Delian Asparouhov on AI gross margins, private credit replacing IPOs, and the White House space EO

Summary

Delian Asparouhov — partner at Founders Fund and co-founder of Varda Space — argues that the AI stack's economics are structurally broken at the application layer, and that the winners so far look a lot like the winners of every prior infrastructure boom.

The negative gross margin problem

Asparouhov frames the problem with a figure passed to him through a group chat that he believes to be roughly accurate: a user pays an application-layer company $1, which pays a foundation model company $5, which pays a hyperscaler $7, which pays a GPU maker $13. Every layer up the stack is subsidising the layer below it.

The 2010s software mantra — marginal cost of distribution is zero, margins are phenomenal — has not held. SaaS companies ended up with bloated sales and marketing spend, and the AI companies replacing them face the same durability questions at far worse unit economics. Alex Clayton at Meritech recently noted that publicly reported AI lab revenues now surpass the entire publicly traded SaaS industry, yet the free cash flow is pooling at Oracle, Microsoft, Nvidia, and energy companies — not at the application layer.

Asparouhov, who has long favoured companies combining bits and atoms over pure software plays, says his Varda Space factory has better gross margins than many of the AI companies he tracks.

The inference cost bull case

The one credible escape route is inference cost deflation. DeepSeek dropped reasoning model inference costs roughly 20x in a short period — a compression speed that simply doesn't happen in physical industries. Asparouhov is cautiously optimistic that continued algorithmic efficiency gains could right-size margins for application-layer companies, but frames it honestly as everyone hoping for a bailout.

Historical precedent

AI capex as a share of GDP is now running at roughly 1.3%, just above the 1.2% reached during the late-1990s fiber rollout and well below the railroad boom of the late 1890s, which hit close to 9% of GDP. The fast-takeoff crowd — he cites Dario Amodei and Satya Nadella — argues it could reach 50% of GDP if scaling keeps working.

The structural difference from prior booms is that fiber and railroad companies depended on revenues from the infrastructure itself to keep building. Today's AI capex is largely funded by free cash flow from entirely unrelated businesses — Meta from social media and targeted ads, Microsoft from SaaS, Amazon from its marketplace. Corning, the dominant fiber rollout company of the late 1990s, only recovered its peak market cap last year, and most of that recovery came from Gorilla Glass for iPhones, not fiber.

The implication Asparouhov leaves hanging is that deep-pocketed incumbents can sustain the build-out far longer than a pure-play infrastructure company could — which makes the current boom more durable than the fiber bust, but doesn't resolve where application-layer margins come from.