US interconnection queues have ballooned from 15 to 45 months; Texas data centers face 2031-2032 wait times for hundreds of MW. The constraint is transmission (poles/wires), not generation…

US grid interconnection wait times have tripled from 15 to 45 months; poles and wires—not generation—are the binding constraint. Regulatory reform (interruptible-load tariffs) and behind-th…

Chip manufacturing schedules imply 230GW power demand by 2030 vs. ~100GW grid additions. Hyperscalers face 'warm shell' shortage (chips bought but no powered racks). This gap drives behind-…

Chip manufacturing implies 230 GW AI power demand vs ~100 GW projected US grid additions. Warm shells (data center buildings) are the immediate limit. GPUs carry ~20x markup over sand-to-ch…

Hyperscalers are procuring 400MW-class gas turbines directly to power data centers, accepting higher $/MWh because energy cost is <5% of AI capex. GE, Hitachi, and new entrants like Boom Su…

Hyperscalers are buying 400 MW gas turbines directly (7-year backlog) and 38 MW trailer-mounted units. Meanwhile, 1 GW 24/7 solar+battery plants in UAE/Chile at $6/W capex can be built in 1…

Wright's Law driving 30% cost decline per cumulative doubling makes solar+battery the fastest deployable 24/7 power (12-month build). UAE 1GW solar+19GWh battery project at $6/W beats US nu…

UAE's 1 GW 24/7 solar+battery plant at $6/W capex vs $15/W for last US nuclear plant. Battery costs down 14x since 2010; sodium-ion (abundant Na vs Li) could drive another 10x drop. Solar+b…

GPU stack has 80% Nvidia markup + 2x TSMC + 2x model-provider markup = ~20x over marginal cost. AI-driven code recompilation (Fable 5, Triton) enables rapid porting to AMD/Cerebras, threate…

AI-assisted code translation (Triton, custom kernels) lets models run efficiently on non-Nvidia hardware. Inference shifting to single-rack coherence reduces need for NVLink. Nvidia's 20x m…

China's massive strategic petroleum reserve releases cushioned oil prices during Middle East conflict. Global GDP/oil spending ratio has quadrupled since 1970s, creating demand elasticity.…

Fission SMRs (Aalo, X-Energy, Natrium) need fleet orders to achieve manufacturing learning rates; first units will be expensive. Fusion has >50 venture-backed startups; Helion's direct-elec…

Construction does not get cheaper; manufacturing does (Wright's Law). SMRs (Aalo, Gradient), ocean data centers (portfolio company), and modular gas turbines all shift to factory production…

Technologies manufactured in factories (SMRs, Agent Gentic batteries, ocean data center pods, 38MW trailer turbines) capture Wright's Law learning rates; construction does not. This manufac…

No new nuclear design is cheap on unit one. France's success came from stamping identical reactors. SMRs (Aalo, X-Energy, Natrium, BWRX-300) must secure multi-unit order books from AI data…

Three fusion paths (tokamak, laser, magneto-inertial) now have venture-backed startups. Helion's direct-electricity capture avoids steam losses; CFS uses high-temp superconductors to shrink…

50+ venture-backed fusion startups exist. CFS (tokamak with HTS magnets) is most scientifically validated, targeting 600 MW demo. Helion (pulsed magneto-inertial) captures energy directly a…

Elon's vision of 100 GW/year in orbit requires 5-6 Starship launches/day (30k/year). Current FAA regulatory regime may cap at ~200 launches/year. AI demand could drive Starship cadence (amo…

Orbital data centers avoid grid, land, and permitting bottlenecks but need launch costs 4-10x lower and airline-like cadence. At 200 launches/year (20k tons), SpaceX would exceed all histor…

Elon's 100GW/year orbital compute vision needs 30,000 Starship launches/year (one every 15 minutes). FAA reliability requirements and orbital debris risks make this a 15-20 year prospect un…

Current scaling follows power-law (compute^4 for intelligence^1). Algorithmic improvements (DeepSeek, Kimi) bend curve but remain concave. Human brain runs inference at 20W vs 20kW for equi…