Sunil Venkataram
Back to garden
•
🌰 seedling
AI Stack Value Accrual - Chip, Infra, Intelligence, App
type/noteaiinvesting

AI Stack Value Accrual — Chip, Infra, Intelligence, App

The four layers

  1. Chip layer (Nvidia) — captured the earliest value wave. The compute bottleneck meant whoever sold the shovels during the dig got paid first.
  2. Hyperscaler / infrastructure layer (AWS, Azure, GCP) — capturing now. They own the datacenter buildout, power contracts, and the distribution channel for foundation models.
  3. Intelligence / model layer (Anthropic, OpenAI, xAI) — capturing in the current moment. Revenue trajectory and the 90% YoY inference cost collapse suggest margins are expanding even as prices fall.
  4. Application layer (Salesforce, HubSpot, Oracle-era SaaS) — open question.

The application-layer question

Two hypotheses, both with believers:

  • Gutted: Model companies eat the functions these apps used to package. Enterprise SaaS gets compressed into thin UI layers over shared model infrastructure. Multi-billion-dollar software franchises lose their pricing power as the intelligence commoditizes their core loop.
  • Turbocharged: Existing apps have distribution, trust, and workflow embedding that models can’t replicate. AI capabilities make existing apps more valuable per seat, not less, because the customer already trusts the vendor with their data and processes.

Sacks’s read (from All-In E230 - Anthropic Mythos OpenClaw Where AI Value Accrues): there is probably some baby-with-the-bathwater selling happening in enterprise software stocks. Value will accumulate at multiple layers of the stack, not just the model tier.

Why this framework matters

The layered view is a positioning tool, not a prediction. Each layer has a different capture window, a different moat structure, and a different buyer. An investor betting on “AI” without specifying a layer is really betting on whichever layer happens to be en vogue when the trade is placed. The disciplined version of the thesis names the layer, names the moat, and names the expected capture window.

Alternative framing — the application company’s three-layer view

Tanay Jaipuria offers a related but differently-sliced frame focused on the application company’s perspective: model / application / service. This is narrower than the full chip→infra→intelligence→app stack above but sharper for thinking about how AI application companies compete and vertically integrate. See Full Stack Down vs Full Stack Up - Two Directions for AI Application Vertical Integration. The two frames are complementary — the chip/infra/intel/app view helps with investment layer selection; the model/app/service view helps with product strategy inside the application layer.

Backlinks (5)
Connected Notes