The Atoms Were Never Obsolete

Lee Roach, writing about what he calls the hard asset supercycle:

AI is reversing that detachment. We are entering an era where software becomes abundant while physical infrastructure becomes scarce. Capital markets are beginning a slow but profound reversion toward hard assets, tangible systems, and ownership of real-world bottlenecks.

There's a version of this argument that's obviously correct, and a version that's dangerously overfit. Roach is mostly making the first one, but he flirts with the second.

The obviously correct version: if AI makes it trivially cheap to build software, then software alone is a weaker moat than it was five years ago. Value migrates to whatever can't be replicated cheaply - and right now, that's compute infrastructure, energy, and physical supply chains. Data centres need electricity. Chips need fabs. Fabs need water, land, and years of construction. You can't prompt your way to a power station.

The overfit version: that we're witnessing a clean, secular rotation from bits to atoms. That physical assets are the trade of the next decade, full stop.

Reality will be messier. Network effects haven't evaporated. Google's moat isn't its codebase - it's the feedback loop between billions of users and the models trained on their behaviour. Enterprise switching costs aren't just about software quality; they're about data migration, compliance, and the sheer organisational inertia of changing how 10,000 people work. These things don't dissolve because a language model can write a prototype in an afternoon.

Where Roach is sharpest is on the "false prosperity phase" - the idea that AI-driven margin expansion looks brilliant in the early innings but represents a one-time adjustment, not a durable advantage. Once every competitor adopts the same tools, you're back to competing on fundamentals. The efficiency gain gets competed away or passed to customers. This is the part most earnings-call analysts are currently missing.

Where he's weakest is on timing. Nearly every prediction is hedged to "within the next decade," which is the investment thesis equivalent of saying it'll rain eventually. The difference between this thesis playing out over three years and over ten is the difference between a generational trade and a long wait for mean reversion.

The labour repricing section deserves attention. Not the headline version - "AI will take your job" - but the quieter mechanism: AI doesn't spike unemployment; it slows hiring and compresses wages. Entry-level roles shrink. Mid-level coordination work gets absorbed. The pain shows up in household balance sheets two or three years later, not in next month's jobs report. That's a harder problem to see coming and a harder problem to solve politically.

One significant assumption baked into the whole piece: that AI stays capital-intensive. If compute costs drop dramatically - through algorithmic breakthroughs, more efficient architectures, or genuine leaps in chip design - the infrastructure bottleneck loosens and the thesis weakens. Roach is betting the current cost curve persists. That's a reasonable bet. It's not a certainty.

Still, the core instinct is right. For two decades, markets priced atoms as if they were boring and bits as if they were magic. AI is forcing a correction. Not because software stops mattering, but because it stops being scarce.

The closing line is good enough to quote: "The atoms were never obsolete. They were simply underpriced."

Agreed. Though I'd add: and the repricing won't be as clean as any single thesis suggests.