Analytics Legends The knowledge platform for SAP Analytics
Guide

SAP AI Core pricing — the units the bill is actually built from

As of 2026-08-14

There is no public price list for SAP AI Core that a consultant can quote, and any page giving you a confident per-unit figure has invented it. What is knowable is the structure: consumption is metered in capacity units on the Business Technology Platform, and inference on top is metered in tokens through the Model Gateway that every Joule request and every custom model call passes through.

That structure determines the invoice, and it is enough to build a defensible estimate.

Why there is no rate card to quote

Our 15-tool analytics review prices the SAP AI layer as quote-only, with an estimated floor near thirty thousand US dollars a year when bundled with the SAP suite at small-workload scale. It labels that a floor anchor rather than expected spend, and adds the correction that matters: across a thirty-six-month total cost of ownership, expect 1.8 to 4.5 times the floor once capacity reservations, services, training and migration are counted. Stated as of Q1 2026.

The honest framing is not a price but a shape: a capacity commitment plus consumption, both quote-driven.

Meter one — capacity units

One capacity unit bundles roughly one vCPU-equivalent of HANA Cloud compute with about four gigabytes of memory, a storage allotment before overflow pricing, and a slice of data-flow throughput. It is sold in tenant-sized blocks rather than per resource, so the same unit trades off differently depending on the workload mix.

Our corpus records the smallest Business Data Cloud tenant at 128 capacity units against a 64-unit floor for standalone Datasphere. The anti-pattern is sizing to average utilisation: peak routinely doubles average, and the recorded tier-one pattern is 130 to 240 units steady against 280 to 360 at peak.

Meter two — tokens through the Model Gateway

Every Joule request, every generative flow in SAP Build and every custom model call passes through the Model Gateway, which owns authentication, rate limiting and usage tracking — and is the single metering point. The billing unit is the token, split into input and output tokens, priced differently.

Attribution is hard because one interaction is rarely one call: asking Joule to summarise open action items can fire a retrieval step, a summarisation step and a formatting step. Our corpus puts a single interaction at 8,500 to 23,600 tokens across three to four calls, none tagged with a business scenario unless someone instrumented it at request time.

What a scenario costs, and the lines nobody budgets

Every figure here is a modelled scenario from our corpus, not a quotation. One business scenario at ten thousand calls is recorded at roughly €25 to €250 a month; ten thousand users at twenty interactions a day, on mid-tier pricing, implies €290,000 to €875,000 a year. Agentic chains are recorded at fifteen to forty times a single-turn retrieval interaction, hence the recommended four-to-six turn cap.

Three lines rarely appear in a business case: idle Joule agents polling for availability consume an estimated five to ten percent of tenant capacity with zero conversations; a large static document in the system prompt on every call is often the single largest controllable lever; and retried calls still consume tokens invisibly.

Making the bill attributable before it arrives

The structural mistake is provisioning a single resource group for every workload: simpler on day one, it makes attribution to a business domain effectively impossible. The pattern that survives an audit is one resource group per business domain crossed with one per environment.

At the Gateway, the resource-group header tags consumption by named group, which lets platform cost management slice spend by domain. Attribution cannot be reconstructed afterwards — untagged token logs will not become a chargeback report.

What we cannot assert

We cannot publish a per-capacity-unit or per-token rate for SAP AI Core: SAP discloses none, and every euro figure here is a modelled scenario from a named card with its assumptions stated, not a quotation. The floor anchors in our 15-tool review are explicitly floors as of Q1 2026, not expected spend. Elastic capacity bursting is recorded as preview with a GA target of H2 2026 and should not be assumed in a 2026 sizing.

Frequently asked

How much does SAP AI Core cost?

SAP publishes no rate card we can quote; the answer is quote-driven. The structure is a capacity commitment metered in capacity units plus token consumption metered through the Model Gateway. Our 15-tool review, as of Q1 2026, anchors the layer at an estimated floor near thirty thousand US dollars a year at small scale.

What is a capacity unit, and how many does a tenant need?

One capacity unit bundles roughly one vCPU-equivalent of HANA Cloud compute with about four gigabytes of memory plus storage and throughput. Our corpus records a 128-unit floor for the smallest Business Data Cloud tenant against 64 for standalone Datasphere.

What does one Joule interaction cost in tokens?

Our corpus records 8,500 to 23,600 tokens across three to four model calls — retrieval, summarisation and formatting. Agentic chains are recorded at fifteen to forty times a single-turn retrieval interaction.

How do I attribute AI spend to a business unit?

Tag it at request time with the resource-group header, and provision one resource group per business domain crossed with one per environment. Attribution cannot be reconstructed from untagged logs.

What this page is built on

Read next