Abhi Dewan

A single training run for a large AI model today consumes electricity equivalent to 100+ U.S. households for an entire year.

My (4-part) Working Thesis

1. As AI becomes increasingly compute-intensive, giants like NVIDIA, Amazon, Google and others will step up to solve the compute bottleneck.
2. The solution for the compute bottleneck will create another critical constraint: energy.
3. Energy will experience incremental improvements around current infrastructure (e.g. transmission, distribution, storage, etc), but the next step-change innovation in energy will be nuclear.
4. Finally, AI (and other) data centers will serve as a beachhead market for nuclear, given the “product” is self-contained, modular, and “generator-like”.

Is this crazy? Maybe. But that’s why I’m sharing it now:

There are a lot of “if”s in there – and even more counter-arguments…

So here’s a handful of key assumptions that I plan to validate (starting with first principles):

Key Assumptions to Validate

1: Energy Will Become a Critical Bottleneck for AI

Current belief: AI-driven energy needs are growing exponentially, outpacing efficiency improvements and grid capacity.

The color: Microsoft is already struggling to secure enough power for its data centers. Unlike compute, you can’t just throw money at the electrical grid and expect Silicon Valley-speed scaling.

Counter-arguments: Algorithmic efficiency improvements may dramatically reduce energy needs; specialized AI chips could achieve order-of-magnitude efficiency gains.

Validation approach: Analyze energy consumption trends of recent model releases; interview data center operators about power planning; research physical limits of AI computing efficiency.

2: Incremental Energy Solutions Won’t Keep Pace

Current belief: Traditional grid expansion, renewables, and batteries won’t deliver the energy density and reliability needed projected 5-10 years out from now.

The color: Grid modernization moves painfully slow – decades, not years. Meanwhile, AI capabilities double every few months. The math doesn’t work.

Counter-arguments: Breakthrough battery technologies could arrive sooner than expected; fusion power might become viable within the relevant timeframe.

Validation approach: Compare renewable deployment timelines with AI energy demand projections; analyze grid capacity expansion rates in key tech hubs; examine energy locality requirements for latency-sensitive AI applications.

3: Modular Nuclear Technology is Viable Soon Enough

Current belief: Technology for safe, small-scale nuclear generators can be commercialized within 5-10 years.

The color: There are designs for nuclear “batteries” – sealed units that could power a data center for years without refueling. These aren’t your grandfather’s nuclear plants.

Counter-arguments: Safety concerns, waste management, and uranium supply could present insurmountable obstacles; technical hurdles might prove more challenging than anticipated.

Validation approach: Research current modular reactor designs and their technological readiness; interview nuclear engineers about feasibility timelines; examine Department of Energy and DARPA research in this area.

4: The Economics Can Work

Current belief: Despite high upfront costs, modular nuclear will be competitive when factoring in reliability and the opportunity cost of AI deployment delays.

The color: The value of reliable, dense power for AI isn’t just measured in cents per kilowatt-hour – it’s measured in competitive advantage worth billions.

Counter-arguments: Capital costs may remain prohibitively high; alternative energy sources might achieve better economics through scale and innovation.

Validation approach: Develop cost models comparing nuclear to alternatives for data center applications; research economies of scale in modular reactor production; interview nuclear industry experts about cost reduction pathways.

5: Regulatory Pathways Exist

Current belief: Despite challenges, there are viable regulatory approaches for data center-specific nuclear deployment.

The color: When national priorities shift, regulatory pathways emerge. If AI is recognized as critical infrastructure for national security, frameworks can adapt faster than conventional wisdom suggests.

Counter-arguments: The regulatory timeline may be too long to matter; public opposition could be insurmountable; political gridlock might prevent necessary reforms.

Validation approach: Map the current regulatory framework for small modular reactors; interview former NRC officials about potential pathways; research international precedents for accelerated nuclear deployment.

6: Data Centers Are the Right Beachhead Market

Current belief: Data centers need massive, reliable power and have both the capital and incentive to pioneer new approaches.

The color: They’re already highly secure, controlled environments operated by sophisticated companies – perfect for demonstrating safety and reliability.

Counter-arguments: Data centers may prefer distributed renewable solutions; major cloud providers may resist nuclear dependencies due to PR concerns and perceived risk.

Validation approach: Interview data center operators about energy constraints and preferences; analyze total cost of ownership models for various energy solutions; research current data center energy strategies and roadmaps.

7: This Is a Venture-Backable Business Model

Current belief: The modular nuclear solution for AI infrastructure represents a venture-scale opportunity with appropriate risk-reward profiles and exit potential.

The color: Traditional energy infrastructure has been challenging for VCs, but the intersection with AI creates a new paradigm. The potential multiples here aren’t utility-like 2-3x returns, but tech-like 10-100x if timed correctly.

Counter-arguments: Capital requirements may be too high for traditional VC; regulatory timelines may exceed venture fund lifespans; established energy players may squeeze out startups.

Validation approach: Model capital efficiency and unit economics for potential startups in this space; interview energy-focused VCs about investment criteria; analyze exits of capital-intensive, regulated-market startups in adjacent sectors.

Can You Help?

If you have expertise in energy infrastructure, AI scaling, nuclear technology, or data centers, I want to hear from you. If you think I’m wrong about any assumptions, I especially want to hear from you.

The future of American innovation may depend on how we solve the energy equation for AI – and that conversation needs to start now.