How the next trillion dollar compute expansion is reshaping global energy markets.
The Energy Hunger of Artificial Intelligence
AI’s growth has become the most energy intensive technological shift since the invention of the semiconductor.
Each layer of the AI stack, from training foundation models to running global inference, depends on firm, affordable, and scalable power.
The industry once measured its progress in parameters and GPUs; it now measures it in megawatts.
Data centres are morphing into energy ecosystems. NVIDIA’s GB200 NVL72 racks demand up to 120 kW each, nearly ten times the load of legacy systems. As hyperscalers race to deploy trillion parameter models and autonomous agents, power, not compute, has emerged as the defining bottleneck.
The Coming Power Crunch
According to the International Energy Agency (IEA), global data centre electricity use is projected to nearly double to ~945 TWh by 2030, with AI responsible for the majority of that increase.
In the U.S., the Department of Energy now forecasts computing and data infrastructure could become the largest single use of commercial electricity by 2050.
At this trajectory, the AI industry alone could consume as much power as Japan within a decade.
This surge is forcing a reorganization of global energy priorities. The world’s next trillion dollar expansion, AI compute, will not be constrained by chips or capital, but by access to reliable electrons.
The Energy Stack of the AI Era
Nuclear: The Firm Backbone
Nuclear power is the only 24/7 zero carbon energy source capable of supporting hyperscale AI loads at global scale.
Existing plants and restarts are seeing a resurgence. Microsoft has signed a 20 year PPA with Constellation Energy to restart the Three Mile Island reactor (~835 MW), while Google and Kairos Power plan advanced nuclear deployments by 2035.
Key data:
- Capacity factor: ~83 – 90%, the highest of any source.
- Cost: $40 – 60/MWh for existing plants; $190 – 284/MWh for new builds.
- Timeline: Existing PPAs = 0 – 3 years; New SMRs = 8 – 12 years.
Nuclear’s reliability makes it the cornerstone for AI baseload power, particularly in regions like Ontario, Tennessee, and Virginia where regulatory frameworks already support nuclear integration.
Solar: The Fastest Builder
Utility scale solar remains the fastest deployable energy source for offsetting daytime and peak AI loads.
Construction timelines average 2 – 3 years, and costs have fallen to $27 – 92/MWh, the cheapest new generation globally.
When paired with 4 hour batteries, solar reduces grid strain and hedges long term energy costs.
However, solar’s 20 – 35% capacity factor means it cannot replace firm baseload power. It excels as part of a hybrid strategy, providing clean daytime energy and enabling geographic diversification for AI clusters.
Batteries: The Bridging Storage Layer
Energy storage is the glue between intermittent renewables and continuous AI demand.
Modern data centres are adopting onsite 4 hour lithium systems for peak shaving, while utilities invest in longer duration flow batteries.
Installed costs hover around $150 – $200/kWh, limiting economic feasibility beyond short term balancing.
But the next wave, thermal, hydrogen, and iron air storage, promises to deliver multi-day firming, enabling fully renewable AI campuses later in the 2030s.
Natural Gas: The Near Term Workhorse
Despite decarbonization commitments, natural gas remains the fastest deployable firm capacity.
New combined cycle plants can be built in 2 – 4 years, offering dispatchable power at $45 – 108/MWh.
Emissions (~0.49 tCO₂e/MWh) are significant, but gas infrastructure provides the reliability buffer that keeps data centres operational when renewable or nuclear sources lag.
In the near term, fossil generation acts as the bridge fuel, a pragmatic necessity until firm clean power scales.
Comparative Energy Outlook
| Attribute | Nuclear (Existing/Restart) | New Nuclear (SMR/Large) | Solar (+ Storage) | Natural Gas |
|---|---|---|---|---|
| Time to Power | 0 – 3 years | 8 – 12+ years | 2 – 3 years | 2 – 4 years |
| 24/7 Firm Power | ✅ | ✅ | ❌ | ✅ |
| Capacity Factor | 80 – 90% | 80 – 90% | 20 – 35% | 50 – 80% |
| LCOE (US) | $40 – 60/MWh | $190 – 284/MWh | $27 – 92/MWh | $45 – 108/MWh |
| Carbon | 0 | 0 | 0 | ~490 kg/MWh |
| AI Relevance | 24/7 clean baseload | Future firm power | Fast, cheap offset | Short-term firming |
The Road to 2035
| Period | Strategy | Key Actions |
|---|---|---|
| 2025–2027 | Stabilize the grid for AI | Secure nuclear-backed PPAs; deploy solar + storage; fast-track interconnects; leverage gas for near-term capacity. |
| 2027–2030 | Bridge to firm clean power | Pilot SMRs and advanced nuclear; expand hybrid solar + battery parks; adopt 24/7 carbon-matching for AI load. |
| 2030–2035 | Fully firmed AI infrastructure | Co-locate SMR, solar, and storage in AI hubs; integrate heat reuse and local microgrids; energy becomes the new moat. |
Investor Takeaway: Power Is the New Compute
The bottleneck for AI is not only capital or chips, it’s energy.
Control over megawatts will define who can scale inference, host training clusters, and ultimately lead the intelligence economy.
The next generation of defensible infrastructure will be measured not just by compute density, but by energy sovereignty:
- Data centers with secured nuclear PPAs will outlast power scarce competitors.
- Developers who integrate solar + long duration storage will own the cost curve.
- Utilities that fast track firm interconnects will define national AI competitiveness.
In the coming decade, the winners won’t just be those who build models, they’ll be those who power them.
Read More
- International Energy Agency (IEA) – Electricity 2024 Report
Global data center demand projections and AI energy scenarios.
https://www.iea.org/reports/electricity-2024 - U.S. Department of Energy (DOE) – Electricity Use in Data Centers Report, 2024
Outlook for U.S. computing energy consumption through 2050.
https://www.energy.gov/ - Lazard Levelized Cost of Energy v17.0 (2024)
Comparative LCOE and cost data for nuclear, solar, and gas.
https://www.lazard.com/perspective/lcoe/ - World Nuclear Association (2025) – Performance Report
Updated capacity factors, plant lifetimes, and global new-build timelines.
https://world-nuclear.org - Reuters & Bloomberg – Microsoft, Google, Meta nuclear PPAs
The corporate pivot toward firm clean energy for AI operations.
https://www.reuters.com | https://www.bloomberg.com