General5h ago

AI Profitability Despite High Costs

LessWrongMay 11, 20262 min brief

In brief

AI companies are burning through cash at alarming rates, with OpenAI alone spending $25 billion in the first half of 2025 while generating only $4 billion in revenue.
The cost of training advanced models continues to rise exponentially due to scaling laws, raising concerns about sustainability.
However, despite these high upfront costs, serving existing AI models is highly profitable.
For instance, DeepSeek-V4-Pro, a 1.6 trillion-parameter model, costs around $1.74 per million input tokens and $3.48 per million output tokens, yet leading labs charge significantly more-$2-5 per million input tokens and $12-25 per million output tokens.
The infrastructure expenses, such as building data centers, are substantial but necessary for future growth.
Even if funding dried up, AI would remain a profitable business due to the low marginal cost of serving API calls compared to their pricing.
While individual companies face pressure to constantly train new models to stay competitive, the industry as a whole can sustain itself by improving existing models at a steady pace.
Looking ahead, the key question is whether current scaling rates can be maintained.
If investment dries up, older models will continue to generate profits, potentially allowing incremental improvements to keep up with demand.
However, without sufficient funding for new models, competition could erode market share over time.

Terms in this brief

scaling laws: Rules that describe how the performance of AI models improves with their size and computational resources. These laws help predict costs and benefits as models grow, guiding investments in larger models for better performance.
marginal cost: The additional cost incurred by producing one more unit of a product. In AI, it refers to the cost of generating each extra output token after initial model training has been completed.

Read full story at LessWrong →

More briefs