Choose an AI provider by verifying the uptime commitment for the exact model service and deployment you plan to use, then checking its measurement rules, exclusions, data terms and model-change policy. Separately, write down how your team would export its assets, replace provider-specific components and keep the service running during a move. Contractual uptime targets are not evidence of which provider will perform best for your workload, and portability is most useful when it is proportionate to the risk you need to manage.
Start with the workload, not the provider’s headline SLA
An availability commitment applies to the services and conditions named in its contract—not automatically to every model, endpoint, region or application a provider offers. Before comparing providers, specify the production workload you need covered:
As an Amazon Associate I earn from qualifying purchases.
- The model, API and deployment type, including whether you use a managed endpoint, batch processing or a training service.
- The region, account or project, and any cross-region routing involved.
- The production traffic pattern, expected request volume and consequences of an outage.
- The application-level availability you need, including acceptable latency, error rates and recovery time.
Then check whether the relevant SLA names that service and configuration. If the contract covers a different product or only a particular deployment mode, its percentage is not a commitment for your workload.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →Read the SLA’s measurement rules and remedy
A percentage alone leaves out the details that determine whether an incident qualifies. Compare each contract’s covered service, error definition, measurement interval, aggregation period, exclusions, claim procedure and remedy. In particular, establish how the provider treats partial failures, idle periods, regional incidents and errors outside its own infrastructure.
#1 Best Overall
| Provider example | Published contractual terms | What to verify |
|---|---|---|
| Google Vertex AI | Google’s SLA lists 99.9% for training, deployment and batch prediction; 99.5% for certain custom online prediction configurations and Vertex Pipelines; and 99% for the training cluster control-plane API. These are targets for the named covered services, not a blanket commitment for every Vertex AI model or endpoint. For covered services, the document defines “Downtime” as more than a 5% server-side error rate. A credit request must be made within 30 days of eligibility. | Match the exact service and configuration to the SLA row. Review the full definitions, exclusions and credit conditions before relying on a target. |
| Amazon Bedrock | AWS’s SLA, last updated October 4, 2023, calculates monthly uptime per AWS Region, defines an error as a request returning HTTP 500, and describes service-credit tiers when monthly uptime falls below 99.9%. It lists exclusions and a claim process; absent another agreement, service credits are the stated remedy. | Because the cited SLA page is dated, recheck its current terms and your governing customer agreement. Confirm how its regional calculation and exclusions apply to the specific workload. |
These are contract terms, not like-for-like outage histories. Their definitions, covered services and exclusions differ, so the percentages do not establish which provider has better observed uptime. The cited materials do not provide a comparable historical uptime ranking for these providers.
For the current contract language, see Google Cloud’s Vertex AI SLA and AWS’s Amazon Bedrock SLA.
Assess operational availability for your own service
Even a clear SLA is only one part of reliability. A service can satisfy its contractual definition while your users experience a problem the definition does not capture—for example, an application-level failure caused by a dependency, a quota limit or a latency spike. Credits may also be a poor substitute for the business value lost during an interruption.
Before selection, check the provider’s status and incident information, relevant quotas, regional and endpoint availability, and the service’s recovery options. Instrument your own application so you can see successful requests, errors, latency and dependency failures from the user’s perspective. Set service targets that reflect your actual need; do not infer them from a provider’s SLA percentage alone.
Check model lifecycle, data controls and geography
Plan for model changes
Models can be retired or replaced, and migration considerations may vary by region, cloud environment and security requirements. Microsoft’s Foundry lifecycle guidance recommends evaluating newer models rather than waiting for an official replacement. Treat lifecycle notices as an input to a repeatable evaluation process: test candidate models against representative tasks, review quality and safety, and check integrations before switching. See Microsoft’s Foundry Models lifecycle and support policy.
Verify retention and residency for the exact deployment
Data terms can depend on the model as well as the service. Amazon Bedrock documents model-specific allowed retention modes; if the effective setting does not meet a model’s requirement, that model may be unavailable or invocations may fail. Check the effective retention setting, model eligibility and applicable terms before assuming your preferred configuration works across models. See AWS’s Bedrock data-retention documentation.
For OpenAI models used through Bedrock specifically, OpenAI says AWS maintains Bedrock deployment options and availability, and directs customers to check model and endpoint availability, cross-Region inference and quotas. It also cautions that an AWS Region is not, by itself, an OpenAI data-residency jurisdiction. This is a deployment-specific qualification; confirm the terms for the exact service, model, deployment and contract. See OpenAI’s guide to OpenAI on Amazon Bedrock.
Free tools Windows power users keep installed
One-click scans. No signup required.
Make the exit path concrete before procurement
An exit plan should answer what can move, what must be rebuilt, how the transition will be funded and how the service will continue during the change. Ask the vendor and internal service owner to walk through the following points together:
- Inventory dependencies. Record provider-specific model endpoints, prompt or orchestration services, retrieval stores, identity controls, monitoring and data pipelines. Note why each dependency is accepted and what would replace it.
- Specify exports. For each dataset and asset, identify the export method and documented format, who can perform it, how long it takes, and any charges or bandwidth limits. Include backup location and how long the data remains available after the contract ends. AWS’s portability guidance recommends understanding these practical details.
- Cover more than raw data. Identify which metadata, logs, configuration, evaluation assets and application code must be returned, retained or recreated. Confirm contractual return and deletion obligations, including what evidence of deletion is available.
- Separate movable from rebuildable components. Version prompts and modifiable application assets. Mark which components can be transferred directly and which need reimplementation or revalidation. A data export does not prove that a replacement model will behave equivalently.
- Assign owners and continuity steps. Name the accountable service and business owners, estimate transition duration and cost, and document how the service will operate during migration. Decide when the plan will be reviewed or exercised, based on workload criticality and sensitivity.
- Price the flexibility. Compare the ongoing engineering, operating and redundancy costs of a more portable design with the risk it reduces. Record provider dependencies you deliberately accept and why.
The Australian Government Architecture guide to managing lock-in, portability and exit planning calls out export, data-return obligations, migration costs, continuity, ownership and testing. It also cautions that “Maximum portability is not always the best outcome.” UK government guidance on managing technical lock-in in the cloud similarly frames exit planning as a trade-off: weigh the plan’s cost against the value of staying with the service.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Keep portability proportional to the risk
There is no requirement to make every component interchangeable. A critical service handling sensitive data may justify tested exports, a fallback route and regular migration exercises. A lower-impact workload may need only documented dependencies, a verified export route and a named owner. Abstraction layers, multi-provider support and duplicated infrastructure can reduce some dependencies, but they also add engineering and operational complexity.
Use the workload’s continuity requirements, sensitivity and provider dependence to decide how much portability to buy. Capture the dependencies you accept, the controls that offset them and the conditions that would trigger a new review. For practical operating considerations, Google Cloud’s guide to deploying and operating generative AI applications discusses the operational work around these systems.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallOutdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchTurn the review into a provider decision
For each candidate, retain the actual SLA and contract, map the covered service to your intended deployment, and document any gaps between contractual availability and your application target. Record data and lifecycle constraints alongside export steps, transition ownership and estimated continuity needs. If you cannot establish a comparable operational history, say so in the decision record rather than treating different SLA percentages as a performance ranking.
Approve the provider only when the team can explain both how it will operate the workload under normal conditions and what it will do if the model, terms or service no longer meet its needs.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
