SAP AI Core & AI Launchpad

What you will learn

• Explain the AI API object model — scenario, executable, configuration, execution, deployment — and map each object to what you see in the AI Launchpad ML Operations app
• Choose between the Standard and Extended plans, knowing that Extended cannot be downgraded, and design resource groups within the documented quotas
• Explain both cost models from SAP documentation (tokens → GenAI tokens → capacity units for generative AI; node hours, storage and a capped hourly baseline for custom models) and estimate with the cost calculator and BTP Usage export
• Describe how a bring-your-own model runs: Argo workflow templates, KServe serving templates with their mandatory metadata, Git repository, application, docker and object-store secrets
• Track the platform's lifecycle — model deprecations in SAP Note 3437766, orchestration v1 decommissioning on 31 October 2026, and new SAP foundation models (SAP-RPT-1.5/1.6, TabPFN-3.5 Plus)

Module overview

SAP AI Core is the runtime; SAP AI Launchpad is the control room. AI Core is a BTP service that runs AI workloads in resource-isolated tenants: it gives you managed access to foundation models through the generative AI hub, and it runs your own models through Git-synced training workflows and serving templates. AI Launchpad is a separate BTP SaaS application that connects to one or more AI Core instances and gives administrators, prompt engineers and ML engineers a UI over the same API. Keep the split clear on delivery: why is inference failing? is a runtime question; who deployed this, with which configuration, and when? is a control-room question — and both are answered from the same objects (scenarios, executables, configurations, deployments, executions) exposed by the AI API.

Prerequisites

• Intermediate hands-on experience on SAP BTP (subaccounts, entitlements, service keys)
• Review core concepts first: C026, C119, C118

Outcomes

• Walk a client through the AI Core object model and show each object live in AI Launchpad.
• Produce a platform design note: plan, resource groups, model routes, pinned versions with deprecation dates, observability and cost monitoring.
• Estimate and explain consumption for a generative and a custom-model workload without quoting invented prices.
• Recognise and avoid the five recurring mistakes: one resource group for everything, idle serving deployments, model-named configurations, floating model versions, 'the platform is compliance'.