Key Takeaways

  • One dollar of Nvidia fabrication cost compounds into eighteen dollars of model software revenue as value stacks across hardware, compute clusters, and foundation models.
  • Dario Amodei separates raw compute scaling from algorithmic compute multipliers, which allow foundation models to drop execution costs tenfold while lifting software gross margins to 85%.
  • Downstream productivity gains are already showing up in corporate filings: Microsoft grew revenue from $198 billion to $332 billion over four years while adding only 2,000 employees to total headcount.
  • Token consumption will expand by up to 500x due to Jevons paradox as automated software agents replace manual human prompting.
  • Tracking margins across each layer requires David Friedberg's AI Value Stack Economic Model.

The David Friedberg's AI Value Stack Economic Model

Layer 1: Chip Fabrication (Hardware Component)

Nvidia spends $1.00 of Cost of Goods Sold (COGS) to fabricate silicon, which sells for $4.00 of revenue at a 75% gross margin.

Layer 2: Infrastructure and Neocloud Depreciation

Infrastructure providers (such as SpaceX or Colossus) acquire the silicon and deploy compute clusters, amortizing capex over a 5-year useful depreciation cycle to generate rental revenue at an estimated 83% operating margin.

Layer 3: Model Layer Software Monetization

Foundation model developers (such as Anthropic) apply compute multipliers and algorithmic efficiencies to earn an 85% incremental gross margin, turning that initial $1.00 of hardware COGS into $18.00 of customer-facing token revenue.

Layer 4: End-User Enterprise Value Realization

Downstream enterprise customers realize an estimated 2x to 10x ROI on token spend through productivity gains and workforce efficiency.

When This Works (and When It Doesn't)

This framework tracks gross margin capture and value compounding across each step of the AI supply chain, from semiconductor manufacturing to end-user software delivery. It applies directly when assessing foundation model developers and neocloud operators that sell compute as a metered utility. By separating the tiers, founders can pinpoint where gross margins hold steady and where capital expenditure gets absorbed.

Where the model breaks down is in assuming linear five-year depreciation schedules. If algorithmic advancements cause older GPU clusters to obsolete in two years rather than five, infrastructure rental margins will contract. Neoclouds will face severe write-downs before their hardware fully amortizes. Additionally, if commoditization hits the model tier through open-weight alternatives, developers will fail to capture that 85% incremental gross margin. They will be forced to compete on pure inference price instead.

What to Do With This

Assume you are a founder running a vertical SaaS business spending $12,000 a month on commercial model APIs. Stop treating token bills as an operational overhead cost. Map your expenses against the four tiers of this model to audit your actual unit economics.

First, isolate your prompt runs into two buckets: standard queries and agentic loops. Second, calculate your cost per successfully resolved customer workflow. If your software charges a flat seat price of $50 per month, an autonomous agent loop that burns $18 in raw model tokens consumes more than a third of your gross margin.

Third, counter this squeeze by taking advantage of algorithmic compute multipliers. Swap out top-tier frontier models on routine extraction tasks for smaller, task-specific models that cost a tenth of the price. If you do not rebuild your software architecture to capture these cost reductions, your foundation model supplier keeps an 85% margin while your own software margins drop to zero.