AI (artificial intelligence) is the fastest-growing expense in industry technology budgets today, with it consuming up to nearly half the IT spend in some organisations.
As generative AI (GenAI) becomes central to operations, cloud computing bills are also rising sharply, up to as much as 19% for many organisations in 2025. Even then, returns can remain elusive. According to the Deloitte US Tech Value survey for 2025, only 28% global finance leaders have reported measurable value from their AI investments, and nearly 50% expect it to take up to 3 years to see any return on investments from basic AI automation.
This disconnect is more than just a financial headache; rather, it’s a strategic reckoning. For many enterprises, adopting AI is more about maintaining competitive advantage and staving off existential threats, and less about immediate returns. As such, the new economic reality is defined by nonlinear, volatile dynamics and not necessarily by traditional metrics.
So, the focus is now shifting from whether or not AI delivers value to how its economics are managed and measured for enterprises to thrive in a structurally different environment. Amidst this, a fundamental concept has emerged that defines operational capacity, cost, and value: the AI token.
The Anatomy of an AI Token: The True Currency of AI
Firstly, AI tokens shouldn’t be confused with cryptocurrency tokens, such as bitcoin, or model parameters. They’re actually tiny units of data that are generated when larger chunks of information are broken down. These tokens are processed by AI models that learn the relationships between them, unlocking capabilities including reasoning, generation, and prediction.
The faster the tokens can be processed, the faster these models can analyse, understand, and respond. The objective is to achieve the lowest cost per token and the fastest processing time to optimise AI infrastructure, thus maximising revenue generation.
Unlike earlier, when costs were tied to virtual machines or subscriptions, AI economics now revolves around tokens, which are now widely considered as the fundamental unit of AI work. Every interaction, from training AI models to inferences, is measured in tokens, thus making costs inherently variable — and often unpredictable. This volatility is a result of variable pricing, fluctuating usage, and non-linear demand.
Contrary to what you might think, a single AI token doesn’t always equate to a single word. At the core of LLMs (Large Language Models), tokenization is breaking down text into sub-word units to process them efficiently. So, a word such as “decentralisation” could be broken down into multiple tokens (“de,” “central,” “isation”).
On the other hand, shorter, commoner words such as “and” or “the” could be single tokens. What makes this granularity crucial is that AI models operate not at the word level, but at the token level.
How Do Tokens Drive AI Economics?
When the pre-training and post-training process takes places, tokens are equated to “investment into intelligence.” When the inference happens, they drive revenue and cost. So, with AI applications proliferating, newer principles of AI economics are emerging.
AI models are designed to sustain high-volume inference, generating intelligence for users by turning tokens into monetizable analysis. That’s why an increasing number of AI services are measuring their model values based on how many tokens are being consumed and generated, thus offering pricing plans based on the rates of token inputs and outputs for individual models.
Furthermore, some token pricing plans offer a set number of tokens to their users, which are shared between inputs and outputs. Based on these limits, users could use short text prompts (using just a few tokens) as inputs to generate lengthy, AI-generated responses that consumed thousands of tokens for the output. Or, they could spend a major chunk of their tokens on inputs, providing AI models with a set of documents to be summarised into just a few bullet points.
Secondly, some AI services also set token limits, which is the maximum number of tokens an individual user generates per minute, to serve a high volume of concurrent users. Next, tokens also define user experiences for AI services. Token-to-token or inter-token latency (the rate at which subsequent output tokens are generated) and time to first token (latency between users submitting prompts and AI models responding) determine how end users experience the outputs of AI applications.
So where does economics come in? For one, every metric has trade-offs, and the right balance differs by use case. For instance, shortening the time to the first token in LLM-based chatbots could help improve user engagement by keeping a conversational pace without any odd pauses. Moreover, optimising inter-token latency could allow video-generating models to achieve desired frame rates or text generation models to match average reading speeds.
When it comes to AI models engaging in research and long thinking, the emphasis is on generating high-quality tokens, even if latency increases.
AI – An Economic System?
In this entire system, there has arisen a paradox: one of rising consumption and falling prices. When AI tokens’ unit prices fall, the enterprise’s overall spending on and scaling AI systems is increasing. With more intense workloads, complex models, and an increasing number of users, token consumption will be greater, and, consequently, costs will be higher.
So, AI can’t and shouldn’t be managed with outdated cost models. Rather, AI tokens need to be recognised as the new digital currency, working alongside FinOps and a hybrid infrastructure, and allowing enterprises to deploy workloads where they can be most strategically and economically advantageous. Fluency in token economics will ultimately dictate the efficiency, scalability, and profitability of the most advanced solutions.
Leave a comment