AI is driving emissions up and GPU utilization down. Learn why sustainability is an infrastructure problem and how OpenStack and Kubernetes solve it.
§1 AI Has an Energy Problem
The biggest names in tech are watching their climate commitments unravel, and AI is the reason.
Google’s emissions rose nearly 50% since 2019, while Microsoft’s increased 23.4% versus its 2020 baseline as AI-driven data center demand surged. (The Guardian)
And the problem is accelerating. Global data center electricity consumption is projected to double by 2030, with U.S. data centers potentially consuming 12% of total electricity by 2028.
The issue isn’t just where energy comes from. It’s how much gets wasted. Idle GPUs, overprovisioned infrastructure, and poor workload scheduling burn power constantly even when utilization stays low.
This is an infrastructure architecture problem.
Open platforms built on OpenStack and Kubernetes are designed for exactly this. They enable right-sized GPU allocation, intelligent workload scheduling, dynamic power management, and full observability, the capabilities that turn sustainability from a reporting exercise into an operational reality.
The greenest GPU is the one that’s actually being used. The question is whether your infrastructure makes that possible.
§2 The Numbers: How Big Is the Problem?
AI’s energy impact is no longer theoretical.
Global data center electricity consumption reached roughly 415 TWh in 2024 and is projected to more than double to around 945 TWh by 2030. AI-driven server demand is one of the primary reasons.
And much of that power is wasted.
CAST AI’s 2026 Kubernetes Optimization Report found average GPU utilization across analyzed clusters was just 5%, meaning most provisioned GPU capacity sits idle while still consuming power and cooling resources.
The issue isn’t careless teams. It’s infrastructure designed around overprovisioning, scarcity anxiety, and poor workload visibility.
Every idle GPU still draws power, generates heat, and increases cooling demand. This isn’t just a cost problem. It’s an energy efficiency problem hiding behind cloud spend.
§3 Hyperscalers Can’t Solve This
Hyperscalers have invested billions into renewables, carbon offsets, and data center efficiency. But that doesn’t solve the waste happening inside the infrastructure itself.
The problem is structural.
GPU instances come in fixed sizes. If a workload needs 40% of a GPU, organizations still provision and power the entire device. Across hundreds of workloads, utilization drops while energy consumption stays high.
Customers also have limited control over infrastructure-level power management. When workloads finish, the physical hardware remains online because the provider manages consolidation and capacity availability at scale.
Visibility is limited too. Most cloud sustainability dashboards provide estimated emissions rather than direct workload-level measurements. That makes optimization difficult at the GPU or job level.
The business model itself rewards higher resource consumption. Reserved capacity, always-available infrastructure, and overprovisioning increase utilization from the provider perspective, not necessarily from the customer workload perspective.
For a deeper look at how these structural incentives work, read What Your Cloud Provider Doesn't Want You to Think About.
This doesn’t mean hyperscalers are ignoring sustainability. Major providers continue investing heavily in clean energy and efficiency initiatives.
But cleaner electricity powering idle infrastructure is still waste.
When organizations don’t control infrastructure efficiency, sustainability remains a reporting metric instead of an operational outcome.
§4 Utilization: The Overlooked Sustainability Lever
The sustainability conversation focuses on energy sources: solar, wind, nuclear, carbon credits. Those matter. But one of the biggest levers organizations control is simpler: use the infrastructure they already provisioned more efficiently.
When GPU utilization sits at 5%, most of the energy powering those GPUs produces nothing useful. Improving utilization doesn’t require new hardware or new energy sources. It requires better infrastructure management.
Flexible GPU allocation through MIG, vGPU, and passthrough helps match resources to actual workload requirements instead of overprovisioning full devices.
Kubernetes scheduling based on workload priority and GPU topology keeps hardware active instead of leaving systems idle across clusters.
Dynamic power management allows unused nodes to power down when workloads consolidate onto fewer machines.
Right sizing also reduces structural waste. When organizations control the infrastructure, they deploy resources based on workload requirements rather than fixed instance sizes from a provider catalog.
These are infrastructure capabilities, not sustainability slogans. And they depend on how much control organizations have over the stack underneath their AI workloads.
For a closer look at how these capabilities affect AI workloads, read What Actually Matters in AI Infrastructure (Beyond GPUs).
§5 Regulation Is Coming
Sustainability reporting for AI infrastructure is moving from voluntary to mandatory.
The EU Energy Efficiency Directive already requires large data centers to report energy consumption, water usage, and renewable energy metrics. The EU AI Act also includes provisions around energy transparency and efficiency for AI systems.
In the U.S., California’s Climate Corporate Data Accountability Act (SB-253) will require large companies doing business in the state to report Scope 1 and 2 emissions beginning in 2026.
The direction is clear: organizations will need to measure, report, and reduce infrastructure energy consumption.
The challenge is visibility.
On proprietary platforms, sustainability data is typically exposed through provider dashboards and aggregate estimates. On open infrastructure, organizations can monitor utilization, scheduling, and resource allocation directly through tools like OpenStack and Prometheus-based observability stacks.
When reporting requirements tighten, operational visibility becomes an infrastructure requirement, not just a sustainability feature.
For more on how open infrastructure supports regulatory readiness, read Sovereign by Architecture: Building AI Infrastructure for the EU AI Act.
§6 What Energy Efficient AI Infrastructure Looks Like
Efficiency isn’t a feature you add on. It’s the result of how infrastructure is designed and operated.
OpenStack enables energy aware infrastructure management through workload consolidation, granular resource allocation, and automated control over idle infrastructure. NUMA aware placement, CPU pinning, and GPU scheduling help ensure workloads land on the right hardware without wasting capacity elsewhere.
Kubernetes complements this at the workload layer. Bin packing, GPU aware scheduling, and autoscaling help keep hardware utilization high while reducing unnecessary capacity.
Ceph keeps storage local to compute, reducing unnecessary data movement and the energy overhead that comes with it.
Full observability ties it together. Prometheus and open monitoring tools provide real time visibility into utilization, power draw, and infrastructure efficiency across the stack. When teams can identify where waste exists, they can reduce it and report on it directly.
Deployment flexibility matters too. On premise and colocation environments let organizations choose efficient facilities, cooling strategies, and energy sources instead of inheriting infrastructure decisions from a provider.
None of these requires sacrificing performance. It requires infrastructure that gives organizations control over how resources are allocated, scheduled, and optimized.
To evaluate how your current stack measures up across compute, storage, networking, and orchestration, read How to Evaluate Whether Your Infrastructure Is AI-Ready.
§7 How Atmosphere Delivers It
Atmosphere by VEXXHOST brings every capability in this post together into a single, production-ready platform.
GPU efficiency: MIG, vGPU, and passthrough with NUMA-aware placement. Allocate exactly what workloads need. No fixed instance types. No forced overprovisioning. Fewer idle GPUs means less wasted energy.
Dynamic power management: OpenStack's workload consolidation strategies migrate VMs onto fewer hosts and power down idle nodes. You set the policies. You control when infrastructure scales down — not just up.
Observability: Prometheus-based monitoring built in. Real-time visibility into utilization, resource allocation, and efficiency at every layer. When regulators or leadership ask for data, it's already there.
Storage: Ceph deployed alongside compute. Data stays local to GPUs, reducing network traffic and the energy overhead of moving data between zones. To see how this works in practice, read Deploying Atmosphere: A Guide to Storage Integration.
Deployment flexibility: On-premise, colocation, or hosted. Choose facilities with clean energy access, efficient cooling, and the right density for your workloads. The platform adapts to your sustainability requirements, not the other way around.
Every component is open source, auditable, and replaceable. No proprietary forks. No black-box control planes. The efficiency gains are measurable because the infrastructure is transparent.
For teams without deep infrastructure expertise, VEXXHOST also offers fully managed operations, so you get efficient, right-sized AI infrastructure without building the ops team from scratch. Learn more in The Complete Guide to Managed OpenStack with Atmosphere.
Conclusion
AI's energy problem isn't about where electricity comes from. It's about how much gets wasted.
GPU utilization in single digits. Over-provisioned instances running 24/7. Infrastructure designed for peak demand idling the rest of the time. No amount of renewable energy fixes an architecture that wastes most of what it consumes.
The solution is infrastructure you control, where GPUs are allocated efficiently, idle nodes power down, and every watt is visible and accountable.
That's what open infrastructure delivers. That's what Atmosphere is built for.
Explore Atmosphere, AI infrastructure that performs and doesn't waste.
