88% of CFOs say cloud costs are rising. The board wants answers. Learn how to present a plan built on infrastructure control, not just cost optimization.
Cloud costs used to be an IT line item. In 2026, they are a boardroom problem.
According to Azul’s 2026 CFO Cloud Cost Optimization Report, 88% of CFOs say cloud spending is increasing, while 90% are concerned about its impact on profitability. The average estimated cloud waste now sits at 23% of total spend.
AI is accelerating the pressure. As those costs rise upstream, enterprises inherit the downstream effect through premium AI pricing, consumption based billing, and growing infrastructure complexity.
Most optimization advice focuses on symptoms: reserved instances, right sizing, or FinOps dashboards. Useful, but limited. When the platform provider profits from higher consumption, efficiency has a ceiling.
The better answer starts with architecture.
OpenStack gives organizations control over the infrastructure layer. Kubernetes keeps workloads portable and schedulable across environments. Together, they reduce dependency on opaque pricing models, minimize lock-in, and create a foundation where efficiency is operational, not aspirational.
This post is for the person who has to explain the bill and present the plan. Here's what to say.
§1 Why the Bill Keeps Growing
The short answer: AI workloads are fundamentally more expensive than the environments most cloud platforms were designed for.
Traditional workloads like web apps, databases, and CI/CD pipelines consume CPU, memory, and storage in relatively predictable ways. AI workloads do not. Training runs can occupy GPU clusters for days or weeks. Datasets expand into hundreds of terabytes. Checkpoints, artifacts, and inference logs accumulate continuously. Every layer of the stack scales faster and costs more.
GPU compute is only part of the problem. The surrounding costs are often even larger.
Storage growth compounds constantly. Every experiment generates checkpoints, logs, and model artifacts, much of which is never cleaned up.
Egress fees scale with AI’s appetite for data movement. Moving datasets between regions, providers, or services quickly becomes expensive at scale.
Idle capacity is one of the biggest sources of waste. GPUs provisioned for peak demand often sit underutilized between jobs. According to Flexera’s 2026 State of the Cloud Report, organizations estimate that 27% of cloud spend is wasted, largely driven by overprovisioning and underutilized resources.
Managed AI services add another layer of cost. Model registries, orchestration platforms, observability tooling, and proprietary ML services each become separate recurring line items.
Cloud pricing itself is also rising. GPU shortages, infrastructure expansion, and energy costs are pushing providers toward higher pricing and premium AI service tiers. The assumption that cloud infrastructure becomes cheaper every year no longer holds for AI workloads.
The result is not a single alarming charge. It is continuous growth across every infrastructure category, often without clear visibility into which workloads are driving the increase or whether the spend is generating proportional value.
For a deeper look at how AI infrastructure costs compound, see The GPU Capacity Crisis: Why Enterprises Are Rethinking AI Infrastructure.
§2 Why Traditional Cost Optimization Isn’t Working
The typical response to rising cloud costs is FinOps: right sizing infrastructure, purchasing reserved capacity, reducing obvious waste, and negotiating better pricing.
For traditional workloads, that can work. For AI, it reaches a limit quickly.
Right sizing assumes flexible infrastructure choices. GPU instances rarely work that way. Most providers offer fixed configurations, forcing teams to either overprovision or compromise on performance. Reserved capacity assumes predictable demand, but AI workloads are often experimental, bursty, and constantly changing.
Most FinOps tooling also operates within the provider’s ecosystem. It helps reduce spend on the same platform, but it does not question whether the platform itself is driving the problem. When pricing is opaque, egress fees discourage movement, and managed services create proprietary dependencies, optimization inside the same environment produces diminishing returns.
Leadership is not asking whether the team reduced last quarter’s bill by a few percentage points. They are asking why infrastructure costs continue growing at this pace and whether the architecture supports a fundamentally different financial model.
That is not an operational question. It is an architectural one.
For more on how lock-in, pricing opacity, and proprietary AI services affect long-term infrastructure costs, read What Your Cloud Provider Doesn't Want You to Think About.
§3 The Questions the Board Is Actually Asking
Strip away the technical language and leadership is asking four things.
What are we spending, and what are we getting in return? They want costs tied to workloads, teams, and business outcomes, not a single aggregate cloud bill with limited attribution.
Why is this difficult to predict? Infrastructure costs that fluctuate significantly quarter to quarter without a clear explanation erode confidence. AI workloads are inherently bursty and experimental, but leadership still expects visibility and planning.
What happens if priorities change? New regulations, new financial pressures, or a shift in strategy can change infrastructure requirements quickly. The question is whether the organization can adapt without being trapped by contracts, proprietary services, or expensive migration paths.
Are we actually in control? Not simply whether costs are being monitored, but whether the organization controls the infrastructure decisions that determine those costs in the first place.
Most teams can answer the first question at a high level. Far fewer can answer the last one honestly.
§4 What a Defensible Answer Looks Like
The board doesn't need a technical architecture diagram. They need confidence that the team has a plan that addresses cost, visibility, flexibility, and control.
Full visibility into cost drivers. Spend attributable to specific workloads, teams, and projects not aggregate invoices. When the board asks why costs rose 15% this quarter, the answer is immediate and specific.
Intentional placement. Not every workload belongs in the same environment. Training jobs, inference endpoints, and experimentation have different cost profiles. A defensible strategy places each workload where the economics make sense, public cloud for burst, private infrastructure for sustained workloads, hybrid when both apply.
Architecture that preserves flexibility. No multi-year contracts that lock in the wrong capacity. No proprietary dependencies that make migration a rewrite. No vendor whose pricing you can't walk away from. The board wants to know there's an exit path — even if you never use it.
Open infrastructure as the foundation. OpenStack provides control over compute, storage, networking, and identity through open APIs. Kubernetes keeps workloads portable across environments. Together, they give the organization the ability to optimize infrastructure decisions based on business needs, not provider constraints.
Ultimately, the answer leadership wants is straightforward: we understand where the money is going, we control the infrastructure decisions behind it, and we retain the ability to adapt without major financial or operational penalties.
For a closer look at how this model compares to hyperscaler economics, read The Real Cost of Running AI on Hyperscalers vs. Open Infrastructure.
§5 How Atmosphere Fits the Conversation
Atmosphere by VEXXHOST combines OpenStack and Kubernetes into a platform designed for organizations looking for more visibility, flexibility, and control over growing AI infrastructure costs. Built on upstream OpenStack and CNCF certified Kubernetes, the platform supports AI ready infrastructure features including GPU passthrough, MIG, and SR IOV while avoiding proprietary lock in and opaque pricing layers.
Organizations can deploy Atmosphere on premise, in colocation environments, or as a hosted platform while maintaining the same operational model across environments. Built in Prometheus-based monitoring provides visibility into GPU utilization, workload attribution, and infrastructure consumption in real time, making it easier to understand where resources are being used and how costs evolve over time.
For teams that want infrastructure control without managing day to day platform operations internally, VEXXHOST also offers a managed approach while preserving architectural flexibility and workload portability.
This is less about short term cost cutting and more about creating an infrastructure foundation that remains transparent, adaptable, and operationally sustainable as AI workloads continue to scale.
For a full overview of the platform, read The Complete Guide to Managed OpenStack with Atmosphere.
Conclusion
The answer that builds confidence is not “we are optimizing.” It is “we control the platform, we can see where every dollar goes, and we can change direction when we need to.”
OpenStack for infrastructure control. Kubernetes for workload portability. Atmosphere for both, with the visibility and cost predictability organizations are increasingly being asked to deliver.
Explore Atmosphere!
