# AI Subsidy Economics

The generative AI industry as of early 2026 runs on subsidies at every level. No link in the chain between GPU manufacturing and end users is sustainably profitable except NVIDIA. The spending is large enough to distort global hardware markets — [[hold-on-to-your-hardware]] documents how AI data center demand has consumed 40% of global DRAM output, sold out WD's entire 2026 HDD production, and pushed consumer hardware into shortage.

The subsidy takes different forms depending on where you sit:

**AI labs** (Anthropic, OpenAI) spend roughly $2 on compute for every $1 of revenue. Anthropic: $5B revenue vs $10B compute. OpenAI: $4.3B revenue vs $8.67B inference cost through September 2025. Both fund the gap with VC and debt — Anthropic raised $30B in February 2026 alone. Their consumer subscriptions ($20-$200/mo) allow users to burn 3-13x their subscription cost in tokens. The subsidy is hidden behind opaque rate limits and percentage gauges rather than transparent token metering.

**AI startups** (Cursor, Perplexity, Harvey, Replit, Lovable) buy API access from labs and resell it via subscriptions. Every one is unprofitable. Cursor raised $3.36B and generated roughly $1B in revenue. Harvey has $190M ARR on an $11B valuation. These companies survive on VC money that is itself facing a historic liquidity crisis.

**Data center developers** borrow heavily ($178.5B in US debt in 2025) to build GPU capacity. CoreWeave, the largest independent, posted -29% net margin in 2025 despite having Microsoft, OpenAI, and NVIDIA as customers. Of 200GW of announced capacity, only 5GW is under construction.

The structural problem is that flat-rate subscriptions trained users to ignore token costs entirely. When labs tighten limits — as Anthropic did with peak hours in March 2026 — users react with fury, not understanding why a $200/mo subscription can't sustain $2,000/mo of compute. But if labs charged real costs, demand would evaporate, because even at subsidized prices, AI hasn't demonstrated productivity gains that justify 10x higher pricing.

This creates what [[subprime-ai-crisis|Zitron calls]] a "variable rate mortgage" dynamic: new models burn more tokens (chain-of-thought, coding agents, deep research), so costs rise invisibly even at the same subscription price. Users demand access to the newest models. The cost adjusts upward but the price label doesn't — until it must.

See [[ai-bubble-pale-horses]] for the specific warning signs that the correction is underway. For the case that local open-source models are the rational exit from this subsidy chain, see [[titit-local-ai]].