SAP's CFO on Token Sprawl: 58 Models, and Sometimes No Model at All

SAP CFO Dominik Asam argues the smartest AI cost move is often to route work to a cheaper model, or skip the model entirely and just run the math. In a CNBC earnings-alert segment filmed at SAP headquarters in Walldorf, Germany, Asam frames rising token spend as a real constraint that enterprises are only now waking up to. His answer is architectural: SAP has wired 58 large language models into its platform so it can pick the right one for each task, and fall back to plain deterministic code when no model is needed. The clip pairs his remarks with an on-screen chyron reading "SAP CFO: recognition that tokens for AI are not for free."

Tokens Are Not Free, and Enterprises Now Feel It

Asam opens with the shift in how companies think about AI cost. "There is now a recognition that these tokens for AI are not for free," he says, and "every enterprise we talk to is grappling with the fact that the token spend is going through the roof." The framing matters because it comes from a CFO, not an engineer. The concern is the budget line, not the benchmark score. His point is that throwing tokens "in a probabilistic way at any problem there is might not be the most efficient way." That is a direct challenge to the reflex of pointing a frontier model at everything.

Lock-In and Rising Prices Push Toward Optionality

Beyond raw spend, Asam names a second worry: dependence on a single vendor. He describes "a fear of lock-in, that certain models are starting to increase prices." This is the commercial logic behind SAP's approach. If you build your product on one model provider, you inherit that provider's pricing power. Asam presents SAP's multi-model design as the hedge. By keeping many models available, SAP can "play that vibrant competition, both on performance and on cost," and move workloads if any one supplier raises prices.

58 Models Bolted Into One Platform

The concrete number is the anchor of the clip. "I just checked this morning, we have now 58 large language models bolted into our AI platform," Asam says. The design goal is task routing: SAP can "channel the tasks to the most appropriate model" rather than defaulting to the biggest one. This matches SAP's public description of its Generative AI Hub and orchestration service, which give customers access to frontier models from providers including OpenAI, Anthropic, Mistral, and Google Gemini, and let them control which model handles a given deployment for cost and compliance reasons.

The Cheapest Reliable Tool Wins

Asam's cost discipline has a clear rule. "Sometimes you don't need the most expensive model," he says. In related reporting around the same SAP results, he put it as choosing "the cheapest reliable tool that can deliver the required outcome safely," whether that is simple software, an open-source model, or an expensive frontier model. The most advanced model is not automatically the right one. For a finance leader, the outcome and its reliability set the requirement, and the model is chosen to meet that requirement at the lowest defensible cost.

When the Right Model Is No Model

The sharpest line in the clip pushes past model selection entirely. "Even more importantly, sometimes you don't need any model at all," Asam says. "You just, in a deterministic, algorithmic, easy-to-audit way, crunch the numbers." For many enterprise tasks, especially in finance, a coded rule is cheaper, faster, and fully auditable in ways a probabilistic model is not. SAP's own efficiency work points the same direction: its SAP-RPT-1 model for tabular business data is described as using dramatically fewer compute resources than a general LLM. The message is that the token bill is a design choice, and the cheapest correct path sometimes has no model in it.

Key Takeaways

Enterprises are hitting real token-cost ceilings, and Asam says token spend is "going through the roof."
SAP's stated hedge against model lock-in and rising prices is keeping many models available and routing tasks to the best fit.
SAP has 58 large language models integrated into its AI platform as of the recording.
The operating rule is the cheapest reliable tool for the required outcome, not the most advanced model by default.
For deterministic, auditable work like crunching numbers, plain algorithmic code can beat any LLM on cost and trust.
Treating AI cost as an architecture decision, not a fixed input, is the CFO framing behind SAP's design.