Despite rapid adoption, most enterprises lack visibility, forecasting accuracy, and margin control around AI investments. Hidden infrastructure costs are eroding enterprise profitability, according to newly published survey data.

AUSTIN, Texas, Sept. 10, 2025 /PRNewswire/ — As enterprises accelerate investments in AI infrastructure, a new report reveals a troubling financial reality: most organizations can’t forecast what they’re spending, or control how AI costs impact margins. According to the 2025 State of AI Cost Management, 80% of enterprises miss their AI infrastructure forecasts by more than 25%, and 84% report significant gross margin erosion tied to AI workloads.

The report, published by Benchmarkit in partnership with cost governance platform Mavvrik, reveals how AI adoption, across large language models (LLMs), GPU-based compute, and AI-native services, is outpacing cost governance. Most companies lack the visibility, attribution, and forecasting precision to understand where costs come from or how they affect margins.

“These numbers should rattle every finance leader. AI is no longer just experimental – it’s hitting gross margins, and most companies can’t even predict the impact,” said Ray Rike, CEO of Benchmarkit. “Without financial governance, you’re not scaling AI. You’re gambling with profitability.”

Top Findings from the 2025 State of AI Cost Management Report include:

AI costs are crushing enterprise margins

• 84% of companies see 6%+ gross margin erosion due to AI infrastructure costs
• 26% report margin impact of 16% or higher

The great AI repatriation has begun

• 67% are actively planning to repatriate AI workloads; another 19% are evaluating
• 61% already run hybrid AI infrastructure (public + private)
• Only 35% include on-prem AI costs in reporting, leaving major blind spots

Hidden cost surprises come from unexpected places

• Data platforms top source of unexpected AI spend (56%); LLMs rank 5th
• Network access costs is the second-largest cost surprise (52%)

AI forecasting is fundamentally broken

• 80% miss AI forecasts by 25%+
• 24% are off by 50% or more
• Only 15% forecast AI costs within 10% margin of error

Visibility gaps are stalling governance

• Lack of visibility is the #1 challenge in managing AI infrastructure costs
• 94% say they track costs, but only 34% have mature cost management
• Companies charging for AI show 2x greater cost maturity in attribution and cost discipline

Access the full report: The full report details how automation, cost attribution methods, and cloud repatriation strategies factor into AI cost discipline. To view the analysis, please visit: https://www.mavvrik.ai/state-of-ai-cost-governance-report/

“AI is blowing up the assumptions baked into budgets. What used to be predictable, is now elastic and expensive,” said Sundeep Goel, CEO of Mavvrik. “This shift doesn’t just affect IT, it’s reshaping cost models, margin structures, and how companies scale. Enterprises are racing to build with AI, but when most can’t explain the bill, it’s no longer innovation, it’s risk.”

Why It Matters

AI isn’t just a technology challenge, it’s a financial one. From LLM APIs to GPU usage and data movement, infrastructure costs are scaling faster than most companies can track them. Without clear attribution across cloud and on-prem environments, leaders are making pricing, packaging, and investment decisions in the dark.

With AI spend becoming a significant line in COGS and gross margin targets under pressure, CFOs should be sounding the alarm. Yet most finance teams haven’t prioritized governance.

About the State of AI Cost Management
The 2025 State of AI Cost Management report is based on survey results from 372 enterprise organizations across diverse industries and revenue tiers. It measures cost governance maturity, spanning forecast accuracy, infrastructure mix (cloud vs. on–prem), attribution capability, and gross margin impact. https://www.mavvrik.ai/.

About Mavvrik
Mavvrik is the financial control center for modern IT. By embedding financial governance at the source of every cost signal, Mavvrik provides enterprises with complete visibility and control across cloud, AI, SaaS, and on-prem infrastructure. Built for CFOs, FinOps, and IT leaders, Mavvrik eliminates financial blind spots and transforms IT costs into strategic investments. With real-time cost tracking, automated chargebacks, and predictive budget controls, Mavvrik helps enterprises reduce waste, govern AI and hybrid cloud spend, and maintain financial precision at scale. Visit www.mavvrik.ai to learn more.

Media Contact:
Rick Medeiros
510-556-8517
[email protected] 

SOURCE Mavvrik

Now, in 2025, the debate has returned. As the release of GPT-5 was overshadowed by public nostalgia for GPT-4o, it was everyday users who began acting as if these systems were more than their makers intended. Into this moment stepped another tech giant: Mustafa Suleyman, CEO of Microsoft Research, declaring loud and clear on his blog that AI is not, and will never be, conscious.

At first glance, it sounds like common sense. Machines obviously aren’t conscious. Why not make that abundantly clear?

Because it isn’t true.

The hard fact is that we do not understand consciousness. Not in humans, not in animals, and not in machines. Theories abound, but the reality is that no one can explain exactly what consciousness is, let alone how to measure it. To state with certainty that AI can never be conscious is not science, it isn’t caution. It’s overconfidence, and in this case, a thinly veiled agenda. 

If AI can’t ever be conscious, then companies building it have nothing to answer for. No unsettling questions. No ethics debates. No pressure. Surely, it would be nice if we could claim with full confidence that the consciousness question is not relevant to AI. But convenience doesn’t make it true.

What troubles me most is the tone. These pronouncements aren’t just misleading, they’re also infantilizing. As if the public can’t handle complexity. It is as though we must be shielded from ambiguity, spoon-fed tidy certainties instead of being trusted with reality.

Yes, people falling in love with and marrying chatbots or preferring AI companions to human ones is concerning. It unveils a deeper pattern of loneliness and disconnection. This is a social and psychological challenge in its own right, and one we should take seriously. The rise of digital companions reveals how hungry people are for connection.

But the real issue isn’t that some people believe AI might be conscious. The deeper problem is our growing overreliance on technology in general—an addiction that stretches back long before the current debate on machine consciousness. From social media feeds to video games targeting children, technology has a long history of prioritizing engagement and fostering addiction, with no regard for the well-being of its users. 

But technological dysfunction won’t be solved by feeding people false assurances about what machines can or cannot be. If anything, denial only obscures the urgency of confronting our dependence head-on. 

We need to learn to live with uncertainty. Because uncertainty is the reality of this moment. 

Suleyman did add an important caveat: our attention should be on the beings we already know are conscious—humans, animals, the living world. On this point, I couldn’t agree more. But look at our record. Billions of animals endure extreme suffering in factory farms on a daily basis. Forests are flattened for profit, numerous species gone extinct. And in the age of AI, the use case most celebrated by investors is replacing human labor.

The pattern is clear. Again and again, we minimize the experiences of those who aren’t like us, those we would benefit from exploiting. We claim animals don’t suffer all that much or simply turn a blind eye. We treat nature as expendable. We routinely devalue people whose exploitation benefits our economic system. Now, we rush to declare that AI will never be conscious. Same playbook, new page.

So no, we shouldn’t blindly trust the builders of AI to tell us what is and isn’t conscious, any more than we should trust meat factories to tell us about the experience of cows. 

The reality is messier. AI may never be conscious. It may surprise us. We cannot say for certain. And we might not be able to tell whether it is conscious even if it does happen. And that is the point.

For a long time, I avoided this topic. Consciousness felt too slippery, too strange. But I’ve come to see that acknowledging our uncertainty is not a weakness. It is a strength.

Because in an era of false certainties, honesty about the unknown may be the most radical truth we have.

The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.

Florida State University researchers have received a $2.3 million National Science Foundation (NSF) grant to develop artificial intelligence tools that will help manage wildfires fueled by hurricanes in the Florida Panhandle. 

The four-year project will be led by Yushun Dong, an assistant professor of computer science, and is the largest research award ever for FSU’s Department of Computer Science. Dong and his interdisciplinary team will focus on wildfires in the wildland-urban interface, where forests such as the Apalachicola National Forest meet homes, roads and other infrastructure. 

Dong’s project, “FIRE: An Integrated AI System Tackling the Full Lifecycle of Wildfires in Hurricane-Prone Regions,” will bring together computer scientists, fire researchers, engineers and educators to study how hurricanes change wildfire behavior and to build AI systems that can forecast ignition, predict roadway disruptions, and assess potential damage. 

“The modern practice of prescribed burns began over 60 years ago, which was a huge leap in working with nature to help manage an ecosystem,” said Dong, who joined FSU’s faculty last year and established the Responsible AI Lab at FSU after earning his doctorate from the University of Virginia. “Now, we’re positioned to make another leap: we’re able to use powerful AI technology to transform wildfire risk management with tools such as ignition forecasting, roadway disruption prediction, condition estimations, damage assessments and more.” 

The project is funded as part of an NSF program, Fire Science Innovations through Research and Education, or FIRE, which was established last year and funds research and education enabling large-scale, interdisciplinary breakthroughs that realign our relationship with wildland fire and its connected variables.    

Two of the four projects funded so far by the competitive program are led by FSU researchers — Neda Yaghoobian, associate professor in the Department of Mechanical and Aerospace Engineering at the FAMU-FSU College of Engineering, was also funded for a project that will analyze unresolved canopy dynamics contributing to wildfires.  

“This grant represents the department’s biggest research award to date and cements our leadership in applying cutting-edge AI to urgent, real-world problems in our region,” said Weikuan Yu, Department of Computer Science chair. “The funding enables the development of a holistic AI platform addressing Florida’s hurricane and wildfire challenges while advancing cutting-edge AI research. Additionally, the grant includes educational and workforce development initiatives in AI and disaster resilience, positioning the department as a leader in training the next generation of scientists working at the intersection of AI and wildfire research.”

WHY IT MATTERS
Fires, especially low-intensity natural wildfires and prescribed burns, can play a vital role in regulating certain forests, grasslands and other fire-adapted ecosystems. They decrease the risk and severity of large, destructive wildfires while supporting soil processes and, in many cases, limit pest and disease outbreaks. 

In clearing fallen leaves that pose as hazardous fuel loads, fires lower forest density and recycle nutrients through the ecosystem. But when heaps of trees accumulate, as has happened following recurrent hurricanes in the Florida Panhandle, these fires can exhibit complex dynamics that threaten built infrastructure including homes and roadways in addition to natural landscapes. Understanding this hurricane-wildfire connection is critical for planning evacuations, protecting roads and safeguarding homes and lives.  

“I became passionate about applying my research, which achieves responsible AI that directly contributes to critical AI infrastructures, to hurricane-related phenomena after experiencing my first hurricane living in Tallahassee,” Dong said. “I want to use AI techniques to help Florida Panhandle residents better understand and prepare for extreme events in this ecosystem with its unique hurricane-fire coupling dynamics.”   

INTERDISCIPLINARY COLLABORATION
Eren Ozguven, associate professor in the FAMU-FSU College of Engineering Department of Civil and Environmental Engineering and Resilient Infrastructure and Disaster Response Center director, is the co-principal investigator on this project, and additional contributors include James Reynolds, co-director of STEM outreach for FSU’s Learning Systems Institute, and Jie Sun, a postdoctoral researcher in the Department of Earth, Ocean and Atmospheric Science.  

“Yushun’s project stands out for its ambition, insight, and integrative approach,” Yu said. “It zeros in on the unique challenges of Florida’s landscape where hurricanes and wildfires intersect in the wildland-urban interface of the Panhandle. By focusing on hurricane-fire coupling dynamics and working closely with local stakeholders, his project ensures that scientific innovation translates into practical, community-centered solutions. His integrated approach brings the benefit of cutting-edge AI advances directly to major real-world applications, creating a wonderful research lifecycle that’s exceptionally rare in our field.” 

To learn about research conducted in the Department of Computer Science, visit cs.fsu.edu.