AI capex sprint → power-supply gap → grid-component + materials bottleneck → nuclear/copper/transformer beneficiary cascade

Hyperscaler capex hit $725B-$830B for 2026 (+79% YoY). Power, not GPUs, is now the binding constraint — most announced gigawatts aren't being built because the grid components (transformers, turbines) are themselves on backlog. Nuclear emerges as the consensus baseload answer; copper deficit and US-China critical-mineral exposure compound. Beneficiaries split into three chains: nuclear (CCJ, CEG, BWXT), copper miners (FCX, SCCO), grid infra (transformer / cable makers).

• Step 2: If hyperscaler earnings beat without power-supply caveats and announced gigawatts come online on schedule, the power-as-binding-constraint framing is wrong.
• Step 3: If transformer / turbine backlogs clear within 12 months (suggests OEM capacity ramps faster than expected), grid-infra beneficiaries lose their scarcity premium.
• Step 4: If copper mine output growth surprises to the upside in 2026 (>5% YoY), the deficit thesis weakens.
• **Step 1 / Step 5 — real overbuild risk (added 2026-06-29).** michael-cembalest in 2026-06-23-podcast-columbia-energy-exchange-michael-cembalest-does-the-math-on-the-energy "takes the under" on the McKinsey / Bain / BCG datacenter-power demand projections and flags genuine overbuild risk: "there's a chance that we end up with some of this stuff that gets overbuilt." Falsifier: if datacenter-power demand undershoots the consultant projections, the power/component scarcity premium deflates — note the OEMs themselves under-expanding capacity only ~**10–20%** (Mitsubishi/Siemens/GE Vernova) is partly an attempt to avoid building into that bust, so an OEM capacity surge *above* that 10–20% range would be the tell that even the suppliers no longer fear an air-pocket.
• **Step 5(a) — Western nuclear cost-gap / execution risk (added 2026-06-29).** michael-cembalest in 2026-06-23-podcast-columbia-energy-exchange-michael-cembalest-does-the-math-on-the-energy: the West builds nuclear at ~**4–5x** the cost of China / Korea / India and points to "Flamanville, Olkiluoto, Hinkley, Vogtle... white elephants." The demand is real, but **Western execution economics are the weak link** in the nuclear leg — if new Western nuclear keeps coming in 4–5x over comparable-nation cost, the nuclear cascade's beneficiaries (CEG / CCJ / SMR designers) underperform the demand signal because the supply can't be built at a returns-positive cost. Cross-link nuclear-baseload-for-ai-data-centers.

• **⚠ Coal-to-gas conversion is NOT the fast dispatchable bridge (2026-07-10, evidence *against* Step 6).** 2026-06-19-feed-construction-physics-converting-coal-plants-to-natural-gas is the first source in this wiki to examine the coal-to-gas conversion path directly, and Brian Potter's conclusion cuts against the bridge-fuel framing: "It doesn't seem likely that we'll see many more of these coal-to-gas conversions." His reasons are structural, not cyclical: "The most obvious candidates for conversion — smaller, older plants that might be useful for peaking — have probably already been converted," and "as grid-scale batteries change the economic logic of peaking, even new gas plants are looking less attractive than they used to; I can only imagine that a less-efficient converted coal plant is even less compelling." He concedes only a hedged AI carve-out: "the enormous demand for power caused by the AI boom might have some effect... But overall, I suspect that the heyday of coal-to-gas conversions is behind us."
• **What this does and does not touch.** It does *not* weaken Steps 1-3 (capex scale; power as binding constraint; grid-component backlogs) — Potter independently confirms the turbine backlog. It weakens the *specific* claim that converted coal plants are a near-term source of dispatchable capacity, and it introduces **grid-scale batteries** as a competing peaker technology this mechanism does not currently model. Step 6's "gas is the bridge fuel" stays `partial`; the *conversion* sub-path should be treated as closed until contradicted.
• Note: the source **names no turbine maker or utility**, so it cannot be read as a ticker call in either direction. It is a falsifier, not a trade.
• **⚠ CGEP: data centers are NOT the main driver of retail electricity prices — and queued demand is heavily phantom (2026-07-13).** doug-arent and robin-millican (Columbia CGEP) in 2026-06-30-podcast-columbia-energy-exchange-doug-arent-and-robin-millican-on-what-s-really present the LBNL/Brattle-based decomposition: "load growth does not stand out as a national driver of price increases... what really is driving prices dominantly, it's fuel price [gas sets the marginal price]... The second driver is in fact increased distribution costs, transmission costs" (Arent). This does **not** falsify this chain's supply-gap steps (2–3) — Millican independently confirms the physical mismatch ("developers... looking to build data centers in one to two years... transmission lines take 10 years to energize") — but it **refines the retail-price narrative**: bills are rising mostly for non-DC reasons (see electricity-price-drivers-decomposition). Sharper for the demand leg: Arent's phantom-load data — "expected demand is down by a third" once financial commitment is required (TX ~$55k/MW queue fee), only ~75% of queued generation gets built, and the 1,200+ proposed DCs (~100–300 GW vs a 1,400 GW US base) "may, in fact, not turn out to be realized" (see phantom-data-center-load) — corroborates Cembalest's "take the under" falsifier below with a *mechanism* for the overstatement. Also adds ~260 GW of GETs latent capacity (avg grid utilization ~40%) as a partial buildout-deferral valve, same direction as the Siemens Neuedra tension (grid-enhancing-technologies-latent-capacity). Recorded as a refinement + demand-leg caution, not silently reconciled.
• Sacks pushes back on the 'dot-com 2.0' framing: 'there's no dark GPUs today' — demand is real, not over-built. Chain remains a bull thesis.
• **Training cost commoditization (Chamath, 8090)**: chamath-palihapitiya in 2026-05-29-podcast-all-in-podcast-anthropic-s-digital-god-pope-vs-ai-job-loss: "There was an economic and capital moat to training that is going away...domain specific architectures at the silicon layer...Elon was like, we've rewritten the entire training complex in C and it's an order of magnitude increase...why would we stick to the $10 billion training runs when we can have the $10 million training runs?" If training costs collapse from $10B to $10M per run (domain-specific silicon + compiler efficiency), the *size* of step-1 capex in this chain is at risk — future builds may require less compute than the current $725-830B capex wave implies. Falsification test: track whether training compute per FLOP actually falls 10-100x by 2027 or whether scale remains the dominant competence signal.
• **Capex not a bubble — FCF still covers in aggregate (Chisholm, Fidelity data)**: denise-chisholm in 2026-05-29-podcast-the-compound-and-friends-what-if-it-s-still-early-with-denise-chisholm: at the 2000 bubble peak, corporate America spent 3.5-4x FCF on capex; today we're still under 1x in aggregate even including hyperscalers. The FCF-coverage signal argues current capex is sustainable rather than bubble-driven; corroborates Sacks and supports chain durability.
• Permitting reform (US permitting/Inflation Reduction Act follow-ons) could partly relieve the bottleneck.
• Hyperscalers' balance-sheet capacity to keep signing 2x-spot energy contracts is finite — at some point pricing power flips to the energy supplier rather than to hyperscaler beneficiaries downstream.

• The bottleneck is upstream of the GPU. INTC/TSMC/NVDA can be in supply, but if power isn't, the token-factory revenue doesn't materialize.
• Hyperscalers giving up free cash flow / buybacks to fund this cycle (Amazon FCF -97%, Google/MSFT/Meta -8 to -12%). Long-term valuation reframing follows (Friedberg: hyperscalers will look like 'big bulky industrials' in 5 years).
• Friedberg's actionable rule: 'follow the dollars going out of hyperscalers — those companies are underpriced.' Picks-and-shovels logic extends beyond semicap.
• Reinforces the orbital-DC chain (terrestrial-power-flat-to-orbital-dc-arbitrage) — power gap is real and structural, not a market head-fake.
• **AI's marginal cost ≫ 0 is the structural difference vs. the internet (the "money glitch" is gone).** chamath-palihapitiya in 2026-06-13-podcast-all-in-podcast-anthropic-s-fable-backlash-nationalizing-ai: "The marginal cost of production for a new internet user was effectively zero… AI is completely different. There is a real cost for every marginal user — everyone you stand up is taxing a GPU, needs electrons, needs memory." This is the unit-economics root of why the capex cascade is *self-reinforcing* (more users → more compute/power, not free scale) and why ROI scrutiny (mega-issuance-peak-to-ai-capex-derate, csp-capex-cycle-peak-or-sustained) is the right counter-watch — the picks-and-shovels (power, copper, memory) get paid on every marginal token regardless of which model layer wins.
• From 2026-06-04-to-boldly-go-the-case-for-space-datacenters: SemiAnalysis (June 3) quantifies the terrestrial power constraint rigorously: space datacenter TCO $10.91/hr/GPU vs $2.49 terrestrial in 2026 (4.4× cost gap); parity ~2040 base case, early 2030s only if terrestrial power **peaks in 2028** (Musk scenario). The forcing function is terrestrial power exhaustion, not free solar or cooling. Five-layer terrestrial power supply framework: grid-connected → converted capacity (bitcoin mining ~8–10 GW) → behind-the-meter generation (26 GW by 2030) → industrial expansion → semiconductor (universal constraint: "chip manufacturing will be the global constraint before we even worry about supply"). AI-related DRAM demand consumes 70% of total DRAM wafer capacity by 2027. **This analysis confirms the power constraint is the correct binding variable for the 2026–2030 horizon — space datacenters are a 2040 solution, not a 2026–2030 escape valve.**