Think the price you pay for AI tools is the real price? You're actually paying just 2-10% of the true cost. Investors are covering the rest. How long can this last?
What Is This About?
AI engineer Will Taubenheim published a 248,000-word research paper. The core finding: the entire AI industry is running on loss-making subsidies.
The numbers make it crystal clear. OpenAI generates $25 billion in annual revenue as of 2026, but inference costs alone eat up $14.1 billion per year. Add $6.7 billion in model development, $2.5 billion in compensation, and $13 billion owed to Microsoft, and they're still in the red.
For every $1 a business pays to use OpenAI's API, OpenAI spends $0.37 just on the electricity and hardware to generate the response. Research, salaries, and debt repayment are extra.
This isn't just an OpenAI problem. Anthropic disclosed in a court filing that it made over $5 billion in cumulative lifetime revenue while spending over $10 billion on inference and training combined. Cursor (Anysphere) burned through $3 billion in VC to generate $2 billion in annualized revenue.
As Ed Zitron put it: "Every single AI startup without exception does the same thing: turn hundreds of millions of dollars into tens of millions of dollars."
Why Should You Care?
The key issue is the gap between current prices and actual costs. According to ScaleDown's bottom-up analysis, the real cost of inference on H200 GPU servers is about $6.37 per million tokens, while GPT-4o-mini's API price is $0.60. That's roughly a 90% subsidy rate.
| Item | Current (Subsidized) | Post-Subsidy |
|---|---|---|
| ChatGPT Plus subscription | $20/month | $100-200+/month |
| Claude Code developer cost | $100-200/month | $1,000+/month expected |
| API per 1M tokens | $0.40-10 | $40-200 |
| Customer support AI (1M tickets/day) | Low millions/year | Tens of millions/year |
| AI coding tools (individual) | $10-39/month | $100-390+/month |
What makes it scarier is that usage is exploding simultaneously. A basic text response uses 0.3Wh, but reasoning mode takes 1.9Wh, and video generation consumes 1kWh — 3,333 times more power. Agents calling AI 10-20 times per task are sending token consumption through the roof.
Jevons' Paradox: As per-token costs drop, usage explodes and total spending actually increases. Uber initially subsidized rides by 59% to capture the market, then hiked prices 92% between 2018-2021.
Steve Smith (Ardalis) is even more direct: "By the end of 2027, agentic AI subscriptions like Claude Code and Copilot will increase by somewhere between 10x and 100x from their January 2026 levels."
How to Get Started
Here's how to survive even if AI prices jump 10x.
- Build an AI cost dashboard
Start tracking monthly spending per AI tool right now. Aggregate API call counts, token usage, and subscription fees into one view so you can instantly assess the impact of any price increase. - Run a 10x stress test
What happens to your business if AI costs multiply by 10? Audit your most AI-dependent workflows first and identify alternatives like local models and open-source options now. - Prefer flat-rate over per-use pricing
Per-token and per-action pricing models are ticking time bombs in the agentic era. Choose flat subscriptions when possible, and always set monthly spending caps on usage-based plans. - Match the model to the task
Using a top-tier model for simple autocomplete is like driving a truck to pick up groceries. Route simple tasks to smaller models and reserve premium models for complex reasoning. This alone can cut costs 50-67x. - Minimize vendor lock-in
Being dependent on a single AI provider means you're fully exposed to their price hikes. Build abstraction layers that let you switch flexibly between OpenAI, Anthropic, and open-source models.
Go Deeper
248,000-Word Full Report — The Foundations Are Cracking
Will Taubenheim's original report. Beyond inference subsidies, it covers the semiconductor supply chain (TSMC's 92% monopoly), the energy crisis (China building 37 reactors vs. US building zero), and labor market disruption across seven chapters with 35+ interactive data visualizations.
AI Benefits — But at What Cost? (Ardalis)
A deep dive into the actual financials of OpenAI, Anthropic, Cursor, Harvey, and Lovable. The pattern analysis of "every AI startup turning hundreds of millions into tens of millions" is particularly striking.
The Unsustainable Economics of LLM APIs (ScaleDown)
A technical bottom-up cost analysis starting from GPU server costs to calculate the true per-token price. Contains the math behind the 90% subsidy figure, plus strategic implications for business decision-makers.
The Real AI Cost Crisis (Appalach.AI)
Reframes the AI cost conversation: training costs are one-time, but inference costs are ongoing and growing. Connects hyperscaler $600B infrastructure spending with Jevons' Paradox.
