Invest in heavy equipment manufacturers like General Electric (GE) to capitalize on massive multi-year backlogs for modular natural gas turbines, which serve as the primary interim power source for AI data centers through the late 2020s.

Allocate to utility-scale solar and battery storage providers such as Tesla (TSLA), which offer the fastest 12-month deployment timeline and lowest cost structure to bypass multi-year electrical grid queues.

Target power grid infrastructure and utility operators through the Utilities Select Sector SPDR Fund (XLU), focusing on grid-orchestration and battery technologies that reduce connection wait times from seven years down to 12 to 18 months.

Focus nuclear investments on conventional nuclear reactor restarts over the next three to five years, rather than early-stage Small Modular Reactors (SMRs) that face commercial delays until the 2030s.

Maintain a cautious view on NVIDIA (NVDA) over the next one to two years, as severe physical power bottlenecks and emerging compiler tools risk eroding its hardware pricing power and software moat.

NVIDIA Corporation (NVDA)

• AI power bottleneck: GPU manufacturing schedules through 2030 (led by NVIDIA, AMD, and Cerebras) represent roughly 200 to 230 gigawatts (GW) of power demand, while the US grid is only projected to add around 100 GW of capacity in that timeframe.
  • Hyperscalers face "warm shell" constraints where expensive chips are sitting idle in warehouses due to lack of powered data center space.
  • Total data center power draw is nearly double the chip draw when accounting for cooling and supporting IT hardware.
• CUDA software moat: NVIDIA's primary competitive advantage historically has been its CUDA programming layer rather than purely hardware superiority.
  • New AI compiler tools (such as automated kernel generators and startups like Lemurian) are increasingly allowing developers to run code efficiently on competing chips like AMD, potentially eroding NVIDIA's software moat over the next 1–2 years.
  • The shift of AI workloads toward 90%–95% inference reduces the reliance on massive cluster-wide GPU interconnects, favoring local rack-level coherence where competitors can compete.
• Grid flexibility investments: NVIDIA is actively investing in energy and grid-flexibility startups (such as Emerald AI) to optimize workload orchestration and maximize power utilization from existing grid infrastructure.

Takeaways

• While hardware demand remains unprecedented, the primary headwind to NVIDIA's revenue growth is physical power availability rather than chip fabrication capacity.
• Monitor developments in compiler technology and AI model inference shifts, as they present long-term margin risks to NVIDIA's high pricing power and CUDA ecosystem lock-in.

Power Grid Infrastructure & Regulated Utilities (XLU)

• Transmission grid delays: The interconnection queue to connect new power generation to the grid has lengthened from 15 months two decades ago to roughly 45 months today.
  • Large data center power requests in fast-moving regions like ERCOT (Texas) now face lead times stretching to 2031–2032.
  • ERCOT currently peaks at 80 GW but has over 200 GW of speculative large-load interconnection requests submitted.
• Utility business model incentives: Traditional US monopoly utilities operate under a "cost-plus" return model (typically targeting a 10% return on capital approved by regulators), which incentivizes large capital expenditure rather than rapid speed-to-power delivery.
• Grid flexibility and interruptible load unlock: Regulators like FERC and ERCOT have enacted new rules for interruptible loads (PCLR / CLR).
  • Data centers capable of shedding load or operating on battery storage during peak demand periods (roughly 100 hours or 1% of the year) can bypass standard queues, reducing connection wait times from 5–7 years down to 12–18 months.
  • Utilizing existing grid off-peak margins (the 200 GW gap between nighttime lows and summer peak demand) can unlock up to $5 trillion to $10 trillion in AI capital expenditure.
  • Startups mentioned in this space include Agentic Infrastructure (site battery integration) and WeaveGrid (EV load management and time-slicing software).

Takeaways

• Grid equipment, transmission capacity, and load-balancing software represent the single largest "picks and shovels" bottleneck in the AI buildout.
• High near-term commercial value exists in battery time-shifting infrastructure and grid-orchestration software that allow data centers to qualify as fast-tracked interruptible loads.

Natural Gas Turbines & Behind-the-Meter Generation (GE)

• Equipment backlogs: Large-scale natural gas turbines (e.g., 400-megawatt units from GE and Hitachi) are sold out with backlogs of approximately 7 years, prompting manufacturers to expand assembly line capacity.
• Modular turbine pivot: Hyperscalers and data center operators are increasingly deploying behind-the-meter, on-site power generation to bypass grid connection queues entirely.
  • Mobile, trailer-mounted natural gas turbines (such as 38-megawatt units from Solar Turbines, where ~40 units equal 1 GW) are being deployed to power data centers immediately.
  • Startups adjacent to jet propulsion (such as Boom Supersonic) are pivoting engine technologies toward modular natural gas turbine manufacturing for data center power.

Takeaways

• Behind-the-meter modular power generation is the primary interim bridge for data centers waiting on utility grid connections through the late 2020s.
• Heavy industrial turbine manufacturers and modular power equipment providers are positioned to capture substantial capital expenditure from hyperscalers willing to pay a premium for speed-to-power.

Solar Energy & Battery Storage Systems (TSLA)

• Plunging solar costs: Solar panel hardware costs have dropped over 1,000x since 1975 to approximately $0.08 per watt for Chinese-manufactured modules, with global capacity scaling according to Wright's Law (costs dropping ~30% per doubling of cumulative scale).
• Baseload solar plus storage: Combining utility-scale solar with massive battery storage has reached cost parity with conventional baseload power:
  • Example: A 1 GW 24/7 baseload facility in the UAE requires 5 GW of solar panels and 19 GWh of battery storage at an all-in capital cost of roughly $6 per watt (compared to modern US nuclear at $15 per watt and Chinese nuclear at $4 per watt).
  • Solar and battery facilities are the fastest power plants to construct, deployable within 12 months when land permitting is secured.
• Battery cost trajectory: Lithium-ion battery prices have declined 14x since 2010. Emerging chemistries like sodium-ion have the potential to reduce battery manufacturing costs by an additional 10x due to abundant raw materials.
• Manufacturing scale (Tesla): Tesla (TSLA) has outlined strategic ambitions to build up to 100 GW of terrestrial solar and battery manufacturing capacity, potentially serving both terrestrial data centers and long-term space-based power applications.
• Key limitations: Seasonal solar intermittency (especially winter heating loads in higher latitudes like Northern Europe) limits 100% solar penetration without seasonal long-duration storage, requiring complementary firm power sources.

Takeaways

• Solar plus battery storage is currently the fastest, most cost-competitive scalable power solution for AI data centers sited in high-insolation regions (deserts, Southern US, Middle East, Australia, Northern Mexico).
• Keep a close watch on sodium-ion battery commercialization as the next catalyst for drastic cost reductions in diurnal energy storage.

Small Modular Reactors (SMRs) & Advanced Nuclear Fission

• Traditional fission landscape: Large-scale gigawatt reactors (like the Westinghouse AP-1000) face high upfront capital costs and multi-year construction delays in Western nations, though loan guarantee programs are being deployed to build standardized fleets.
• The SMR investment boom: Venture capital and public markets are heavily funding Small Modular Reactors designed for factory manufacturing rather than field construction:
  • Valar Atomics recently raised $1 billion at a $6 billion valuation.
  • X-Energy recently went public to commercialize high-temperature gas-cooled reactors.
  • Oklo and Radiant are developing compact micro-reactors (ranging from 5 MW to 50 MW containerized units).
• Deployment timelines: SMR commercial operation is widely projected between 2030 and the early 2030s, though first-of-a-kind units are expected to face initial cost overruns and timeline slippage before achieving manufacturing scale efficiencies.

Takeaways

• SMRs present a multi-billion-dollar market opportunity driven by AI hyperscaler demand for 24/7 carbon-free power, but commercial revenue at scale remains outside the 5-year investment window.
• The safest near-term nuclear plays involve life extensions and restarts of decommissioned conventional reactors (e.g., Three Mile Island) rather than unproven first-generation SMR builds.

Commercial Nuclear Fusion Startups

• Timeline acceleration: Over 50 venture-backed fusion startups are currently operating, shifting the investment thesis from scientific theory to engineering execution:
  • Helion Energy holds a commercial power purchase agreement with Microsoft targeting 50 MW of grid delivery by 2028. Helion uses a pulsed magneto-inertial system that captures energy directly as electricity, bypassing steam turbine efficiency losses.
  • Commonwealth Fusion Systems (CFS) is utilizing high-temperature superconducting (HTS) magnets to shrink Tokamak designs down to a planned 600 MW commercial scale.
  • Avalanche Fusion is developing highly compact, micro-scale fusion devices.
• Favorable regulatory environment: The US Nuclear Regulatory Commission (NRC) determined that fusion reactors will be regulated under radiological health frameworks (similar to medical imaging equipment) rather than the stringent regulatory rules applied to nuclear fission, significantly reducing licensing friction.

Takeaways

• Fusion is transitioning into an investable sector, with direct power purchase agreements appearing as early as 2028–2030.
• While execution and engineering risks remain high, favorable regulatory classification gives fusion startups a faster path to commercial deployment than next-generation fission if physics milestones are met.

Advanced Geothermal & Wave-Powered Energy

• Advanced Geothermal Systems: Startups including Fervo Energy, Eavor, and Quaise (utilizing millimeter-wave plasma drilling) are unlocking deep geothermal energy, transforming it from a niche geography-dependent resource into a widely deployable 24/7 baseload power source.
• Ocean wave-powered compute: Startups backed by prominent venture investors (including Peter Thiel) are deploying floating ocean data centers:
  • Buoy-based generation captures deep-ocean wave energy in regions with continuous swell (e.g., the Southern Ocean near Antarctica) targeting power costs around $0.02 per kilowatt-hour (kWh).
  • Systems utilize the surrounding cold ocean water for free, passive GPU heat dissipation without complex pumping systems, extending hardware lifespan and eliminating land permitting constraints.

Takeaways

• Advanced geothermal represents a prime dark-horse candidate for continuous, clean, behind-the-meter data center power.
• Off-grid extreme environments (ocean wave buoys) offer novel bypass routes for compute workloads (such as AI batch training or crypto mining) that do not require ultra-low-latency terrestrial fiber connections.

Space-Based Compute & Starship Launch Infrastructure

• Long-term compute in orbit: Placing data centers in space leverages continuous solar radiation and avoids all terrestrial permitting, land acquisition, water use, and grid bottleneck issues.
• Launch cadence and cost hurdles:
  • AI in space becomes economically viable once launch costs fall 4x to 10x below current levels.
  • Siting 10 GW per year of compute in orbit would require approximately 1,500 to 2,000 Starship launches per year (5 to 6 launches per day)—a scale that exceeds cumulative historical human space launches.
  • SpaceX's long-term goal of 100 GW per year in orbit by the 2030s requires airline-like operational cadence (100 launches per day), which remains technically and regulatorily constrained in the near term.

Takeaways

• Space-based data centers serve as a massive future demand driver for heavy-lift launch vehicles like SpaceX Starship, but terrestrial solutions (solar, batteries, gas turbines, and nuclear) will dominate the AI compute infrastructure buildout through at least the mid-2030s.