AI token pricing varies wildly based on hardware, utilization, and model—making per-token cost a poor metric for business value

US private AI investment reached $285.9 billion in 2025, according to Stanford HAI's 2026 Artificial Intelligence Index Report. AI data center power capacity rose to 29.6 GW, comparable to New York state at peak demand, and annual GPT-4o inference water use alone may exceed the drinking water needs of 12 million people.

Token pricing depends on multiple variables: inference on an Nvidia H100 GPU costs roughly $0.0038 per million tokens at 100 percent utilization ($2.50/hour rental, 185 tokens/second), but at 30 percent utilization that rises to ~$0.013/M tokens and at 10 percent to ~$0.038/M tokens. Anthropic charges $5/M tokens (input) and $25/M tokens (output) for Opus 4.7; Google's Gemma 4 26B A4B has a weighted average input price of $0.096/M tokens per OpenRouter.

Machine learning researcher Devansh argues token spend is not directly correlated with productivity and that tracking token usage as a key performance indicator repeats past mistakes of measuring lines of code or words typed. Companies like Meta and Shopify have highlighted token usage, but employees optimizing for token volume may not deliver meaningful business value.