Thesis II: Compute Is A Cascading Energy Problem

{{ AI + Chips + Data Centers + Raw Materials + Nuclear }}

To say compute is an energy problem is not a radical statement. But it is a yet unsolved issue and one that translates down the supply chain. Demand drives down the following cascade: 

  • AI Models 
  • Compute Capacity (for both training and use)
    • Decentralized Compute
      • Networking 
      • Parallelization
    • Data Centers
      • Racks
        • Chips (bespoke chips, GPUs, Memory) 
        • Raw Materials
      • Power
        • Nuclear (Fission)
        • Raw materials
        • Regulatory approvals
        • Nuclear (Fusion… Later, AI helps)
        • Raw materials
        • Regulatory approvals
      • Cooling
        • Water cooling towers / other methods of reuse
        • Closed loop liquid cooling at the box/rack

What Is Already True

  • In 2023-2025 the big demand was for GPUs. Purpose built chips are in development that would offload some of that demand. Memory is part of that demand stack too. Inference chips being developed now prove that the AI demand cascade will result in new chip designs and the evolution of chip design at large as AI accelerates chip design itself.
  • Nuclear as the preferred option is already in development. Major AI providers (Amazon, Microsoft, and Google) first put money into conventional reactor builds near new data centers, and new data centers near old, soon to be restarted reactors. The next move looks like semi-trailer portable nuclear (at least two companies I know of are working on this now). The trend towards on-site nuclear will likely flow down to “bathtub” nukes and “lunchbox” nukes with smaller sizes powering a single facility or building. Community support will be a hurdle and will likely be solved by smaller nuclear packaging, and having data centers provide excess power to the grid, lowering costs for local communities. 
  • Diesel fuel will likely remain the primary onsite fuel for a few more years as we overcome the realities of community sentiment, regulatory approvals, and Nuclear development.
  • Fusion is a hot topic. I do believe it will work at some point in the future, but I know enough about technology prototyping to know when predictions are real and when they are driven by the marketing team. Marketing teams cannot, unfortunately, change the laws of physics or the comfort level of regulators. Underwrite the right teams, but be patient.
  • The raw materials factor is real. Follow the cascade and you will see that AI demand requires transformers and other high-voltage equipment, skilled labor, water, rare earths, metals, etc.

Assumptions To Be Tested

  • Small and medium sized reactors in particular need to prove out their cost and see real delivery happen to data centers.
  • Nuclear on-site with data centers and other facilities is the idea, but we need to see that happen and the regulatory and community sentiment challenges are a significant hurdle.
  • Raw material risk needs to be solved. Rare earths, copper, gold, silver, enriched uranium, tritium, hydrogen or helium (for fusion), and other materials will be critical. Novel techniques for extracting rare earths are in the works with promising results, but are not tested anywhere near scale yet and other raw material needs will have to be repatriated against foreign dominance.
  • Decentralized compute projects are in the works but not yet deployed or tested at scale. These have significant potential if you can solve distribution and tasking (which is doable).

How I Would Underwrite It

I would be very careful of the marketing hype and who-has-invested-in-whom biasing a choice.  Data Center builds are a low-risk bet but Nuclear (fission or fusion) should go through deep diligence because each company’s technology stack and connections ARE the thesis for the deal. 

That being said, I would consider the entire demand cascade. It is a diverse set of plays that accelerate each other, which is my entire third derivative thesis.

UNDERWRITEAVOID
Nuclear plant restarts. Small to medium reactor development from the right teams with sufficient regulatory, manufacturing AND technical savvy. Decentralized compute. Raw materials (all). Novel techniques for rare earth extraction. Fusion only very selectively after extreme diligence; same for lunchbox or bathtub nukes intended for neighborhood/commercial installation. Solar/wind for hyper scale data centers. Any nuclear you have not done on-site diligence of across tech, management, regulatory etc. Data center with no power and community sentiment strategy. Chip development that does not address raw materials, power and cooling. New AI models (unless required to accelerate other bets you might have in Fusion or certain bio-medical plays).