$5 billion. That's approximately how much OpenAI lost in 2024, against revenues of roughly $3.7 billion. Anthropic has raised over $7 billion from Amazon and Google and still hasn't turned a profit. Every major AI lab is currently selling compute at prices that don't reflect what it actually costs to run their models - and the gap between what users pay and what inference actually costs has become one of the defining tensions in the industry.
The standard defense is that token pricing covers ongoing inference costs even if it doesn't recover the billions spent training the model. That argument holds up when usage is relatively low. It gets shakier as AI becomes the default for tasks that used to take human hours: document summaries, data analysis, email drafting, code review. Every one of those tasks runs on expensive GPUs. Unlike traditional software - where distributing to a million users costs barely more than distributing to a hundred - AI inference scales linearly with usage. More tasks means more compute means more cost.
Researchers who've tried to estimate real consumer AI task costs put typical subsidization at somewhere between 5x and 20x, depending on the model and complexity. An agentic task - where the model breaks a problem into steps, calls external tools, generates multiple outputs, and revises its work - costs far more than a single short chat response. The $20/month that feels expensive for a productivity app often covers a fraction of what's actually being consumed.
What Repricing Already Looks Like
The correction is happening, just slowly. ChatGPT Pro jumped from $20 to $200/month for the top tier. API pricing has shifted across models, with efficiency gains making some cheaper while the most capable models stay expensive. Enterprise contracts increasingly include usage caps and overage charges rather than flat unlimited rates.
The consumer tier - unlimited usage for $20/month - is becoming a funnel to premium tiers rather than a sustainable product in its own right. Both OpenAI and Anthropic understand that the path to profitability runs through enterprise contracts with usage-based billing, not consumer subscriptions priced to drive adoption.
The Practical Question
For anyone building workflows on top of AI tools, the useful exercise is separating the tasks that justify their cost at current subsidized prices from the ones that would justify their cost at 2x pricing. Those are different lists.
The workflows with a real answer to "is this worth paying for at actual cost" are the ones worth building habits around. The use cases that only work because AI is currently cheap - and there are more of these than most people admit - are worth treating as temporary advantages, not permanent features of the landscape.
This isn't a prediction of collapse. It's a description of a business model running on investor capital rather than unit economics, and an observation that the gap between what AI costs and what users pay will close in one direction or another.