# The Subprime AI Crisis Is Here

Edward Zitron's March 2026 piece draws a structural parallel between the 2008 housing crisis and the current AI industry. The core claim: every dollar of AI demand exists only because of subsidies, and the entire industry is now entering the phase where those subsidies become impossible to maintain.

## The Analogy

The housing bubble was built on adjustable-rate mortgages sold with teaser rates that hid the real cost. Borrowers made decisions based on flawed information — they assumed housing values would keep rising, refinancing would always be available, and the good times were permanent. When interest rates rose 17 times in two years, the whole chain collapsed: harder to borrow → harder to buy → prices fall → can't refinance → foreclosures → prices fall further.

Zitron argues AI subscriptions work the same way. Flat monthly fees ($20-$200) mask wildly variable token costs. Users build habits around subsidized access. When the subsidy shrinks — via rate limits, pricing changes, or degraded quality — the product becomes less valuable, but there's no price point where the economics work *and* users stay.

## The Chain of Pain

Money flows through six links, and only the first one is profitable:

1. **NVIDIA** sells GPUs to data centers and hyperscalers. The only consistently profitable player.
2. **Data center developers** borrow heavily to build capacity (~$178.5B in US debt in 2025). CoreWeave, the largest independent, ran -6% operating margin and -29% net margin in 2025. Of 200GW announced globally, only 5GW is actually under construction.
3. **Hyperscalers** (Google, Amazon, Microsoft, Meta) rent GPUs and resell access. None break out AI revenues.
4. **AI labs** (Anthropic, OpenAI) rent compute, sell API access and subscriptions. Anthropic: $5B revenue, $10B compute spend. OpenAI: $4.3B revenue, $8.67B inference spend (through Sept 2025). Neither has a path to profitability.
5. **AI startups** (Cursor, Perplexity, Harvey, Replit, Lovable) buy API access, sell subscriptions. Every one is unprofitable. Token burn exceeds subscription revenue by 3-10x.
6. **Consumers** pay flat monthly fees, unaware of actual token costs.

## The Predicted Sequence of Collapse

This is the forward-looking argument — how Zitron expects the crisis to unfold:

1. AI labs' costs keep growing (more compute, upfront capacity commitments). AI startups burn more per customer as they scale.
2. Labs face a cash and compute crush. They must either limit usage or charge more.
3. Labs introduce priority tiers and raise API prices on startups. This already started in June 2025 when Anthropic and OpenAI launched priority service tiers.
4. Startups are forced to degrade service or raise prices. Cursor switched to token-pass-through pricing. Replit moved to "effort-based" pricing. Augment Code went to per-message then credits. All faced user backlash.
5. Labs also tighten consumer subscriptions. Anthropic introduced peak hours in March 2026 after a 2x promo, immediately angering users who hit limits in minutes on $100-200/mo plans.
6. Users trained on unlimited access revolt. They can't be moved to usage-based pricing because they never internalized token costs. Rate limits feel like product degradation, not price changes.
7. Startups die as API costs become unsustainable and VC dries up. This kills a revenue channel for the labs.
8. Labs are left holding compute reservations they can't afford. Dario Amodei himself said "there's no hedge on Earth" that could prevent bankruptcy if Anthropic overbought compute.
9. Data centers built for demand that never materializes become stranded assets.

## The "Variable Rate Mortgage" of AI

A subtle point: labs don't just raise prices explicitly. They release new models that burn more tokens. Chain-of-thought reasoning, coding agents, deep research — each generation costs more per query. Users demand the newest models at the same subscription price. The cost goes up without the price label changing, exactly like an ARM adjusting upward.

## Pale Horses — What to Watch

Zitron's updated list of warning signs (originally from August 2024):

- Further price increases or rate limit tightening from Anthropic/OpenAI
- Capex reductions from big tech (kills NVIDIA's growth story)
- AI startup pricing changes, layoffs, or shutdowns
- Data center deal collapses — unbuilt, mid-construction, or completed
- CoreWeave or other data center players failing to raise debt (already signs: Lancaster PA deal issues)
- Problems with Stargate Abilene (OpenAI's flagship data center, built by Oracle)
- Delays or problems with OpenAI/Anthropic IPOs
- Blue Owl (aggressive AI lender) having trouble with its loan book
- SoftBank instability — $40B debt payable in a year to fund OpenAI round, ARM stock dependency ($15B margin loan, trouble below $80/share)
- NVIDIA customers unable to pay, or NVIDIA missing earnings

## What's Not in the Piece

Zitron explicitly says the AI bubble's economic footprint is much smaller than the 2008 crisis — trillions in CDOs vs tens of billions in AI revenue. The death of AI would devastate VCs, end the Magnificent Seven's hypergrowth era, and potentially kill Oracle, but wouldn't threaten the banking system the way subprime did. The comparison is structural (subsidized demand → inevitable repricing → collapse), not scale.

The piece also doesn't address open-source models as a potential escape valve. Startups fleeing to cheaper open-source alternatives could shift the dynamics, though Zitron briefly notes that lab price increases "push AI startups toward cheaper open source models and death." Ted Neward picks up this thread in [[titit-local-ai]], arguing that local open-source models are the rational response to the subsidy structure Zitron describes.

## Follow-up

Zitron's May 2026 [[ai-too-expensive]] piece updates this analysis with Musk-OpenAI trial testimony (Microsoft's $87B OpenAI infrastructure sink), Anthropic's March affidavit (~$3 spent per $1 earned on compute), the RPO concentration data (>50% of hyperscaler backlog from Anthropic + OpenAI alone), the enterprise [[ai-token-budget-explosion]] data, and the [[business-idiot]] structural argument. The thesis didn't change; the supporting numbers got worse.