Why Compute Costs Could Become 10x More Expensive in the Coming Years

For AI labs' revenue to grow explosively, the value of compute resources must rise sharply. Currently, companies like Anthropic are growing their revenue 10x every year, and sustaining this trend requires either rising compute costs or expanding margins.

At present, AI labs' compute spending is increasing 3x every year. For revenue to grow much faster than compute spending, one of the following three things must happen.

• An increase in labs' Margin
• A rise in Compute prices
• An expansion of the Inference share of compute spending

In practice, labs view compute costs not simply as operating expenses, but as investment assets for training even larger models. In particular, as models become smarter, they can generate more revenue from the same compute resources, which becomes a key driver of rising compute value.

For example, if an AI with H100-level performance could perform the role of a human-level software engineer, that GPU's value would be far higher than its current market price—exceeding $250,000 per year. This is roughly 15 times the current spot price.

Large tech companies are already paying far more than spot prices to secure compute. The cost Google pays to rent GPUs from SpaceX is about twice the spot price, showing that the cost of securing stable infrastructure is already rising.