverage and can stomach the volatility (whole cohort −20–32% on the month). The confirming catalyst is a signed hyperscaler lease. BITF's +89% parabola is a separate one-off to investigate.
- Why "open": the thesis is physically sound, but (a) entry timing is poor on the hottest name (GEV) and (b) the operator leg is unconfirmed until leases are signed. Not a buy-list — a map.
Provenance / method note
Names surfaced by a blind supply-chain trace with prices withheld, and an operator layer added by a separate grounded read. Every shipped fundamental was reconciled against a second source; anything unverifiable is marked as such rather than estimated.