$10 billion spent on compute against $5 billion in revenue. That's Anthropic's ratio, and it's not the worst in the industry. OpenAI burned $8.67 billion on inference alone (the cost of actually running models to serve user requests) against $4.3 billion in revenue through September 2025. These aren't early-stage losses on the path to scale. They're structural deficits baked into how AI products are priced.
The core problem is simple: AI companies have trained users to expect near-unlimited access to expensive compute for flat monthly fees. A Claude Pro subscriber paying $20 per month could historically burn through thousands of dollars in actual compute costs. One Augment Code user reportedly consumed $15,000 in tokens on a $250-a-month plan. These aren't edge cases. They're the business model.
The Subsidy Chain Is Cracking
The cracks started showing in mid-2025. Anthropic and OpenAI launched "priority service tiers" requiring enterprise prepayments and raising API costs on downstream startups. By February 2026, Perplexity slashed deep research queries from 600 per month to 20 for Pro subscribers. In March 2026, Anthropic introduced "peak hours" restrictions so aggressive that users reported hitting weekly limits in 15 minutes.
The pattern is consistent across the industry. Replit, Augment Code, and others have implemented what amounts to austerity pricing - charging more, delivering less, and confusing users in the process.
Downstream AI startups are caught in the same trap. Cursor raised $3.36 billion but converts to roughly $1-2 billion in revenue. Harvey sits at an $11 billion valuation against $190 million in annual recurring revenue. MiniMax generated $79 million in 2025 while losing $250.9 million. None of these companies are profitable.
The Infrastructure Debt Problem
Behind the software layer sits a physical infrastructure bet that makes the financial picture worse. Data center developers borrowed $178.5 billion in the U.S. alone in 2025. But of the 200 gigawatts of capacity announced, only 5 gigawatts is actually under construction. Most of those loans are interest-only, secured against projects that exist primarily on paper.
CoreWeave, the largest AI compute provider and a key NVIDIA partner, posted a -6% operating margin and -29% net loss margin in 2025. This is the company that major AI labs depend on for GPU capacity.
The analogy to 2008's mortgage crisis has obvious limits - there's no securitization chain turning bad AI bets into AAA-rated financial products. But the structural similarity is real: an entire industry priced its products below cost, used cheap capital to cover the gap, and now faces the question of what happens when the subsidy runs out.
What This Means for Users
The practical implication for anyone who depends on AI tools is straightforward: the product you're using today is probably not priced at what it actually costs to run. When pricing adjusts - and the rate limit changes at Anthropic, OpenAI, and Perplexity suggest it already is - the tools will either get more expensive, more restricted, or both.
Building critical workflows around $20/month AI subscriptions without a backup plan is a risk. Not because the technology doesn't work, but because the economics behind it haven't been solved yet. The companies burning $3-10 for every dollar of revenue can't do it forever, and when they stop, users will feel it first.