Egress fees, lock-in, and pricing complexity aren't accidents. Learn the cloud trade-offs most teams miss and how open infrastructure changes the mode
Cloud providers sell a simple idea: use what you need, scale when you want, and let someone else manage the infrastructure. It works, and for many teams it’s the right starting point. But that convenience has trade-offs.
89% of enterprises now operate in multi-cloud environments, driven in part by the need to reduce vendor lock-in. The concern isn’t theoretical. It shows up over time in pricing, architecture, and flexibility.
Costs grow through things like egress fees and long-term commitments. Managed services introduce dependencies that are hard to unwind. Moving later becomes expensive, both technically and financially. This isn’t about whether cloud is useful. It is. The question is whether the terms stay favorable as you scale.
That’s why more teams are looking at open infrastructure. OpenStack provides control at the infrastructure layer, while Kubernetes enables workload portability. Together, they offer a way to keep the benefits of cloud without being tied to a single provider.
The starting point is simple: understand the trade-offs early.
§1 Egress Fees: You Pay to Leave
Egress fees typically range from $0.08 to $0.19 per gigabyte depending on the provider and region. That sounds small until you are moving terabytes of training data, model artifacts, or inference logs. At scale, egress can multiply the effective cost of storage several times over. Public pricing across major providers consistently falls within this range, with standard tiers starting around $0.08 to $0.12 per GB.
The structure is not accidental. Free ingress removes friction at the start. Paid egress introduces friction when you try to move. The more data you store, the more expensive it becomes to leave or even to operate across multiple environments.
For AI workloads, this is particularly visible. Training datasets are large. Checkpoints are frequent. Model artifacts are versioned and shared across teams. Every movement of data between regions, between services, or out of the provider entirely adds cost that many teams did not plan for.
On open infrastructure, this dynamic changes. Platforms like Atmosphere, built on Ceph for storage and OpenStack for infrastructure management, do not apply egress fees in the same way. Data moves within the environment without per gigabyte transfer costs because compute and storage are part of the same system, exposed through open APIs and operated under your control. There is no metered boundary between GPUs and training data, no added cost for replicating datasets across regions, and no surcharge for moving artifacts between environments. To see how this works in practice, read Deploying Atmosphere: A Guide to Storage Integration.
Kubernetes adds another layer of flexibility. Workloads can be scheduled where the data already exists instead of forcing data to move to where compute is available. The infrastructure follows the workload, not the pricing model.
§2 The Managed Services Trap
Managed services are easy to adopt. That's the point.
A managed database here. A managed ML platform there. A model registry, a secrets manager, a managed Kubernetes control plane. Each one solves a real problem and saves time on day one.
But each one is also proprietary. The APIs are provider specific. The data formats are non-standard. The integrations only work inside that ecosystem. And none of them have a portable equivalent you can take with you.
Over time, these services become the connective tissue of your architecture. Your training pipeline calls their feature store. Your inference endpoint depends on their load balancer. Your CI/CD triggers their container registry. Your access control lives in their IAM.
No single service feels like lock-in. But collectively, they make migration a full rewrite, not a lift and shift.
This is worth managing, not avoiding entirely. Managed services have real value. The key is knowing which ones create dependency you can live with, and which ones quietly become structural obligations you can't easily undo.
The difference with open infrastructure is transparency. On Atmosphere the services that support your workloads, storage through Ceph, orchestration through Kubernetes, infrastructure through OpenStack, are all built on open standards with no proprietary forks or vendor specific extensions. If you use them, you understand exactly what you depend on. And if you need to move, every component has a portable equivalent because VEXXHOST builds on upstream projects, not around them. To see how this works in practice, read The Complete Guide to Managed OpenStack with Atmosphere.
Managed services should make operations easier. They shouldn't make leaving harder.
§3 The Pricing Complexity Is a Feature, Not a Bug
Cloud pricing is not complicated by accident. It is complicated by design.
Instance types, storage tiers, API call charges, zone to zone transfer fees, per second vs per hour billing, reserved vs spot vs on demand. The layers add up quickly, and most organizations cannot accurately predict their cloud bill next month, let alone next year.
This is not a failure of tooling. It is a pricing structure that benefits the provider. When you cannot forecast costs, you cannot negotiate effectively. When you cannot isolate where spend is growing, you cannot optimize meaningfully. Complexity keeps you reactive.
On open infrastructure, pricing reflects actual resource consumption, not layered abstractions that obscure it. With Atmosphere you see what you use, and you control what you pay for. No surprise line items. No opaque tiers. No pricing model that requires a dedicated FinOps team to interpret.
Cost clarity is not a nice to have. It is how you make infrastructure decisions with confidence.
§4 What the Alternative Looks Like
The answer isn't avoiding the cloud. It's choosing infrastructure where the business model works for you, not against you.
Open infrastructure built on OpenStack, Kubernetes, and Ceph provides the same core capabilities compute, orchestration, storage without the egress fees, committed use traps, proprietary dependencies, or pricing opacity.
Atmosphere delivers this as a production ready platform:
- OpenStack for infrastructure control: GPU allocation, networking, storage, identity, all through open APIs
- Kubernetes for workload orchestration: upstream, CNCF certified, fully portable
- Ceph for scalable storage: deployed alongside compute, no egress penalties
- GPU passthrough, SR IOV, DPDK for AI ready performance
- On premise, hosted, or colocation: deployed where your business needs it
No multiyear lock in. No egress taxes. No pricing layers are designed to confuse. Every component is open, auditable, and replaceable.
You get the capabilities of a hyperscaler without the terms of one.
For a deeper look at how this architecture comes together, read OpenStack, Kubernetes, and AI: What 2025 Taught Us About the Future of Cloud.
Conclusion
Cloud providers make it easy to start and expensive to leave. Egress fees, committed use contracts, proprietary services, and complex pricing are not accidents. They are part of the model.
Organizations that ask hard questions early tend to build on infrastructure they control. Those that do not usually find the answers later in the invoice.
OpenStack for control. Kubernetes for portability. Atmosphere brings both together without hidden costs.
Explore Atmosphere.
