Signal: "The Subprime AI Crisis"
Signal: “The Subprime AI Crisis”
Date: 2026-04-04. Source: Ed Zitron, Where’s Your Ed At.
The claim
The AI industry mirrors the 2008 housing crisis. A “chain of pain” where money flows through debt-funded data centers → hyperscalers → AI labs → AI startups → consumers, with profitability only at the NVIDIA hardware layer.
Evidence cited
| Metric | Number |
|---|---|
| Anthropic compute spend | $10B |
| Anthropic revenue | $5B |
| OpenAI inference burn (through Sept 2025) | $8.67B |
| OpenAI total revenue | $4.3B |
| Startup burn ratio | $3–13 per $1 of subscription revenue |
| Data center construction (actual) | ~5GW |
| Data center construction (promised) | 12GW+ |
| OpenAI valuation | $850B |
| OpenAI unsold shares | $600M |
| Total AI industry revenue (2025) | ~$65B (not profit) |
Why this matters for the tracking
If the subsidy-dependent pricing is temporary, every tool on the radar should be evaluated for pricing durability. The most defensible investments are in patterns (spec-driven dev, orchestration architecture, MCP integration) rather than specific vendor subscriptions.
Local models become more strategically important in this frame — they’re the hedge against pricing changes in cloud-hosted AI.
My read
The financial analysis is credible and the numbers are publicly verifiable. The tools and workflows are also genuinely productive — these aren’t contradictory claims. The question is what happens to the ecosystem when the subsidies end. Price increases and rate limits (Anthropic mid-2025) are already the canary signals.
This strengthens the case for local models as a hedge against cloud pricing volatility.