The App Layer Is Where Capital Goes to Die

This piece is long, investor-brained, and worth skimming for one specific reason: it names something most AI commentary quietly avoids.

The application layer — ChatGPT, Copilot, the stuff you actually touch — is where most of the capital currently being deployed into AI will be destroyed. The author makes this directly: the companies everyone talks about are burning faster than they earn, structurally, because inference costs scale with quality. The real margin is sitting three layers down, in chips, cloud infrastructure, and power. That inversion — the thing you can see is the worst investment; the thing you've never heard of is printing money — is stated plainly and with enough specificity (Nvidia at 75% gross margins, TSMC at 70% foundry market share) that it actually lands.

Most AI content either ignores the stack entirely or mentions it in passing. This piece builds the whole argument around it.

The thing that doesn't earn its keep is the electricity analogy. The author calls it "Infrastructure Gravity" and presents it as a historical law: infrastructure always wins first, applications capture value later, pattern repeats. But the 1880–1920 electrification era had no open-source competitors, no efficiency breakthroughs that halved compute costs mid-cycle, and no geopolitical fragility baked into a single fabrication plant in Taiwan. The analogy is deployed as inevitability when it's actually a hypothesis. Section III is where the piece stops being analysis and starts being persuasion.

The five-layer framework is worth keeping. The confidence about what it predicts is not.