Cost to run · explainer
Why AI costs less to run than you pay
The cost-to-serve calculator shows a query's compute cost sitting far below its API price — yet the companies selling those queries report billions in losses. That is the profitability paradox of a high-fixed-cost, low-marginal-cost business, and these three questions unpack it.
Why is the price so much higher than the compute cost?
The gap — often roughly 10–20× — is not pure profit. The per-query compute cost is the marginal cost of running your request. It excludes the enormous fixed costs behind the model: training, research and salaries, and the free tiers that most users never pay for.
Why do AI companies still lose money if each query is profitable?
This is the paradox of high-fixed-cost, low-marginal-cost businesses (like pharma or early SaaS): each query more than covers its own compute, yet the company loses money because revenue doesn't yet cover the fixed costs. Training a frontier model is estimated in the tens to hundreds of millions (Epoch AI ~$40M for GPT-4 amortized; Stanford HAI ~$78M by a cloud-rental method), with frontier runs projected toward $1B by 2027. And providers report large losses — OpenAI reportedly ~$5B in 2024 and a reported ~$21B operating loss in 2025 (leaked, not audited).
Will AI get cheaper?
Probably, at the compute level: inference cost for a fixed level of capability has been falling roughly 10× per year (Epoch AI), as GPUs, quantization, and serving efficiency improve. Whether prices fall depends on competition and whether providers keep subsidizing for adoption.
For how the compute cost itself is modeled — the rental formula and the per-GPU throughput caveat — see the cost methodology. This tool's sibling estimates the environmental cost of the same queries.