AI Tokens Are Inflating Enterprise Cloud Bills, ZDNet Reports

Enterprise spending on AI tokens is turning into a new budget line that nobody knows how to control. ZDNet analysts flag a worrying trend: the cost curve mirrors the early history of the cloud, with rapid growth, no clear optimization tools and no answer to what exactly the business gets in return.

Déjà vu from 2012

In the early cloud era, IT directors received bills for EC2, S3 and RDS that grew faster than any forecast. Nobody could calculate cost per unit of business value. It took years, dozens of FinOps tools and entire cloud cost management departments to bring those costs under control.

With AI tokens, the story is repeating at a compressed pace. Developers are plugging GPT-4o, Claude Opus and their own fine-tuned models into product systems. Each agentic workflow with tool calls, RAG search and long context can spend from 20,000 to 100,000 tokens per transaction. Multiply that by thousands of users per day and you get a bill that appeared in no budget and keeps growing every quarter.

"AI tokens will remind many enterprise customers of early cloud pricing,"

ZDNet states.

The unsolved problem: how to measure value

The main problem ZDNet identifies is not the spending itself but the fact that companies cannot measure the value AI tokens create. In the cloud world, things were relatively transparent: an instance performs a task, the task has a cost in person-hours, and the difference is savings. With AI, that logic breaks down at every step:

• An AI assistant speeds up writing an email, but nobody has systematically measured by how much
• An agent automates a process, but the quality of its work is subjective and inconsistent
• A chatbot offloads support, but satisfaction metrics in complex cases are falling
• A code assistant reduces development time, but technical debt does not go away
• A RAG system improves answer accuracy, but benchmarks depend on the dataset and the specific task

Unlike CPU-hours or gigabytes of traffic, tokens have no obvious link to a business outcome. A CFO cannot say: "For this million tokens, we got this specific, measurable value." Until this problem is solved, AI budgets will be approved on faith rather than data.

How the market is responding

Some large enterprise customers of AWS, Azure and Google Cloud already report AI spending growing 3-10 times year over year. Providers are responding with tools for monitoring token consumption, but so far these mostly display numbers rather than help optimize costs. A new specialization is emerging, AI FinOps, whose job is to manage the cost of LLM inference in production.

Among the first tactical tools:

• Prompt caching for repeated requests
• Routing tasks to cheaper models depending on complexity
• Limiting context depth and the number of steps in agentic chains
• Batching requests instead of single real-time calls
• Regular audits of unused or low-value AI integrations

But this is tactics, not strategy. The industry does not yet have an answer to the question of how many tokens it is worth spending to achieve a specific business result.

What this means

Enterprise AI is walking into the same trap the cloud did in 2012: the technology is widely adopted, costs are rising fast, and ROI is hard to measure. Companies that have already deployed LLMs in production should invest in cost visibility tools and build value metrics right now. Otherwise, the conversation with the finance director will turn into an uncomfortable exam nobody prepared for.