IBM POWER hardware is capital-heavy, AIX/PowerVM is licensed per socket, and SWMA/HWMA renewals compound every year — while AIX specialists get harder to hire. An IBM exit moves those workloads onto commodity x86 running an open, Kubernetes-native platform your existing team can operate.
Ænix runs IBM AIX/Power migrations end-to-end. The engineers who created and co-maintain Cozystack — the open-source destination platform — work alongside your team for assessment, sequencing, and execution.
Pairs with: Ænix Private Cloud Platform for regulated banks (air-gapped installs, chargeback, migration delivered as a quoted service), or the OpenShift alternative if you’re specifically replacing IBM Cloud Pak / OpenShift.
Who runs an IBM exit in 2026
Organizations triggered by:
- IBM Power Systems (AIX) hardware at end-of-life — refresh means another capital-heavy POWER purchase, or an exit.
- IBM cost compounding — premium POWER CapEx, socket-based AIX + PowerVM licensing, and SWMA/HWMA renewals, year over year.
- Oracle-on-Power tax — Oracle carries a core-factor of 1.0 on POWER (the maximum). Every non-Oracle workload still sitting on POWER inflates the licensable core count.
- Scarce specialists — AIX/PowerVM expertise is a shrinking, expensive talent pool; Kubernetes/DevOps is not.
- Sovereignty and sanctions exposure — a proprietary, single-vendor stack is a different risk profile from an open, CNCF-governed platform for state-owned and regulated institutions.
- Modernization — a legacy estate where the upgrade path is also the exit path, often alongside a move to microservices.
If two or more apply, a structured exit compounds. If a comfortable POWER refresh is already budgeted and nothing else bites, “stay and tune” is the honest answer.
What an Ænix IBM migration covers
1. Inventory and assessment AIX/Power estate: LPARs, sockets and cores, firmware, PowerVM dependencies, Oracle footprint, Cloud Pak/OpenShift usage. Workload classification: rebuild-now / re-platform-later / keep-on-bare-metal (Oracle) / retire.
2. Destination architecture Target platform on commodity x86. Cozystack default — KubeVirt for VMs, Cilium (eBPF) for networking, LINSTOR/DRBD on ZFS for storage, Tenant CRD for multi-tenancy. Capacity model, HA, and geo design.
3. Migration execution Cohort-based. Microservices and container workloads first; VMs via KubeVirt; databases re-platformed or attached externally. Parallel-run against the IBM estate until validation. Live migration and geo-stretch handled by the platform.
4. Decommission POWER frames retired as cohorts complete; AIX/PowerVM and IBM support contracts wound down. Oracle footprint compressed to dedicated hosts only.
Honest scoping note — endianness. AIX is big-endian on POWER, x86 little-endian: there is no binary lift-and-shift. Modern microservices and standard database/middleware move cleanly; legacy monoliths need re-architecture. We separate the two classes in the assessment, not mid-cutover.
The economics: Cozystack vs IBM
The model below is an illustrative list-price scenario for a mid-size bank (~500 staff) moving the subset of workloads that can leave POWER (microservices, VMs, non-Oracle databases) over a three-year horizon. Figures are order-of-magnitude and recomputed on real estate data during assessment.
| Line item (3 years) | IBM / AIX / Power | Cozystack (x86) |
|---|---|---|
| Hardware (CapEx) | $200,000 — refresh 2 POWER servers | $90,000 — 6 commodity x86 nodes |
| OS / platform licensing | $40,000 — AIX + PowerVM | $0 — Apache 2.0 |
| Support (3 yr) | $180,000 — IBM SWMA/HWMA | $198,000 — Plus support tier list price for 10 nodes (24×7, guided migration, 3 h/month training) |
| Oracle (licence + support) | $300,000 — on shared POWER (core-factor 1.0) | $120,000 — isolated to a minimal dedicated footprint |
| Migration services | — | Quoted after the assessment (not included above) |
| Total (3 years) | $720,000 | $408,000 + migration services |
The support line uses the published Plus tier as an illustration; an Ænix Private Cloud Platform programme for a bank is quoted per RFP, and migration services are quoted separately after the assessment. Model your own numbers with the TCO calculator or on a discovery call.
Oracle: the licensing trap to avoid
The single most expensive mistake in a Power-to-Kubernetes move is running production Oracle inside the cluster.
- Oracle treats Kubernetes and KubeVirt as soft partitioning. CPU limits and pinning do not narrow the licensable scope — “the processors of all nodes in the cluster are subject to Oracle licensing.”
- The node is licensed, not the pod. A whole worker node counts even if Oracle uses a fraction of its cores; a KubeVirt VM does not qualify as Oracle-approved hard partitioning.
- The clean path: keep production Oracle on dedicated, separately-licensed bare-metal and attach it to the platform as an external application (Helm chart / operator wrapping connection points and credentials via external secret reference) over a private network. Tenant workloads reach it like any managed endpoint; the database is never pulled into the cluster.
It compresses the licensable footprint as non-Oracle workloads leave POWER. (Oracle’s partitioning policy is “educational, not contractual” — finalize the model with Oracle and your legal team.)
Cozystack vs OpenStack vs IBM Cloud Pak
| Criterion | Cozystack | OpenStack | IBM Cloud Pak / OpenShift |
|---|---|---|---|
| What it is | Open PaaS framework on Kubernetes for building a cloud | IaaS — modular infrastructure services | Proprietary data/AI software bundle on Red Hat OpenShift |
| VM + containers | One API (KubeVirt + containers, one scheduler) | Separate: VMs via Nova, containers via Zun/Magnum | Container-centric; no native unified VM+container provisioning |
| Licence & cost | Apache 2.0; software free. Ænix support tiers from $1,250/mo per 10 nodes | Apache 2.0; pay for distro/support | Proprietary per-cluster subscription, vCPU-per-pod metric; restricted OpenShift entitlement |
| Vendor lock-in | Low — API-first, CNCF-governed | Medium — at the distro level | High — proprietary stack + bundled-restricted OpenShift |
| Multi-tenancy | Native (Tenant model, eBPF isolation, billing integration) | Native (Keystone, projects, quotas) | Supported (OpenShift namespaces + Zen) |
| On-prem / air-gap | Yes | Yes | Yes (operator-catalog mirroring) |
Cozystack is a CNCF Sandbox project (its Incubation application is in due diligence), released under Apache 2.0 and governed in the open rather than by a single vendor. That removes most of the “vendor changes the licence” risk of proprietary and quasi-open products: a different risk profile for a state-owned bank under a digital-sovereignty mandate.
Storage and scale on x86
The destination architecture is engineered for linear horizontal growth — each x86 node adds both compute and a share of distributed storage, no re-architecture:
- Storage in the kernel. LINSTOR orchestrates per-volume DRBD devices on ZFS; DRBD replicates in the Linux kernel rather than in a userspace daemon, so the write path does not cross into user space on every I/O. After a node returns, DRBD resyncs only the changed chunks by bitmap, not the whole disk — critical at large volume sizes.
- No bottleneck at scale. Each PVC is an independent DRBD device spread across the cluster — 100 volumes means 100 independent devices, not one fat shared device.
- Network. Cilium eBPF replaces kube-proxy with O(1) in-kernel service lookup; latency does not degrade as service count grows.
- Geo-stretch. Stretched-cluster designs can span up to three data centres; replication goes synchronous only for a migrating VM, governed by a hard RTT budget (~15 ms). These designs are delivered as engineering work; there is no automated cross-site VM failover.
How Ænix engages
- Assessment (14 or 28 days) — Platform Readiness Assessment: AIX/Power inventory, destination architecture, workload classification, Oracle plan, cutover sequencing, risk register.
- Pilot — Cozystack stood up as a working framework against your real requirements; success criteria agreed up front. Scope and price are set in the assessment.
- Migration — cohort execution with parallel-run validation, quoted after the assessment; legal/procurement can run on your templates (tenders, forms).
- Operations (optional) — managed Cozystack operations, 24×7, after cutover, quoted separately.
A recurring real-world idea: stand the platform up on the POWER servers being freed at end-of-life (POWER supports Linux) as a live demonstration before committing the wider estate.
Why Ænix specifically
- We created the destination platform. Estimates are calibrated against work we have shipped, not theory.
- Honest about hard parts. Endianness, Oracle licensing, and legacy re-architecture are surfaced in the assessment, not mid-cutover.
- Operable by your team. Kubernetes skills you can hire, not scarce AIX/PowerVM specialists.
- Open destination. Apache 2.0 and CNCF-governed — you run the platform you migrate to without a platform licence fee.
- EU and Central Asia teams. Engineering teams in the EU and Central Asia; EU contracts through AENIX s.r.o. (Czech Republic).
Typical migration timeline
| When | What |
|---|---|
| Day 0 | Discovery call (free) — confirm fit |
| Days 1-14 (or 1-28) | Platform Readiness Assessment, fixed price |
| Day 14 (or 28) | Executive readout — written plan + TCO on real data |
| After the readout | Pilot against real workloads |
| Build phase | Workload cohorts migrate; POWER frames retired as cohorts complete |
| End of programme | IBM/AIX decommission; Oracle compressed to dedicated hosts |
Estate size and the legacy/microservice mix drive the actual schedule; a private-cloud build typically runs 3-12 months depending on scope, and sequencing is set in the assessment.
Companies running platforms built with Ænix
Hosting providers running Ænix Public Cloud Platform in production
Hosting providers running Ænix Public Cloud Platform in production. For regulated finance, see the anonymised bank and financial group case studies.
"Running Nuvolos on sovereign bare metal meant we needed hyperscaler-grade multi-tenancy without the hyperscaler. Cozystack delivered exactly that — isolated tenants, shared GPUs, KubeVirt for legacy workloads — and it's why we can promise our customers infrastructure control, not just a workspace."
Alexandru PopescuCEO, Nuvolos
- OpenShift alternative — replacing Cloud Pak / OpenShift
- Cozystack vs OpenShift — direct comparison
- Private Cloud Platform — turnkey for regulated banks
- Financial services — sector context
- Data sovereignty — open, customer-controlled infrastructure
- Cozystack — the open-source destination platform
- OpenStack alternative and Cozystack vs OpenStack — if OpenStack is on your shortlist
Ænix created Cozystack (a CNCF Sandbox project) and co-maintains it with maintainers from other companies. On top of it, Ænix offers three platforms — Public Cloud, Private Cloud and AI.