First, What Are AI Tokens Anyway?
Think of AI tokens as the currency of intelligence. Every time you ask a generative AI model a question, you're using tokens. A token is a piece of data, roughly equivalent to about three-quarters of a word in English. When you send a prompt to a model like
Anthropic's Claude or one of Amazon's own Nova models, the system breaks your request down into tokens, processes them, and then generates an answer, which is also measured in tokens. The input tokens and output tokens together determine the cost of that single interaction. For businesses building AI features—from a chatbot that answers customer questions to a tool that summarizes legal documents—these token costs can add up quickly, making them a critical metric to watch.
A Race to the Bottom on Price
Over the past couple of years, the price of these tokens has been in a freefall. Fierce competition among the major cloud providers—AWS, Microsoft Azure, and Google Cloud—is a huge factor. Each is fighting to become the go-to platform for developers building with AI, and aggressive pricing is a key weapon. Furthermore, the AI models themselves are becoming more efficient. Newer models from companies like Anthropic and Meta can perform complex tasks using less computational power, allowing providers like AWS to offer them at a lower cost. For instance, some tasks that once required a premium, expensive model can now be handled by a cheaper, faster one, driving the average cost per task down significantly. It's a classic technology deflationary cycle, similar to how computer storage and processing power have become exponentially cheaper over decades.
The Demand Explosion
This price drop is having a predictable effect on demand, a phenomenon known as the Jevons paradox: as a resource becomes cheaper, we tend to use more of it. With lower token costs, the barrier to entry for building AI-powered applications has fallen dramatically. Startups and large enterprises alike are now embedding AI into more products and workflows. A task that was once too expensive to automate, like generating personalized marketing copy for thousands of customers, is now economically viable. This has led to an explosion in usage. Reports suggest that total token consumption is multiplying at a staggering rate as companies find more ways to leverage cheaper intelligence. The result is that while the cost per token is falling, the total number of tokens being consumed is skyrocketing.
The Profit Paradox
Herein lies the paradox for AWS. More customers and more usage should equal more profit, but it’s not that simple. The business of selling raw AI model access is becoming a low-margin, high-volume game. Competing purely on token price is a race to the bottom that can squeeze profits, especially when the underlying hardware (like GPUs) remains incredibly expensive. However, AWS is playing a longer game. The real strategy isn't to get rich selling tokens. It's to use cheap tokens as a gateway to lock customers into its broader, highly profitable ecosystem. Once a developer builds their AI application on AWS Bedrock, they are also likely to use AWS for data storage (S3), databases (DynamoDB), and other cloud services that carry much healthier margins. By making the entry point—AI inference—as affordable as possible, AWS ensures it wins the platform war. The profit doesn't come from the token itself, but from all the other services sold around it.
