Zitron — AI Is Too Expensive / Why Everybody Is Losing Money On AI

Citation

Zitron, Ed. AI Is Too Expensive (2024) and Why Everybody Is Losing Money On AI (2026). Better Offline podcast / wheresyoured.at newsletter. URLs: https://www.wheresyoured.at/ai-is-too-expensive/ and https://www.wheresyoured.at/why-everybody-is-losing-money-on-ai/

Status

Stub. Currently summarized only via bear-case-synthesis. Primary essays not yet ingested directly.

Core argument (as inherited)

The AI application and product layer is a financial ouroboros:

1. Venture capital funds application-layer startups (Cursor, Perplexity, etc).
2. Startups immediately pay foundation models (OpenAI, Anthropic) for API access — many at deeply inverted margins (Perplexity ~164% of revenue on compute in 2024).
3. Foundation models pay hyperscalers (AWS, Azure, GCP) for compute — OpenAI ~50% revenue on inference + ~75% on training.
4. Hyperscalers report this VC cash as “AI revenue” to Wall Street, justifying further capex.
5. ~70–80% of hyperscaler AI capacity is consumed by just two entities: OpenAI and Anthropic.

Genuine non-tech enterprise revenue is barely trickling in. When VC funding tightens, the chain collapses from the application layer up, leaving hyperscalers with rapidly depreciating specialized hardware and stranded power contracts.

Claims it challenges

• That AI software has product-market fit and is generating durable enterprise revenue.
• That the “AI revenue” hyperscalers report on earnings calls reflects sustainable demand.
• The bottom-of-the-U thesis (infrastructure as a safe place to capture margin) — at least under current revenue dynamics.

Why it matters for Where Value Lands

• The single most direct threat to H2_u-curve-of-value: it argues the bottom of the U may also be hollow, propped up by recycled VC. If true, the U-shape is wrong.
• Strongest empirical artillery on the product layer is structurally unprofitable claim.
• Connects to scenario 1 (“Hyperscaler consolidation”) in 03_structure — Zitron’s view implies that scenario ends in a correction, not a stable equilibrium.

To verify on primary read

• The exact 70–80% hyperscaler-capacity number (and whether it includes Microsoft’s OpenAI in-kind equity arrangement).
• Per-company unit economics (Cursor, Perplexity, OpenAI, Anthropic) — these change quickly.
• Whether Zitron names a specific trigger or only describes the dynamic.

Data surfaced by this source

• hyperscaler-capex-trajectory — $800Bbooked+$1.7T planned.
• hyperscaler-customer-concentration — 70–80% concentration to OpenAI + Anthropic.
• perplexity-burn-ratio — ~164% of revenue on compute (2024).

Related

• bear-case-synthesis
• circular-ai-economy
• H2_u-curve-of-value
• wef-ai-in-action-2025 — counterpart bull synthesis; treats the scaling gap as managerial where Zitron treats it as structural.