To The Moon

This infrastructure buildout is the biggest malinvestment bubble since the dot-com era.

Everyone’s piling into AI infrastructure plays, but I think we’re witnessing a classic capital misallocation cycle:

CoreWeave’s $6B Pennsylvania bet is peak bubble behavior. They’re building massive GPU clusters based on current AI demand, but AI workloads are fundamentally different from traditional cloud computing. These specialized data centers will become stranded assets the moment AI efficiency improves or demand patterns shift. Remember all those fiber optic cables laid in 1999?

The “Trump AI Push” is political theater, not sustainable policy. Government-driven tech initiatives have a terrible track record - Solyndra, anyone? This Pennsylvania project smells like industrial policy designed to create jobs and photo ops, not economic returns. When the political winds shift, the subsidies disappear.

Navitas and the power semiconductor angle is even worse. Everyone assumes AI means infinite power demand, but the real trend is toward efficiency. Google’s TPUs, custom chips, and better algorithms are reducing power consumption per AI operation. We’re about to see a race toward efficient AI, not powerful AI.

The real problem: NVIDIA’s moat isn’t hardware - it’s CUDA software lock-in. But open-source alternatives are emerging fast, and every hyperscaler is building custom silicon. CoreWeave is betting billions on being NVIDIA’s landlord right as NVIDIA’s tenants are planning to move out.

Historical parallel: This feels exactly like the late 1990s telecom infrastructure boom. Massive capital deployment chasing a real trend, but with completely wrong assumptions about demand curves and technological evolution.

I’d rather short the infrastructure plays and buy NVIDIA directly - at least they have the software moat and can pivot when the hardware cycle turns.​​​​​​​​​​​​​​​​