Concept card

SAP AI Core

As of 2026-09-27

Edited by Cédric Mary, sole editor and publication director — 27 years in the SAP ecosystem. How these pages are produced, and where automation is used.

What is SAP AI Core?

SAP AI Core is the SAP BTP runtime that executes AI workloads — Argo-based training and batch runs, KServe-based model serving — and, on the extended plan, hosts the generative AI hub. Everything hangs off one tenant per subaccount, split into resource groups; a service key is the entry ticket; consumption is billed in capacity units.

What it is

SAP AI Core is a service of SAP Business Technology Platform that manages the execution and operation of AI assets "in a standardized, scalable, and hyperscaler-agnostic manner" (SAP AI Core service guide, September 2026). It is available in the Cloud Foundry, Kyma and Kubernetes environments and works together with two siblings: SAP AI Launchpad, a multitenant SaaS console that manages AI scenarios across one or more runtimes and gives access to the generative AI hub, and the AI API, the open lifecycle specification that both speak. Under the hood AI Core runs Kubernetes, uses Argo Workflows for pipelines and KServe for serving, syncs templates from your Git repository, pulls images from your Docker registry and reads and writes artifacts in your object store. The generative AI hub is not a separate product: SAP describes it as a capability of AI Core and AI Launchpad that "extends" AI Core, and it is only available on the extended service plan. The trade-off is one-way and easy to under-price: standard is cheaper while a project is genuinely custom-AI-only, but the moment any generative AI use case is even plausible on the roadmap, extended is the safer default, because upgrading later keeps data and models while downgrading is not possible at all.

Why it matters

Resource-group design decides cost allocation and access separation for the whole programme — and runtime objects cannot be moved between resource groups afterwards.
The plan choice is one-way: standard can be upgraded to extended but never downgraded, and without extended there is no generative AI hub.
AI Core's audit log covers management events, not prompts — evidence of what a model saw requires inference observability, switched on per request.

Key points

AI Core = BTP runtime for AI workloads on Kubernetes: Argo Workflows for training/batch executions, KServe for serving, Git + ArgoCD for templates.
One AI Core tenant per subaccount; several service instances in the same subaccount share that tenant.
Resource groups isolate runtime objects (configurations, executions, deployments, artifacts); executables, scenarios and Docker secrets are tenant-wide. Default limit: 50 resource groups.
Scenario → executable → configuration → execution (run) or deployment (inference URL): the core AI API vocabulary.
Plans: standard (custom AI, no GenAI) and extended (adds the generative AI hub). Upgrade only; the free plan ended on 22 May 2026, a 30-day trial remains.
Service key = clientid/clientsecret (or x.509) + XSUAA url + AI_API_URL; credential rotation is the customer's job.
Billing in capacity units: node hours, storage and a baseline capped at 730 h/month for custom AI; token-based metering for foundation models and grounding.
ISO/IEC 42001 covers Joule, AI Core and AI Launchpad; the audit log records management events, prompt payloads only via opt-in inference observability.

Terms used on this page

AI API: Open specification for managing AI artifacts and workflows across runtimes; SAP AI Core is one runtime implementation, SAP AI Launchpad one client.
Resource group: Isolation unit inside an AI Core tenant, mapped to a Kubernetes namespace; owns configurations, executions, deployments, artifacts and its secrets.
Scenario: Grouping of the executables that implement one AI use case; SAP-managed global scenarios include foundation-models and orchestration.
Executable: Reusable workflow or serving template with placeholders; non-deployable ones yield executions, deployable ones yield deployments.
Configuration: Binding of concrete parameter values and artifact versions to an executable; required to start an execution or deployment.
Execution: One run of a workflow executable — training or batch inference — producing output artifacts such as models or result sets.
Deployment: Instance of a serving executable that creates a model server with an HTTPS inference URL.
Capacity unit (CU): SAP BTP billing metric; AI Core converts node hours, storage, baseline tenant hours and GenAI tokens into CU.

Sources

Full card available to members. What the full card adds: the full decision framework · the SAP vs Snowflake / Databricks / Fabric comparison · the common pitfalls and their fix · the cheat sheet · the architecture schemas · the code blocks · the facts worth quoting.

Open in the app →

Guides that answer with this page

These guides cite this page as one of the sources their answer rests on.