Apache private cloud tools: what a small business pays

October 30, 2026 update: added setup time and costs when an AI agent does the typing.

Rent the few servers your business needs before building a private cloud. CloudStack, OpenWhisk and ZooKeeper belong with an IT provider; Karaf and the remaining runtimes mainly matter inside software your developer already maintains.

A private cloud gives you control over how virtual machines, storage and networks are offered to users. It also gives you responsibility for the machines that offer them. That can make sense for a hosting provider or a business with a real internal infrastructure team. Two business applications are rarely enough reason on their own.

Start with what must keep working when a server fails, how long recovery may take and who answers the phone. A dashboard that creates machines does not answer those questions. We read the project documentation and price pages on October 29, 2026; we did not install these systems or test a production topology.

Every project in this group

Registry dates may lag current releases. File dates are artifact modification dates. The Aries row represents its Proxy component and the Felix row its Framework component, not every separately released package. CloudStack, OpenWhisk and Karaf received fresh release checks. Incubating projects are not yet fully endorsed by the ASF; incubation is not a guarantee of technical quality.

CloudStack: only when operating infrastructure is part of the job

CloudStack manages virtual machines, accounts, networks and storage. It makes sense when your business must offer repeatable computing environments to several teams or tenants, rather than maintain a couple of isolated applications. The download page and release announcement record identify 4.23.0.0, published September 2, 2026. It is a Regular release; do not assume every latest release is an LTS release.

For example, a 12 person machine shop develops software for equipment it sells. Its developers need clean test environments that do not disturb production. CloudStack could let an administrator provide those environments from owned capacity. On a normal day, a developer requests a known template instead of rebuilding a test machine. The low cost case below is only a hardware and operations floor; the shop needs a network and storage quote before approving a cloud.

A 15 person ecommerce business wants separate staging and production environments for its custom store connector. CloudStack could allocate them, but renting two ordinary machines may already solve the problem. The high labor case shows what more involved operation can cost. The daily benefit of isolation does not require a private cloud brand or a new cluster if rented machines meet the requirement.

The smallest layout has more than one role

The installation overview allows management, database and storage services to share a machine. It specifies 4 GB memory and 250 GB local disk, with 500 GB recommended, for that management-side layout. A separate hypervisor host needs hardware virtualization, 4 GB memory and 36 GB disk, plus suitable networking. New hosts must not arrive carrying existing guest workloads that the installation is expected to preserve automatically.

That gives a two-machine trial, with single points of failure. The selected DigitalOcean tiers in our shared assumptions do not meet the management disk minimum: their included disks are 80, 160 and 100 GiB. Quoting a $24 server as the whole cloud would hide a real requirement.

Use the official quick installation guide only after an operator maps addresses, names, hypervisors, storage and recovery. The required skills include networking, Linux, the database and virtualization. Give that operator a drawing of what happens if the storage machine stops, not just a list of VM sizes. Skip CloudStack when nobody owns those dependencies.

A cost floor, not a deployment quote

Prices are US dollars before tax, read October 29, 2026. Our shared labor proxy is $68.09/hour, from the BLS sysadmin wage of $47.66 divided by the 70.0% wage share. An outside IT provider bills more; get a written quote. DigitalOcean Small is $24/month for 2 vCPU, 4 GiB RAM and 80 GiB disk; Medium is $48 for 4 vCPU, 8 GiB and 160 GiB.

We hold retained backups at 100 GB at both ends. B2 costs (100 minus 10) x $0.00695 = $0.6255/month before rounding. We assume no excess egress or chargeable Class D requests. Setup = setup hours x $68.09. Monthly = server charges + backup + monthly administration hours x $68.09. A 36 month total is setup + 36 x monthly. Savings = alternative total minus self hosted total; break even is ceiling(setup difference divided by monthly savings), or never if ongoing self hosting costs more.

Our IT platform range is 24 to 60 setup hours and 6 to 16 administration hours/month. A packaged single-runtime service uses 4 to 12 and 1 to 3. These are assumptions, not project staffing promises. Application development, migration, training, tax, major incidents and common application maintenance are excluded. Alternative incremental setup is assumed zero for existing managed-service accounts and prepared applications, not zero real adoption effort. The unmanaged two-VM comparison instead includes its own setup, backups and operating labor.

The shared Linode plan catalog prices a 32 GB shared-CPU VM with 640 GB disk at $192/month. Vultr's published bare-metal starting offer lists an Intel E3-1270 machine with 32 GB RAM and two 240 GB SSDs from $120/month. Using one as a management-side size reference and one as a hypervisor size reference yields $192 + $120 = $312/month before backup and labor.

These two offers are a capacity cost floor, not a verified cross-provider CloudStack installation. Network placement, private connectivity, storage traffic, availability and actual stock need a provider quote. Do not put shared storage across the public internet because two prices happen to add up. A connected or colocated design may cost more. We show the floor because it already fails the small-business comparison before those missing costs are added.

The narrower alternative is two DigitalOcean Medium application VMs at 2 x $48 = $96/month. That buys two machines, not CloudStack's private-cloud control or equivalent capacity. We hold the two-VM count fixed at both ends and apply 24 to 60 setup hours and 6 to 16 monthly hours to the CloudStack floor. The two-VM alternative also needs OS maintenance: we apply 4 to 12 setup hours and 1 to 3 monthly hours to each VM, totaling 8 to 24 setup and 2 to 6 monthly hours, plus the same 100 GB backup allowance. Its monthly operating cost is $96 + $0.6255 + (2 to 6) x $68.09, or $233 to $505 rounded. Common application work remains outside both figures.

This is not a claim of measured performance or a complete production budget. If the shop only needs two application machines, the extra cloud control must justify a much larger bill. If it does, commission a topology and recovery design before committing to servers.

Setup time when Claude or ChatGPT does the typing

We keep the machine shop's $312 CloudStack server floor and 100 GB backups; topology still needs a quote. Hours are our assumptions, including prompting, review, testing and corrections. Studies do not measure Apache setup savings; METR found slower work in its 2025 experiment.

We charge one separate setup month at $100, then allocate $50/month for 36 operating months to Claude Max 5x or ChatGPT Pro 100. Other work must use the remaining half; total allocations cannot exceed the subscription. Dedicated use adds $1,800. Allowances differ; check limits and charge extra usage.

API assumptions total 1 to 5 million setup input tokens and 100,000 to 500,000 output; monthly input is 500,000 to 2 million and output 50,000 to 200,000, including retries. Low input is 30% ordinary, 10% short cache writes and 60% reads; high input is uncached. Per million input/write/read/output rates are $4/$5/$0.20/$20 for Opus and $10/$12.50/$1/$50 for Astra, using standard short-context pricing. Each call must fit that band. T0/Tm are setup/monthly tool fees; round totals after calculation.

L = $68.09/hour; H = $312 + $0.6255/month
S = manual setup hours; A = manual monthly admin hours
q = 0.65 low or 0.90 high; r = 0.80 low or 1.00 high
Manual setup = S x L; Manual monthly = H + A x L
AI setup = S x q x L + T0; AI monthly = H + A x r x L + Tm
36 month total = setup + 36 x monthly; saving = manual total minus AI total
API fee = (ordinary input x input rate + cache writes x write rate
           + cache reads x read rate + output x output rate) / 1,000,000

The conclusion stays: two rented application VMs remain cheaper, with no break even against that alternative. High-case subscriptions save no monthly labor and exceed manual CloudStack cost from month 7. Low-case subscriptions save $1,613 against manual operation.

Claude Code or Codex can draft CloudStack configuration and recovery commands. Desktop Commander, on the new server or through authorized SSH, can install packages, edit configuration and read management logs. Its free local service is assumed; optional $20/month Remote Pro adds $380 with the same allocation. Claude in Chrome can inspect provider, Cloudflare and CloudStack dashboards while we watch. ChatGPT's cloud browser uses a separate browser where available.

Use new test machines, approve commands or restrict the sandbox, and keep secrets out of chat. Watch dashboard actions and review every change, including network isolation and a restore test. Use subscriptions for a person directing supported tools and API keys for custom automation. We still own backups, updates and topology approval.

OpenWhisk: events can run code without your own platform

OpenWhisk runs application actions when an event or request arrives. A developer supplies the action code; an operator supplies the platform that schedules it, stores its state and exposes the endpoint. Calling it serverless does not remove those servers when you host it yourself.

The download page and distribution directory identify core release 2.0.0 with an April 7, 2024 artifact date. Runtimes and clients have separate releases, so that core date alone does not establish that every part is inactive.

For example, an eight person print shop wants uploaded artwork to produce thumbnail proofs for its quoting clerk. A developer writes a function that creates the image after upload; the clerk still approves the proof. The low scenario assumes 100,000 monthly invocations, each using 1 GB for one second. That is an assumption to replace with a measurement, not a claim that any particular artwork conversion will finish in a second.

A 12 person delivery company wants scanner and vehicle-status events accumulated into draft route summaries for office review. The high example assumes one million monthly invocations with the same resource shape. A person checks the result before sending it. Both examples need developer work and storage outside the runtime bill.

Keep the local demonstration in its proper role

The Kubernetes deployment README documents a small default environment with a worker of at least 2 virtual CPUs and 4 GB RAM. It explicitly calls the default a testing and exploration platform. The guide also describes a local Docker environment. Our Small all-in-one evaluation machine follows that small-test scale; it is not a production minimum, availability promise or quote for a managed Kubernetes worker.

Production planning must account for control-plane capacity, workers, persistent storage, ingress and scaling. The technical requirements need compatibility review against current Kubernetes versions. The separate standalone runner is for local development and tests; its default in-memory store can lose changes when stopped.

Start with an action that reads a test file and produces another test file. Test retries and duplicate events before allowing it to affect a customer record. Keep credentials scoped to that action. Skip self hosted OpenWhisk when the platform itself would become the business's largest systems-administration task.

Cost against Lambda Event Functions

AWS Lambda pricing lists US East (N. Virginia) x86 Event Functions at $0.20 per million requests and $0.0000166667 per GB-second, with monthly allowances of one million requests and 400,000 GB-seconds. Our low volume fits both allowances. At the high volume, compute is (1,000,000 minus 400,000) x $0.0000166667 = $10/month rounded, with no request charge after the allowance. This calculation assumes that the account's allowance has not already been used elsewhere.

Logs, triggers, networking, persistent storage and application work can add charges. We compare the stated runtime layer only. The OpenWhisk pilot uses one Small server and the IT labor range of 24 to 60 setup hours and 6 to 16 monthly hours. Any production cluster would be a separate larger quote.

For intermittent small jobs, the managed runtime wins this cost comparison. A need for private execution or a particular integration may still justify OpenWhisk, but it should be a written requirement rather than an assumed saving.

Karaf: ask about the application inside the runtime

Karaf hosts modular Java applications and provides management facilities. An owner normally encounters it when a supplier builds an integration service. You do not install Karaf and receive a finished stock-control or maintenance-request system.

The download page lists OSGi Runtime 4.4.12, dated September 27, 2026. The getting-started page describes installing Java, downloading and extracting the distribution, then starting the runtime. The manual states Java 8 or later and 50 MB disk for the binary, but gives no universal workload RAM or CPU floor. Fifty megabytes is not a complete application storage budget.

For example, a seven person wholesaler has a custom service translating stock changes into shipping-system updates. Its developer might package that service in Karaf. On a normal day, staff see failed transfers in an exception list instead of retyping every successful order. The low case below assumes the application already exists and has been tested on a 4 GiB Small server.

A ten person property manager has a custom maintenance-request API used by its office application. A supplier might run that API in Karaf and store records in an external database. The high case allows more monthly administration. It does not pay for writing the API or supplying its database, and staff still decide how to handle urgent reports.

Compare an appropriate hosting shape

DigitalOcean App Platform lists a shared 2 vCPU, 4 GiB container at $50/month. That can be a useful comparison for an already-built stateless HTTP application. It is not a managed Karaf administration contract.

The platform's limits include ephemeral local storage, no persistent volumes and restrictions on exposed protocols. The comparison therefore assumes an immutable container deployment, HTTP traffic and state already held externally on both routes. Do not move a live administrative shell and local state there unchanged. High availability requires a larger design than either single-instance figure below.

Our self hosted case uses one Small server, 4 to 12 setup hours and 1 to 3 monthly hours. Application updates, external state and business logic maintenance remain separately owned on both paths.

The hosted container is cheaper under these assumptions. Karaf can still be appropriate software inside that container if the application's design fits. Ask the supplier to explain its deployment model, security boundaries and recovery process before deciding where it should run.

Needs your IT department or a provider

These remaining tools solve infrastructure problems. Our monthly hours are platform-level allowances, not a promise of support or amounts to add once for each component in the same environment. Where no usable managed price or production minimum was verified, the missing quote is stated.

Airavata. A small materials-testing firm with ten technical staff might need a portal into scientific computing resources. The installation guide describes an Ansible standalone inventory covering a database, identity service, message broker, coordination, application and portal. One inventory is not one simple application. No universal hardware floor is published there, and old operating-system examples need review. Our allowance is 6 to 16 hours/month across the platform, plus custom integration. SciGaP offers hosted science-gateway infrastructure; no public price was verified, so request terms. Ask who controls research-resource credentials and how a failed job is reconciled.

Brooklyn. A six person software integrator supporting retailer clients might use infrastructure blueprints to reproduce environments. The requirements page gives a development example of 1 vCPU, 3.75 GiB RAM and 4 GB disk; larger production guidance is 2 or more cores, 8 GB RAM and 100 GB disk. A Medium server fits that runtime allowance; the machines it manages cost extra. Use the archive-based start guide. Allow 6 to 16 platform hours/month because credentials and deployment changes are powerful. A simpler HTTP application could use the $50 App Platform container below instead; that is not a multi-cloud orchestration replacement. No fixed managed Brooklyn price was verified.

Helix. A 20 person delivery-software business might need to distribute partitions across its application servers. The quickstart starts ZooKeeper and multiple participants to demonstrate failures. It is an older architectural example, not a hardware minimum for the current release. Production needs the application participants, controllers and resilient coordination sized together. Budget 6 to 16 hours/month for the platform, with application engineering separate. No fixed managed Helix price was verified; obtain a quote or ask whether a managed database removes the requirement. Different coordination APIs are not interchangeable without code changes.

VCL. An eight person training company might reserve remote desktop environments for classes. The 2.5.2 installation guide can put its database, web portal and management node together or apart; guest hosts and images are additional. It gives no universal RAM minimum. Use a supported OS and reserve 6 to 16 administration hours/month for images, access and recovery. Before storing student education records, check applicable FERPA requirements. Amazon WorkSpaces pricing includes an Oregon Windows Standard example at $44/desktop/month with 175 GB root and 100 GB user volumes. Eight desktops would be 8 x $44 = $352/month before extras; this is hosted desktops, not a full VCL equivalent. Ask about image licensing, export and classroom peaks.

YuniKorn. A 20 person ecommerce business with an existing Kubernetes cluster might need to keep overnight data jobs from competing with order services. The deployment modes guide puts this scheduler inside Kubernetes; it is not an independent server to buy. No universal RAM minimum is stated. Our shared cluster oversight allowance is 6 to 16 hours/month, not an extra charge per scheduler component. DigitalOcean Kubernetes lists a free standard control plane or $40/month for high availability, with workers and storage extra. Its default scheduler may already suffice; that offering is not managed YuniKorn. Ask how scheduling changes are tested and rolled back.

ZooKeeper. A 15 person manufacturer may inherit a vendor application requiring shared configuration and leader coordination. The administrator guide recommends at least three servers on separate machines for a fault-tolerant ensemble. Three Small servers are 3 x $24 = $72/month before backups and 6 to 16 monthly administration hours. The 4 GiB allocation is our assumption; load-test memory and disk behavior and avoid swapping. A quorum is not a backup, and putting all replicas on one host defeats the purpose. No standalone managed ZooKeeper price was verified. Ask the vendor for a supported managed option and a maintenance quote; do not substitute a different coordination service without its approval.

Your developer's toolbox

• Aries: your Java supplier may use these modular enterprise components; ask for the actual component versions, not just the umbrella name.
• Curator: this client library helps application code use ZooKeeper; installing it does not supply the three-server ensemble.
• Felix: modular Java applications may use the framework and its related components, which have separate release schedules.
• Libcloud: your developer can call different cloud APIs through Python; the credentials it uses still need limited permissions and rotation.
• OpenServerless (Incubating): its CLI instructions say convenience builds are not official Apache releases or production ready. Treat it as development work, with no verified numbered release in this catalog.

Risks and support

Cloud credentials, machine images and service keys can reach far beyond one application. Keep administrative access separate from ordinary application use, record who can create resources, and rehearse recovering a service without its original administrator. Backups need a protected destination outside the failed machine's storage, with a restore someone has actually tested.

Check CloudStack security, OpenWhisk security, Karaf security and ZooKeeper security, alongside ASF guidance. An old guide can describe a valid architecture while naming obsolete dependency versions. Confirm release-specific compatibility rather than treating old commands as a supported stack.

Use mailing-list activity, recent fixes and Attic status when assessing a quiet project. Core, client and plugin versions often move separately. For commercial help, Karaf's community page lists providers including Codecentric, Yupiik and Savoir Technologies without endorsing them. No price is stated there. Request an explicit maintenance and response-time quote; community discussion is not an on-call contract.

The Apache License 2.0 permits use and modification under its conditions, including applicable license and notice obligations when redistributing. It provides no warranty and no right to imply ASF endorsement. This explanation is not legal advice, and third-party software in a platform can have separate terms.

Download the kit

Download the private-cloud and cluster-tools kit for operator questions, setup boundaries and the same cost worksheets. The restraint behind The $100 Network applies here: spend on a specific business requirement before buying a platform to discover one.

Related reading

• Apache projects for small business
• Web servers and API gateways
• Monitoring and logging
• Big data and lakehouse
• Self hosted AI server guide

Fact-check notes and sources

Project checks and new comparator prices were read October 29, 2026, from 21:38 to 21:42 UTC. Catalog and shared infrastructure prices were supplied by earlier reads that day. No installation, cross-provider network test, capacity benchmark or production support quote is claimed.

1. Apache Airavata. Read October 29, 2026.

2. Index of /airavata/0.17. Read October 29, 2026.

3. Apache Aries :: Apache Aries. Read October 29, 2026.

4. Index of /aries. Read October 29, 2026.

5. Home, Apache Brooklyn. Read October 29, 2026.

6. Index of /brooklyn/apache-brooklyn-1.1.0. Read October 29, 2026.

7. Apache CloudStack, Apache CloudStack. Read October 29, 2026.

8. cloudstack release 4.23.0.0 on GitHub. Read October 29, 2026.

9. Apache Curator. Read October 29, 2026.

10. curator release apache-curator-5.9.0 on GitHub. Read October 29, 2026.

11. Welcome to Apache Felix :: Apache Felix. Read October 29, 2026.

12. Index of /felix. Read October 29, 2026.

13. Apache Helix , Home. Read October 29, 2026.

14. helix release helix-2.0.1 on GitHub. Read October 29, 2026.

15. Apache Karaf, The Modulith Runtime. Read October 29, 2026.

16. Apache Karaf, Download. Read October 29, 2026.

17. Apache Libcloud is a standard Python library that abstracts away differences among. Read October 29, 2026.

18. projects.apache.org, projects.json. Read October 29, 2026.

19. Apache OpenServerless™. Read October 29, 2026.

20. Apache OpenWhisk is a serverless, open source cloud platform. Read October 29, 2026.

21. Index of /openwhisk. Read October 29, 2026.

22. Apache VCL, Apache VCL. Read October 29, 2026.

23. Apache VCL, Downloads. Read October 29, 2026.

24. Welcome to Apache YuniKorn, Apache YuniKorn. Read October 29, 2026.

25. Apache ZooKeeper. Read October 29, 2026.

26. Index of /zookeeper/zookeeper-3.9.6. Read October 29, 2026.

27. Incubation Policy. Read October 29, 2026.

28. Apache CloudStack Downloads, Apache CloudStack. Read October 29, 2026.

29. Installation overview , Apache CloudStack 4.23.0.0 documentation. Read October 29, 2026.

30. Quick Installation Guide , Apache CloudStack 4.23.0.0 documentation. Read October 29, 2026.

31. Occupational Employment and Wage Statistics Profiles. Read October 29, 2026.

32. Bureau of Labor Statistics Data. Read October 29, 2026.

33. Droplet Pricing, DigitalOcean. Read October 29, 2026.

34. B2 Cloud Storage Pricing, Backblaze. Read October 29, 2026.

35. api.linode.com, types. Read October 29, 2026.

36. vultr.com, pricing. Read October 29, 2026.

37. Apache OpenWhisk Downloads. Read October 29, 2026.

38. openwhisk-deploy-kube file README.md on GitHub. Read October 29, 2026.

39. openwhisk-deploy-kube file docs/k8s-technical-requirements.md on GitHub. Read October 29, 2026.

40. openwhisk file core/standalone/README.md on GitHub. Read October 29, 2026.

41. AWS Lambda Pricing. Read October 29, 2026.

42. Apache Karaf, Get started. Read October 29, 2026.

43. Apache Karaf Container 4.x, Documentation. Read October 29, 2026.

44. App Platform Pricing, DigitalOcean. Read October 29, 2026.

45. App Platform Limits, DigitalOcean Documentation. Read October 29, 2026.

46. Apache Airavata Installation, Apache Airavata 0.18. Read October 29, 2026.

47. SciGaP Admin Portal. Read October 29, 2026.

48. Requirements · GitBook. Read October 29, 2026.

49. Running Apache Brooklyn · GitBook. Read October 29, 2026.

50. Apache Helix , Quickstart. Read October 29, 2026.

51. Apache VCL, VCL 2.5.2 Installation Guide. Read October 29, 2026.

52. Desktop as a Service Pricing, Amazon WorkSpaces, AWS. Read October 29, 2026.

53. Deployment Modes, Apache YuniKorn. Read October 29, 2026.

54. Kubernetes Pricing, DigitalOcean. Read October 29, 2026.

55. ZooKeeper: Because Coordinating Distributed Systems is a Zoo. Read October 29, 2026.

56. Install, Apache OpenServerless™. Read October 29, 2026.

57. Apache CloudStack Security, Apache CloudStack. Read October 29, 2026.

58. Apache OpenWhisk Security. Read October 29, 2026.

59. Index of /security. Read October 29, 2026.

60. Security, Apache ZooKeeper. Read October 29, 2026.

61. ASF Security Team, Apache Software Foundation. Read October 29, 2026.

62. The Apache Attic, The Apache Attic was created in November 2008 to provide process and. Read October 29, 2026.

63. Apache Karaf, Community. Read October 29, 2026.

64. apache.org, LICENSE 2.0. Read October 29, 2026.

65. Claude Max plan. Read October 29, 2026, 23:31:40 UTC; supplied research.

66. ChatGPT pricing. Read October 29, 2026, about 23:27 UTC; exact minute not independently recorded; supplied research.

67. Claude API pricing. Read October 29, 2026, 23:31:14 UTC; supplied research.

68. OpenAI API pricing. Read October 29, 2026, 23:31:13 UTC; supplied research.

69. OpenAI prompt caching. Read October 29, 2026, 23:32:00 UTC; supplied research.

70. Codex CLI. Read October 29, 2026, 23:32:01 UTC; supplied research.

71. Codex with a ChatGPT plan. Read October 29, 2026, 23:31:16 UTC; supplied research.

72. Desktop Commander documentation. Read October 29, 2026, 23:31:41 UTC; supplied research.

73. Desktop Commander pricing. Read October 29, 2026, 23:31:08 UTC; supplied research.

74. Claude in Chrome. Read October 29, 2026, 23:31:11 UTC; supplied research.

75. Claude in Chrome permissions. Read October 29, 2026, 23:31:09 UTC; supplied research.

76. ChatGPT cloud browser. Read October 29, 2026, 23:32:01 UTC; supplied research.

77. METR early 2025 developer experiment. Read October 29, 2026, 23:30:58 UTC; supplied research.

78. Anthropic Consumer Terms. Read October 29, 2026, 23:31:08 UTC; supplied research.

79. OpenAI Terms of Use. Read October 29, 2026, about 23:27 UTC; exact minute not independently recorded; supplied research.

Apache, the Apache feather logo and the Apache project names are trademarks of The Apache Software Foundation. This site is not affiliated with or endorsed by the ASF. Other product names are used for identification only.