# Relearning cloud lessons from runaway AI token costs > Source: > Published: 2026-07-10 09:00:00+00:00 Every few years, some new technology comes along that promises to revolutionize how we do business, and enterprises pile in headfirst without asking how much it’s going to cost. I’ve been watching this movie for 30 years. [Cloud computing](https://www.infoworld.com/article/2238873/what-is-cloud-computing.html) was the first act. Now it’s [generative AI](https://www.infoworld.com/article/2338115/what-is-generative-ai-artificial-intelligence-that-creates.html), and the bill is arriving faster than anyone expected. The latest data shows that many enterprises are seeing their AI token costs run 10 to 20 times higher than initial projections. That’s not a rounding error. That’s a strategic miscalculation that CFOs are starting to notice, and they’re not happy about it. Here’s the thing: This crisis was entirely predictable. We’ve been through this before with cloud computing, and we learned some hard lessons about what happens when you deploy technology without rigorous cost management. The good news is that enterprises are finally applying those lessons, reaching back to their cloud finops playbooks to wrangle this new breed of spending. Let me explain the scale of what’s happening. Goldman Sachs has estimated that [AI agents](https://www.infoworld.com/article/3611465/how-ai-agents-will-transform-the-future-of-work.html) consume roughly 50 times more computing power per task than traditional prompt-based chatbots. That’s a fundamental shift in how resources get consumed. When you multiply that across an enterprise that’s deploying dozens or hundreds of AI agents, the math gets ugly fast. The token problem compounds because AI costs are inherently variable. Unlike traditional software licensing or infrastructure contracts, you pay per token, and per-token usage can fluctuate wildly based on user behavior, query complexity, and the sheer volume of requests flowing through these systems. This is exactly the same problem we faced with cloud computing. Every time someone spins up a new instance or stores data in the wrong tier, the bill goes up. Enterprises expected to deploy AI and see costs stabilize. Instead, costs are climbing month after month, often exceeding projections by an order of magnitude. The business case that looked compelling in the conference room is looking considerably less attractive in the finance committee. Here’s where it gets interesting. Cloud providers and the managed service providers who work with them have spent the better part of two decades building disciplines around financial operations—[finops](https://www.infoworld.com/article/2338592/6-finops-best-practices-to-reduce-cloud-costs.html), if you want to use the buzzword. These are the practices, tools, and organizational structures that make cloud spending visible, controllable, and ultimately justifiable to the business. Those same disciplines are now being applied to AI token costs, and enterprises with mature finops programs are faring better than those without. The playbook is essentially the same: Companies like Priceline have deployed dashboards that provide executives with real-time visibility into token consumption, with monthly reports delivered directly to the CFO and CTO. Smartsheet has implemented similar approaches, providing department-level dashboards that let managers see exactly how their teams are consuming tokens, with automated alerts when consumption approaches predefined thresholds. The accountability piece is critical. When developers and business users can see exactly how their AI usage translates to dollars, they tend to make better decisions about which models to use, how to structure prompts, and when to rely on human judgment instead of AI processing. One of the most effective techniques emerging from this crisis is the “show back” approach to AI cost management. Rather than simply reporting costs to individual departments, companies are now attributing AI spending to the teams and individuals responsible for driving that consumption. This creates accountability without the organizational complexity of full chargeback models. OpenText has reported that implementing show-back and chargeback approaches can reduce token costs by 20% to 30% within a few months. That’s not trivial. If you’re spending $5 million a month on AI tokens, that’s a $1.5 million savings just by making people aware of what they’re spending. The mechanism is straightforward: When development leaders understand that their team has consumed $200,000 in tokens this month, they start asking questions. Why are we using the most expensive model for that task? What if a smaller model could handle 80% of these queries? Are prompts being repeated unnecessarily? These questions lead to optimization, and optimization leads to savings. Another lesson from the cloud experience is that the most expensive option is rarely the best option. This sounds obvious, but organizations tend to default to the [largest, most capable AI model](https://www.infoworld.com/article/2335213/large-language-models-the-foundations-of-generative-ai.html) for every task, regardless of whether that capability is actually required. The emerging best practice is to match model capability to task requirements. A simple classification task doesn’t need a frontier model. A straightforward text-generation job might be handled perfectly by a smaller, cheaper model running locally or via a less expensive API tier. The efficiency gains from this approach can be substantial. Some enterprises are going further, adopting older models or open source alternatives for appropriate use cases. Qualcomm, for instance, has invested in running models on its own hardware rather than relying exclusively on cloud-based model providers. This approach requires more technical sophistication but can dramatically reduce per-token costs for high-volume applications. Here’s what concerns me most about the current situation. Many enterprises deployed [AI](https://www.infoworld.com/article/4061121/a-brief-history-of-ai.html) without putting adequate cost management infrastructure in place up front. They got caught up in the excitement of the technology, the competitive pressure to move fast, and the belief that the benefits would justify whatever the costs turned out to be. That approach worked when AI projects were small-scale experiments. Now that AI is becoming core to business operations, the lack of financial controls is becoming a serious problem. We need to bring the same rigor to AI procurement and deployment that we’ve brought to every other significant technology investment. The organizations that succeed will treat AI token costs as a managed operational expense rather than an unpredictable variable. That means deploying the same tools and disciplines that have worked for cloud cost management: visibility, accountability, optimization, and continuous improvement. Cloud providers and the managed service partners who work with them have been doing this for years. They built the tools, developed the best practices, and trained the workforce that can now apply those skills to the AI cost challenge. If your organization is struggling with AI spending, finding partners with deep finops experience might be the fastest path to control. The good news is that this crisis is solvable. But it requires acknowledging the problem, investing in the right capabilities, and accepting that technology deployment without financial discipline is a path to trouble. Get smart about your AI spending. The CFO will thank you.