For five straight years, enterprises got better at managing their cloud budgets. Waste went down. Optimization improved. Finance teams and engineering departments found a fragile but real détente over how much to spend on infrastructure and how much of that spending actually produced value. That streak is over.
A new report from TechRadar, drawing on findings from HashiCorp’s 2025 State of Cloud Strategy Survey, reveals that AI workloads have reversed the long-running trend of declining cloud waste. For the first time since 2020, organizations are reporting that their wasted cloud spend is climbing — and the culprit isn’t traditional compute or storage sprawl. It’s artificial intelligence.
The numbers are striking. According to HashiCorp’s survey of more than 3,400 respondents across multiple industries, 57% of enterprises say they have increased their cloud spending specifically to support AI workloads. Meanwhile, estimated cloud waste has ticked upward after years of steady decline. The correlation isn’t subtle. Companies are throwing money at GPU instances, large-scale training jobs, inference endpoints, and vector databases — often without the kind of cost governance frameworks that matured around conventional cloud workloads over the past decade.
This isn’t a small-dollar problem. Gartner has projected that worldwide public cloud end-user spending will exceed $723 billion in 2025. Even a few percentage points of waste across that total represents tens of billions of dollars evaporating into underutilized infrastructure, orphaned resources, and poorly scoped AI experiments that never make it to production.
So what happened?
The short answer: AI workloads behave differently than the workloads enterprises spent years learning to optimize. Traditional cloud cost management relied on right-sizing virtual machines, purchasing reserved instances, shutting down idle resources, and building internal chargeback models. Those techniques work when workloads are relatively predictable — web servers, databases, batch processing jobs with known resource profiles. AI workloads don’t fit that mold. Training a large language model can consume thousands of GPU-hours in bursts that are difficult to forecast. Inference workloads scale unpredictably based on user demand. Data pipelines feeding AI systems often require expensive high-throughput storage and networking configurations that sit idle between runs.
And then there’s the experimentation problem. Many organizations are still in the early stages of figuring out which AI initiatives will generate real business value. That means dozens — sometimes hundreds — of pilot projects running simultaneously, each provisioning cloud resources that may never be reclaimed once the experiment concludes or the team moves on to the next idea. It’s the cloud equivalent of leaving every light in the house on because you’re not sure which room you’ll walk into next.
The HashiCorp report highlights a potential remedy: dedicated AI governance teams. According to the survey data cited by TechRadar, organizations that have established formal governance structures around their AI initiatives report significantly better control over cloud costs. These teams typically sit at the intersection of finance, engineering, and data science, with a mandate to evaluate AI projects not just on technical merit but on resource efficiency and return on investment.
The idea isn’t new in principle. FinOps — the practice of bringing financial accountability to cloud spending — has been a growing discipline for years, championed by the FinOps Foundation and adopted by major enterprises worldwide. But the AI wave has exposed gaps in existing FinOps practices. Most FinOps tooling and processes were designed for CPU-based workloads with well-understood pricing models. GPU pricing is different. Spot instance availability for AI-class hardware is more volatile. And the cost structures of managed AI services from AWS, Azure, and Google Cloud don’t always map cleanly onto the frameworks FinOps teams have built.
That’s the governance gap AI teams are meant to fill.
Consider the scale of the shift. A single training run for a frontier-class large language model can cost millions of dollars in compute alone. Even smaller fine-tuning jobs on enterprise-specific data can run into the tens or hundreds of thousands. When those jobs are initiated by data scientists who have deep expertise in model architecture but limited visibility into cloud billing, costs escalate fast. A governance team with cross-functional authority can impose guardrails — budget caps on experiments, automated shutdown policies for idle GPU clusters, mandatory cost reviews before scaling training jobs — without slowing down legitimate innovation.
Some organizations are already moving in this direction. According to recent reporting, companies including major financial institutions and healthcare systems have begun embedding cost-awareness metrics directly into their AI development pipelines. The goal is to make cost a first-class consideration alongside accuracy, latency, and fairness — not an afterthought discovered when the monthly cloud bill arrives.
But governance alone won’t solve the problem if the underlying infrastructure strategy is flawed. The HashiCorp survey also found that multi-cloud adoption continues to grow, with 76% of respondents operating across two or more cloud providers. Multi-cloud introduces its own cost complexities: different pricing models, different discount structures, different monitoring tools. Add AI workloads on top, and the management burden multiplies. An organization running training jobs on AWS while serving inference from Google Cloud and storing data in Azure faces three separate billing systems, three separate sets of reserved capacity options, and three separate cost optimization toolkits. Without a unified strategy, waste is almost inevitable.
Infrastructure-as-code tools — HashiCorp’s core business, it should be noted — offer one path toward consistency. By codifying cloud resource provisioning in version-controlled templates, organizations can enforce standardized configurations across providers and workload types. This makes it easier to audit what’s been deployed, identify resources that have drifted from their intended specifications, and tear down environments that are no longer needed. It’s not a silver bullet. But it’s a necessary foundation.
The timing of all this matters. Enterprise AI spending is accelerating even as macroeconomic pressures push CFOs to scrutinize every line item. That tension — spend more on AI to stay competitive, spend less overall to protect margins — creates an environment where waste isn’t just an operational nuisance. It’s a strategic liability. Boards and investors are asking hard questions about AI ROI. If a company can’t demonstrate that its AI cloud spending is generating proportionate value, the backlash could be severe, leading not just to budget cuts but to organizational skepticism about AI initiatives broadly.
There’s a historical parallel worth considering. In the early 2010s, the first wave of cloud migration produced a similar pattern: enterprises rushed to move workloads to AWS, Azure, and Google Cloud without fully understanding the cost implications. Shadow IT proliferated. Departments spun up resources without central oversight. Cloud bills ballooned. It took years — and the emergence of dedicated cloud cost management vendors like CloudHealth, Spot.io, and Apptio — to bring discipline to the process. AI is replaying that cycle, compressed into a shorter timeframe and with higher stakes.
The question is whether enterprises will learn faster this time.
Early signs are mixed. On the positive side, awareness of the problem is high. The HashiCorp survey data suggests that most organizations recognize AI is driving up their cloud costs and are actively seeking solutions. AI governance teams are one response. Improved tagging and attribution of AI-specific resources is another. Some companies are experimenting with internal AI marketplaces where teams must “purchase” compute time from a centralized budget, creating natural cost pressure.
On the negative side, the pace of AI adoption often outstrips the pace of governance implementation. New models, new frameworks, new use cases emerge weekly. By the time a governance team has established policies for one class of AI workload, three more have appeared. The result is a perpetual game of catch-up — and in the meantime, the meter keeps running.
There’s also a talent dimension. Building an effective AI governance function requires people who understand machine learning operations, cloud infrastructure economics, and corporate finance. That Venn diagram doesn’t describe a large talent pool. Organizations that can’t hire or develop these cross-disciplinary professionals will struggle to implement governance frameworks that actually work, regardless of how many policies they write.
Cloud providers themselves have a role to play, and they know it. AWS, Microsoft, and Google have all introduced AI-specific cost management features in recent months — usage dashboards tailored to GPU workloads, savings plans designed for inference endpoints, recommendations engines that suggest cheaper model-serving configurations. These tools help. But they’re also self-interested: cloud providers benefit when customers spend more, not less. The optimization tools they offer tend to focus on efficiency within their own platform, not across the multi-cloud environments where most enterprises actually operate.
Independent vendors are stepping into that gap. Companies like Vantage, Kubecost, and Anodot have expanded their platforms to address AI-specific cost visibility. Some newer startups are building tools specifically for tracking the cost of model training, comparing the economics of different serving strategies, and forecasting AI infrastructure budgets based on planned development roadmaps. It’s a young market with no dominant player yet. But the demand signal is unmistakable.
For industry insiders watching this space, several things are worth tracking in the months ahead. First, whether the cloud waste trend identified in the HashiCorp report continues to worsen or stabilizes as governance practices catch up. Second, how the major cloud providers adjust their pricing and discount structures in response to growing AI cost sensitivity among enterprise customers. Third, whether the AI governance team model — a dedicated, cross-functional group with budget authority — becomes a standard organizational pattern or remains an ad hoc response adopted by a minority of firms.
And finally, there’s the fundamental question that underlies all of this: how much of today’s AI cloud spending is genuinely productive, and how much is speculative — companies paying for infrastructure to support AI initiatives whose business value has not been demonstrated and may never materialize? That question doesn’t have a clean answer yet. It might not for years. But the organizations that develop rigorous frameworks for answering it — governance teams, cost attribution systems, ROI measurement methodologies — will be the ones that turn AI from an expense line into a competitive advantage.
The rest will just be burning cash in the cloud.
WebProNews is an iEntry Publication