SAP Cheaper AI Models Target Over €2 Billion in Gains
Key Takeaways
- SAP is replacing some expensive AI models with cheaper alternatives where performance remains strong, chief financial officer Dominik Asam told Reuters on October 28, 2026.
- SAP expects the move to help generate more than €2 billion in AI-driven efficiency gains over the next two years.
- SAP uses more than 70 AI models from providers including OpenAI, Google and Anthropic, and increasingly chooses the cheapest model capable of a task to the required standard.
- Unlike many AI providers, SAP charges customers per completed task, not per unit of compute, so lower operating costs can improve profitability.
- Overall AI spending is still expected to rise as adoption grows; the goal is healthy margins at scale, not an absolute reduction in AI expenditure.
SAP has a refreshingly unfashionable message for the AI industry: the best model for a task is not always the most powerful one, and sometimes it is the cheapest one that does the job well enough. In comments to Reuters on October 28, 2026, chief financial officer Dominik Asam said SAP is replacing some expensive AI models powering its products with cheaper alternatives, a move it expects will help generate more than €2 billion in AI-driven efficiency gains over the next two years.
The wording of that expectation matters, and Reuters corrected its own initial wording to say the move will help generate, not generate, the gains. The €2 billion figure is therefore a contribution to efficiency gains over two years, not a standalone profit figure, and it is a company expectation, not a reported result. Readers should keep both qualifications in mind when quoting it.
How SAP chooses a model
SAP uses more than 70 AI models from providers including OpenAI, Anthropic and Google. Asam said the company had removed some of the most expensive AI models from products including its internal consulting assistant and replaced them with lower-cost alternatives where performance remained strong. The company increasingly chooses the cheapest model capable of performing a task to the required standard.
That “capable to the required standard” qualifier is the whole strategy. A model that is cheaper but wrong, slow or non-compliant is not a saving. A model that is cheaper and good enough for a routine task — drafting, summarising, classifying — frees budget for the tasks that genuinely need a frontier model. The discipline is in knowing which tasks fall into which category, and in testing performance before swapping, rather than swapping on price alone.
SAP’s pricing model makes that discipline profitable in a way it would not be for a compute-based provider. Unlike many AI providers, SAP charges customers for completed tasks rather than for the underlying computing resources used to generate responses, so lower operating costs can help improve profitability. If the customer pays per task and SAP’s cost per task falls, the margin improves — provided quality does not fall with it.
Why spending still rises
The most misunderstood part of the announcement is that overall AI spending is still expected to rise as adoption grows. Asam said the goal is healthy margins as scale grows, not an absolute reduction in AI expenditure. In other words, SAP expects to spend more on AI in total, because more customers will use more AI features, while spending less per task than it would have if it had kept the most expensive models everywhere.
That distinction is important for anyone tracking AI economics. A company can be both cost-disciplined and spend-growing, if volume growth outpaces per-unit savings. The €2 billion efficiency figure should therefore be read as a margin story — better economics per unit of AI work — not as a claim that SAP is cutting its AI budget.
Model origin is a second theme in the announcement. Customers are paying closer attention to where AI models come from. Asam said government agencies and companies in regulated industries increasingly ask specific requirements and exclude models based on origin, while European customers increasingly demand local solutions. That demand sits alongside the cost question: the cheapest capable model may not be usable if its origin disqualifies it for a regulated customer, and a local model may be chosen even at a premium for sovereignty reasons.
How this differs from SAP’s robotics story
This article is about model economics, and it should not be confused with the separate Intel-SAP robotics story in this batch. The Intel-SAP proof of concept, demonstrated at Intel’s Chengdu site, is about warehouse robots receiving jobs from SAP Extended Warehouse Management and reporting results back — a physical-AI integration question. This Reuters story is about which language models SAP runs inside its software, how much they cost per task, and what that means for margins. One is about robots moving materials; the other is about models answering questions. Both involve SAP, but they answer different questions.
Readers interested in the broader platform context can read SAP Physical AI and the autonomous enterprise. The inference-cost angle is also illustrated by FuriosaAI’s inference chips, where cheaper, faster inference is itself the product.
Limitations and unanswered questions
The €2 billion figure is an expectation over two years, not a reported saving, and no breakdown was provided of how much comes from model swaps versus other efficiency measures. No specific models were named as removed or substituted, and no performance comparisons were published, so readers cannot independently judge whether quality was maintained. The model-origin discussion was described in general terms, without naming which models are excluded for which customers. Until SAP reports actual efficiency gains and the models involved, this story is best read as a stated strategy with a quantified ambition, not as a verified result.
