Does self-hosting Onyx cost less than Amazon Q Business?
The answer depends on user count, index size, model choice and the staff effort needed to run the stack, not on license price alone. A cost review triggered by a renewal, a licensing change or a residency requirement should compare those four drivers side by side. Amazon Q Business charges through its subscription tiers and index capacity, so the comparison has to be built from your own quote and usage figures, which you should request from AWS directly before modeling anything. Onyx Community Edition is MIT licensed, so the software itself carries no per-user license fee. The costs that remain are recurring and sit with you: compute and storage for the document index (Onyx's documented infrastructure includes Postgres, Vespa or OpenSearch, Redis and model servers), the model source, whether that is GPUs you run with a serving engine such as vLLM or API calls to a proprietary provider, backup, upgrades and monitoring. The model line is often the largest variable, and it scales with query volume and the models you choose, so ask your team to size it against real query logs rather than assumptions. The cost that is easiest to underestimate is people. Someone must own deployment, connector credentials, permission sync checks, upgrades and incident response. Ask any implementation partner for a fixed-scope written quote that lists each component, who pays for it, which items are one-time and which recur, and what changes the price if user count or index size grows. Compare that against the full Amazon Q Business bill, including any index capacity charges, before deciding.
Which cost components change when you move from Amazon Q Business to Onyx
This table sets each Amazon Q Business cost component against its self-hosted Onyx counterpart and marks whether it is one-time or recurring. Read it as a checklist for a quote: confirm each row with AWS billing and with your own infrastructure and staff before comparing totals, because the Q Business tier prices and index unit-hour rates are not stated here.
| Cost component | Amazon Q Business (managed) | Self-hosted Onyx (your infrastructure) | One-time or recurring |
|---|---|---|---|
| User subscription | Lite or Pro per-user subscription, billed by AWS; confirm current rates in the AWS pricing terms for your region | No per-user license for Onyx Community Edition (MIT); Enterprise Edition features such as RBAC are a separate license to confirm | Recurring |
| Index capacity | Starter or Enterprise index tiers billed in unit-hours; ask AWS how many unit-hours your data volume implies | Index sized by your own Vespa or OpenSearch nodes, storage and replicas, so cost tracks hardware and storage size | Recurring |
| Core infrastructure | Included in the AWS service price; no database, cache or queue to run | Postgres, Redis, the index engine, object storage and model servers run in your cloud account or data center | Recurring |
| Model inference | Model usage is covered inside the Q Business service; confirm what is metered and what is excluded | GPU capacity for self-hosted models through vLLM or Ollama, or API token charges if you use a proprietary provider such as OpenAI or Anthropic | Recurring |
| Implementation and migration | Connector setup and app rebuilds are largely configuration work within AWS | Scoped deployment, connector reconnection, permission sync testing and cutover, usually quoted as fixed-scope work | One-time |
| Support and upgrades | Included in the AWS service and support plan you already hold | Vendor support is paid separately from the open-source project; upgrades and patching are your team's or a retainer's work | Recurring |
| Internal skills and staff time | Administration happens in the AWS console with limited operations overhead | Staff who run Kubernetes or Docker, databases, GPU nodes, monitoring and access control; ask who will own each layer | Recurring |
What moves the comparison most
Six variables decide whether a self-hosted Onyx deployment costs more or less than Amazon Q Business at your scale. Each one pulls the total in a predictable direction, so the quote you request should price every one of them explicitly.
Number of named users
Amazon Q Business charges per user on its Lite and Pro subscriptions, so seat count scales its license line directly. Self-hosted Onyx has no per-seat license, so user growth mainly adds load on chat, search and model capacity rather than a new fee per person. Ask the quote to state peak concurrent users, not only named accounts, because concurrency drives serving capacity. Confirm current Q Business subscription terms with AWS before comparing.
Volume and churn of indexed content
The more documents you index and the faster they change, the more work the index, sync workers and embedding step perform. Amazon Q Business meters index capacity in unit-hours on its Starter and Enterprise tiers, so heavy churn raises that line. With Onyx, the same volume raises storage, index size and compute on your own servers. Ask for the number of sources, the expected document count and how often each source changes.
Onyx Standard or Onyx Lite
Onyx Standard includes vector and keyword indexing, background sync workers, model inference servers, object storage and Redis, so it carries the largest infrastructure footprint. Onyx Lite drops the vector database and background workers, which lowers infrastructure cost but removes the sync and index capabilities that replace connector-based search. Choose by whether you need permission-synced search over company content or only chat and agents. Confirm which edition the quote assumes.
Hosted API or self-hosted vLLM
Calling a hosted model API turns model cost into per-use charges that rise with query volume and token length. Running vLLM on your own GPUs moves that cost into fixed accelerator capacity, plus the staff who operate it. vLLM itself has no license fee, so the decision is about utilization: low, bursty traffic often favors APIs, while steady high demand can favor owned or reserved GPUs. Ask the quote to model both paths at your expected query volume.
Availability and residency requirements
Multi-region, high availability or data residency requirements add duplicate infrastructure, backups and failover testing to a self-hosted stack, which raises cost. Amazon Q Business carries these controls within AWS regions, so the comparison depends on which regions and recovery targets you must meet. Ask the quote to list each environment, the backup schedule and the failover design, and confirm which residency obligations apply to model inference as well as stored data.
Team skills and operating time
A self-hosted stack moves patching, upgrades, monitoring and on-call from AWS to your staff or an implementation partner. Teams already running Kubernetes, PostgreSQL and GPU clusters carry this cost at lower marginal effort than teams starting from scratch. Estimate the people hours for installation, connector maintenance and routine upgrades, and ask any partner to separate one-time setup from recurring support. Those hours, not the software, usually make up the largest part of the difference.
How the self-hosted Onyx stack is built and where its costs sit
The cost of a self-hosted Onyx deployment depends first on which Onyx shape you choose. Onyx Standard runs vector and keyword indexing, background sync workers, model inference servers, object storage and Redis, so every connector you keep syncing and every query you run lands on infrastructure you pay for. Onyx Lite is a lighter chat and agents deployment without a vector database or background workers. It suits a team that needs chat over a small, mostly static set of sources, but it does not give you the continuous connector sync that replaces an always-current Q Business index. Decide which shape you need before you price anything, because it changes the number of services you run and patch.
Map each service to a cost line. The index and metadata store is Postgres, plus Vespa or OpenSearch for the search index, so storage and memory for the index grow with document volume and sync frequency. Redis and object storage are smaller lines, but they still need backup, monitoring and capacity planning. Model serving is the largest variable. If you host open-weight models, vLLM is Apache-2.0 licensed with no paid edition, so its cost is GPU or other accelerator capacity, the surrounding infrastructure, implementation and the people who operate it. Its vllm serve process exposes OpenAI-compatible endpoints such as /v1/chat/completions and /v1/embeddings, which lets Onyx call your own server the same way it would call a hosted API. If you use a proprietary API instead, the per-token charges replace the GPU line.
Ask your implementation partner and your cloud team to produce a quote that separates these lines: index and database hosting, Redis and object storage, GPU or API model consumption, backup and monitoring, and operator time for upgrades. Confirm with Onyx which components your chosen shape requires and which Helm chart version your deployment targets, since the project publishes releases often. Ask who patches the vLLM nodes, because vLLM's security documentation says inter-node traffic in multi-node runs is unauthenticated and unencrypted by default and must sit on an isolated network. Those controls are recurring work, and they belong in the comparison beside Amazon Q Business subscription fees.
GPU capacity, API models or Onyx Lite: which model cost path fits
Running open-weight models on your own GPUs with vLLM moves the main recurring cost into accelerator capacity. You pay for the hardware or cloud instances that stay on whether or not anyone asks a question, plus the networking, storage and staff who keep the serving stack alive. vLLM itself has no paid edition, so the license is not the cost. Onyx connects to vLLM through its OpenAI-compatible endpoints, which keeps prompts and documents inside your environment.
- Recurring driver: GPU or accelerator hours for the reserved capacity, sized to peak concurrent users rather than average traffic.
- Operating effort: driver and CUDA or ROCm version management, model upgrades, capacity planning and monitoring through the Prometheus metrics at
/metrics. - Security work: in multi-node runs, inter-node traffic is unauthenticated by default, and the
--api-keyoption does not cover every route, so a reverse proxy allowlist is needed. - Release pace: vLLM ships new minor releases about every two weeks, so you need a rule for which versions you test before rolling out.
- Ask your team or vendor: what GPU class and count serve your target concurrency, and who is on call for the serving tier.
Confirm the hardware requirements for your chosen model with the model publisher and your hardware vendor before you size anything.
Onyx can call hosted proprietary models from Anthropic, OpenAI and Gemini, so you pay the providers per use instead of for idle capacity. The cost scales with the questions asked and the context each answer pulls in, which means retrieval settings and document volume matter as much as the model choice. Your own platform stays in your infrastructure, but prompts and retrieved content leave it for the provider, so the provider's contract and data terms become part of your review.
- Recurring driver: token consumption for prompts, retrieved passages and answers, multiplied by the number of queries.
- Operating effort: lower than running GPUs, mainly key rotation, spend monitoring and model version changes that the provider announces.
- Data review: check retention, training use and regional processing terms in each provider's enterprise documentation before connecting sensitive sources.
- Cost control: set per-user or per-team usage limits at the provider and watch how many chunks each query retrieves.
- Ask the provider: what usage tiers, commitments and data processing terms apply to your region and volume.
Ask each provider for a written statement of retention and training terms that covers your specific contract, not only its public pages.
Onyx Lite is a lighter chat and agents deployment without a vector database or background workers, according to Onyx's deployment documentation. It reduces the number of services you run and patch, which lowers the fixed infrastructure cost and the operating effort. It does not replace the need for a model path, so you still pay for GPUs or API usage. Lite is a fit when chat and agents matter more than broad indexing of company content, because connector-based search depends on the components Lite leaves out.
- Recurring driver: the model path you choose, plus a smaller set of always-on services than Onyx Standard.
- Trade-off: connector sync, vector and keyword indexing and background workers belong to Onyx Standard, so Lite is not a like-for-like replacement for permission-aware search across sources.
- Operating effort: fewer components to back up, upgrade and monitor, but a different architecture to validate against your requirements.
- Upgrade path: plan how you would move from Lite to Standard if your indexing needs grow.
- Ask Onyx: which features are absent in Lite, how the database and model server are sized, and whether Lite supports the permission and SSO controls you need.
Confirm the feature list for the Onyx version you plan to deploy, because the deployment documentation changes with releases.
How to build the cost comparison with your own numbers
This sequence turns a vague sense that self-hosting might be cheaper into a three-year comparison built from your own Amazon Q Business billing, your own source inventory and your own infrastructure rates. Each step produces a figure you can defend in a budget review.
Pull current Amazon Q Business subscription and index records
The Amazon Q Business administrator and the finance or AWS billing owner export what you pay today. Collect the number of Lite and Pro subscriptions by user group, the index capacity you provision (Starter or Enterprise index unit-hours), and the billing line items for the last full billing period, using AWS Cost Explorer or your invoice. Confirm each unit price with your AWS account team, because list rates and negotiated rates can differ. Done when every line of the current monthly cost traces to a usage record you can show a reviewer.
Count indexed sources, documents and active users
The Q Business admin and the content owners list each connected data source, such as S3 buckets, SharePoint, Confluence or Salesforce, and record the approximate document count, total storage size and rate of change for each. Count active users over the same period and note peak concurrency, since query volume drives model cost. Record which sources need permission sync so the Onyx side reflects the same access scope. Done when you have one table of sources, sizes and user counts that the rest of the model reads from.
Define the Onyx tier and model path
The architecture owner chooses between Onyx Standard and Onyx Lite, and between a self-hosted model served with vLLM or Ollama and a proprietary API provider. Standard includes vector and keyword indexing, background sync workers, model inference servers, object storage and Redis, while Lite omits the vector database and background workers. Decide which edition you need, whether you need the Enterprise Edition features such as RBAC, and whether you will self-host the model. Write the choice down, because each option changes the infrastructure and GPU lines in the next steps. Done when the tier and model path are recorded as explicit assumptions.
Price infrastructure at your own cloud or data center rates
The platform or cloud team prices each component against your own contract or hardware cost: compute for the Onyx services, Postgres, Vespa or OpenSearch for the index, Redis, object storage, and for self-hosted models, GPU or accelerator capacity sized for your expected concurrency. Use your organization's cloud discount, reserved capacity or depreciation schedule, not public list prices. Ask your cloud provider or hardware vendor for quotes, and if you use API models instead, price tokens with the provider's current rate card. Done when each infrastructure line has a monthly figure and a stated rate source.
Estimate staff hours for operations and upgrades
The platform engineering lead estimates recurring operations effort: monitoring, backups, connector maintenance, patching, Onyx and model server upgrades, and on-call coverage. Count hours per month for each activity, then convert them with your internal loaded cost per hour. Note that vLLM ships minor releases about every two weeks, so upgrade effort is a recurring line, not a one-time one. If you retain DeployOpen or another partner for support or upgrades, add that quote as a separate line. Done when recurring staff hours and cost are in the model with their assumptions listed.
Add one-time implementation and migration costs
Separate the one-time work from the recurring costs. This includes deployment, connector and permission sync configuration, identity integration with your SSO provider, security hardening, testing, user onboarding and migrating or rebuilding Q Apps and guardrails. Ask each implementation provider for a fixed-scope written quote, and record internal effort in hours. Done when one-time costs sit in their own column, so they can be amortized over the comparison period without being mixed into monthly run rates.
Compare three-year totals and test the assumptions
Build one table with the current Amazon Q Business cost and the self-hosted Onyx cost across the same three years, including one-time and recurring lines. Then vary the two inputs with the widest uncertainty, usually user count and GPU or API usage, and check whether the conclusion holds. Share the assumptions with finance, security and the AWS account team so the comparison reflects the organization's own position. Done when the three-year totals, the sensitivity results and the list of confirmed quotes are approved in writing by the budget owner.
Common questions about the cost of replacing Amazon Q Business with Onyx
Does Onyx Community Edition carry a software license fee?
No license fee applies to the Onyx Community Edition itself, because it is MIT licensed and covers RAG, AI chat, agents and actions. Code in the ee directories falls under the Onyx Enterprise License, so confirm with the project which features you need and whether the Enterprise Edition, with premium features such as RBAC, is required for your controls. Your recurring costs then come from infrastructure, model usage and the people who operate the stack, not from a per-seat subscription to Onyx. Ask the vendor or DeployOpen to list which features map to the Community Edition and which need the Enterprise License before you compare totals.
What drives the infrastructure cost of a self-hosted Onyx deployment?
Infrastructure cost follows the components you choose and how much content you index. Onyx Standard runs vector and keyword indexing, background sync workers, model inference servers, object storage and Redis, alongside Postgres and Vespa or OpenSearch. Onyx Lite drops the vector database and background workers, so it costs less to run but serves a narrower purpose. Drivers include the number of connected sources, document volume, sync frequency, query load and whether you run model servers locally. Ask for a sizing estimate tied to your document count, user count and sync schedule.
Do I need GPUs, or can Onyx use API models instead?
You do not need GPUs to run Onyx, because it supports proprietary providers such as Anthropic, OpenAI and Gemini through API keys. GPUs become relevant when you host open models yourself through Ollama, LiteLLM or vLLM. vLLM runs on NVIDIA CUDA, AMD ROCm and Intel XPU GPUs and on some CPUs, and its cost comes from accelerator capacity, surrounding infrastructure and the people who operate it. Choosing API models shifts cost toward per-token usage, while self-hosted models shift it toward fixed hardware and operations. Model your expected token volume against both options before deciding.
Can our own team run Onyx without outside help?
A team with container, Kubernetes or Terraform experience can run Onyx, since it deploys with Docker, Kubernetes, Helm or Terraform and documents an air-gappable architecture. The effort lies in ongoing work: upgrades (the project publishes frequent releases), connector maintenance, index tuning, backups, monitoring, access reviews and, if you host models, GPU operations and inference hardening. If you run vLLM, its security documentation requires isolating multi-node traffic and placing a reverse proxy in front of user endpoints. Estimate the staff hours per month for these tasks and compare them with the cost of outside support.
How long does a migration from Amazon Q Business take?
The duration depends on scope: the number of data sources, how much permission mapping needs to be verified, the number of users, whether you host models and how many Q Apps or guardrails must be replaced. No duration figure is supplied for this work, so ask for a timeline in the written scope rather than relying on a generic estimate. Staff needs include an owner for identity and SSO, people who know each data source's permission model, and someone to validate answers against source-application access. Agree on acceptance tests before the work starts.
What should we ask for in a quote for this replacement?
Ask for a fixed-scope quote that lists each deliverable, the connectors in scope, the identity integration (Google OAuth, OIDC or SAML with SCIM), the deployment target and the model providers. Request the one-time costs separately from recurring ones, including infrastructure sizing, licenses for any Enterprise features, model usage assumptions and support or retainer terms. Also ask who pays for GPUs, backups, monitoring and upgrades, and how permission sync will be tested. DeployOpen works from a written scope and does not publish a price list, so the scope document is what makes the comparison possible.
Does Onyx keep answers within each user's existing source permissions?
Yes, Onyx syncs permissions from the connected applications, so answers only draw on documents the user can already see in the source system. This is central to comparing it with Amazon Q Business, because the cost of a replacement includes testing that permission sync works for each connector. Ask the implementation partner to define test cases per source, including revoked access and group changes, and to record the results as evidence. Permission sync does not replace your own access reviews, so budget for them as recurring work.
When the self-hosted case is weak
Amazon Q Business can remain the simpler and cheaper choice for some organizations, and a self-hosted Onyx deployment is not automatically the better answer. A small user base is the clearest case. Every self-hosted stack carries fixed components, including the index, Postgres, Redis, model serving and the object storage and backups around them, and those costs do not shrink when only a few dozen people ask questions. If your estate already runs on AWS IAM Identity Center, your sources are mostly AWS-native, and usage is light, the managed service may cost less to operate than a platform your team has to patch, monitor and upgrade. Confirm the current per-user and index unit-hour rates directly with AWS and Amazon, then compare them against a full cost estimate for your own environment rather than against the cost of a single server.
The self-hosted case also weakens when no one on your team can run the parts that need specialist skill. Onyx Standard depends on Postgres, a vector index (Vespa or OpenSearch), Redis and model inference servers, and self-hosted models add GPU capacity and an inference engine such as vLLM, which needs its own security hardening, such as a reverse proxy that exposes only user-facing endpoints. If you lack operators for GPUs and databases, ask a partner or your own staff to quote that operating load in writing before you commit. Onyx Lite, which drops the vector database and background workers, may suit a smaller pilot, but confirm its feature limits against your connector and permission requirements first.
Before committing to a multi-year model, run a pilot with a written scope. Name the user group, the three to five connectors that matter most, the permission behavior you must verify, the model options you will test and the exit criteria, including a measured monthly run cost and the staff hours spent. Agree in writing what happens to the data and the workspace if the pilot fails, and keep your Amazon Q Business contract on a short term while the comparison runs. If the pilot shows that the managed service meets your needs at lower total effort, that is a valid result and a reason to stay.
How DeployOpen scopes a fixed-price Onyx implementation
DeployOpen plans, implements, hardens, documents and supports self-hosted open-source platforms inside your own cloud accounts, data centers or air-gapped networks. For a move off Amazon Q Business, we start by agreeing in writing what the Onyx deployment covers, then build and hand it over with the evidence your team needs to run it.
How we scope
- Engagements are fixed-scope, with scope, deliverables and timelines agreed in writing before work starts.
- The written scope lists which Amazon Q Business data sources, users and permissions Onyx must reproduce.
- Cost components such as Onyx hosting, model serving, GPU or API usage and staff time are itemized in the quote.
How we build
- We deploy Onyx with Docker, Kubernetes, Helm or Terraform in your environment, matching the target infrastructure you specify.
- Connectors are configured for the sources in scope, and permission sync is tested against real user accounts before cutover.
- Model serving choices such as Ollama, LiteLLM, vLLM or a hosted API are set up and load-tested against your expected usage.
How we hand over
- You receive architecture documentation, configuration records, runbooks and the test results agreed in the scope.
- Administrators are trained on user provisioning through SSO and SCIM, connector management and query history review.
- Evidence for compliance assessments is prepared, while independent auditors and certification bodies issue any audit reports.
How we support
- Ongoing support, upgrades and continuous testing are available as a retainer after handover.
- Upgrades to Onyx releases and Helm chart versions are planned and tested in a staging environment first.
- Ask us to confirm how licensing applies to your edition, since Onyx Community Edition and Enterprise Edition features differ.
