# Stacked Margins in the AI Chain Every layer of the AI chain now earns a 75-85% margin. That is a scarcity price, not a stable structure. Base: the "$1 of Nvidia COGS" slide. It calls itself an illustrative model. Karan shared it; the original author is unknown. | Layer | Per $1 of Nvidia COGS | Margin | |---|---|---| | Nvidia | $4.00, one time | 75% gross | | SpaceXAI | $8.60 per year | ~83% | | Anthropic | $17.80 per year | 85% gross, 52% after all compute | | Customers | $36-178 per year | Assumption | The slide arithmetic reconciles. Its inputs are the weak part. ## Insights - **Tokens get cheaper while GPUs get dearer.** - Cheaper tokens drive more usage. More usage needs more GPU-hours. - This is [[Jevons Paradox]]: efficiency raises total demand for the input. - The token price falls. The rent for compute rises. - **One contract carries the chain.** - The slide divides $15B per year by ~$10B of Colossus 1 capex. - Press reports say the deal covers Colossus 1 and Colossus 2. If true, the $8.60 is too high.^[https://www.techloy.com/spacex-ipo-filing-reveals-anthropic-pays-41-million-a-day/] - On 220,000 GPUs, the price is about $7.80 per GPU-hour.^[https://www.constellationr.com/index%2Ephp/insights/news/anthropic-buys-all-spacexs-colossus-1-data-center-capacity] - [[IREN thread]] shows older H100 pricing of $2-2.50. The clusters differ, so this is a rough check. - **Price is the risk, not GPU life.** - A 3-year life moves the margin from 83% to 74%. - A $2.25 price moves the margin to about 42%. - **The contract is shorter than the asset.** - The S-1 reportedly allows termination on 90 days' notice.^[https://www.techloy.com/spacex-ipo-filing-reveals-anthropic-pays-41-million-a-day/] - The GPUs have an assumed 5-year life. - The buyer may pay a premium for speed and the right to leave. - **Non-serving compute is the largest wedge.** - Serving costs 15 points of Anthropic revenue. - Non-serving compute costs 33 points. - One analyst puts channel fees at 17 points. - **Nvidia is the only one-time box.** - Other layers earn per year. Nvidia must sell again. - The chain mixes a one-time sale and annual flows. Do not read it as one multiplier. - **The only unproven box pays for everything.** - Customer value is an assumption. - Every margin to its right depends on it. ## Concept map ```mermaid graph LR subgraph Cheap["Gets cheap"] K(Tokens) S(Serving cost) G(Compute, once capacity lands) end subgraph Scarce["Stays scarce"] H(GPU-hours, for now) W(Non-serving compute) V(Proven customer value) end subgraph Notes["Concept"] c1[Jevons Paradox] c2[Systems and Compiler Depth] c3[Own the Margin You Create] end K --> H --> c1 S --> W --> c2 G --> V --> c3 class c1,c2,c3 internal-link; ``` ## So what for us - Sell into a wedge, not a layer. High supplier margins give cost-cutters room. - [[Scarcity Migration]] applies: ask what stays scarce when GPU supply catches up. - Treat measured customer value as the scarce evidence. > [!question] > Which layer keeps its margin when compute stops being scarce? ---- tags: #boat #AIstrategy links: [[The Studio Insights]] [[Token Economics of Inference]] [[Neoclouds]] [[Own the Margin You Create]]