The announcement came straight from Brussels: the European Commission has closed its investigation into alleged anticompetitive practices by SAP, after the German company committed to abolishing reinstatement fees and reducing back-maintenance costs for customers returning to its support ecosystem. The decision concerns the aftermarket for on-premise business software services, but its ripples extend far beyond the ERP universe, touching a raw nerve for anyone designing a long-term IT infrastructure—including the infrastructure used to run Large Language Models locally.

The case stems from the thorny issue of migration from SAP ECC to S/4HANA. With mainstream support for ECC ending in December 2027 (extendable to 2030 at an extra two percentage points annually), over 35,000 global customers face a crossroads. According to Gartner, only 39 percent had purchased or subscribed to licenses for the transition by Q4 2024. Some, like UK retailer Kingfisher, opted to turn to independent providers such as Rimini Street to keep ECC 6.0 running, deeming the migration value insufficient.

It was precisely this choice—dropping the vendor’s support and possibly reconsidering later—that was made prohibitive by the reinstatement fees and back-maintenance SAP imposed. The Commission deemed such clauses could restrict competition, and SAP is now eliminating them, also committing to clarify conditions for alternating maintenance providers and different support levels. The agreement will stay in force globally for ten years.

For those assembling GPU racks today to perform inference of language models, the lesson is immediate. Hardware and software support contracts for AI—from NVIDIA AI Enterprise licenses to service plans for MLOps platforms—may conceal similar mechanisms: reactivation fees that make it uneconomical to return to an official vendor after trying alternative solutions. The SAP case sets a precedent, as it establishes that regulators are ready to intervene when exit and re-entry costs harm competition in the on-premise aftermarket.

Commissioner Teresa Ribera explicitly warned that the commitment “should serve as a warning against practices with similar effects in the cloud markets.” This is crucial: while the decision formally concerns only on-premise maintenance, the regulator’s direct link to cloud dynamics signals it will not hesitate to examine support contracts for hybrid or private cloud AI services through the same lens. For teams evaluating self-hosting of LLMs, it becomes essential to negotiate upfront clear exit clauses and the absence of re-entry penalties, because future regulation may impose them anyway.

SAP, for its part, stressed that the agreement strengthens predictability for customers and that its maintenance practices have always aligned with industry standards. The company added that the decision does not affect its cloud offering, but the increased transparency will support customers on their journey toward the AI-enabled “autonomous enterprise.” This distancing, however, does not change the fact: those investing in on-premise infrastructure now have an additional argument to demand flexibility from vendors—whether it’s servers for model quantization or appliances dedicated to low-latency inference.

The message from Brussels is, at its core, simple: software maintenance cannot become a positional rent. For the on-premise AI ecosystem, where switching costs between frameworks and accelerators can already be high, the SAP ruling provides a guiding principle—the freedom to repair and switch sides without retroactive penalties—that could redefine contracts for the next decade.