SAP Bets on Tabular AI, From the Supply Chain to the Space Station
The pitch is that structured data needs a different kind of model from an LLM. The orbital deployment is planned for early 2027, and the governance layer is the part shipping now.
SAP used its TechEd conference in Berlin to announce a broad expansion of its enterprise AI platform, and the most interesting part is not the part with the rocket. The company added capabilities across Joule Studio, Joule Work and the SAP Business AI Platform: no-code and pro-code agent building with a choice of more than 70 models, an ontology manager in SAP Business Data Cloud, a knowledge graph extended to non-SAP sources, an enterprise AI gateway for governing AI traffic, and an AI Governance Assistant aimed at the EU AI Act. It also confirmed a commitment that matters commercially: customer data stays the customer’s and is not shared to train or improve large language models.
The space station is the marketing, and it is good marketing. SAP says its tabular foundation model, SAP-RPT, will run on HPE’s Spaceborne Computer-2 aboard the International Space Station to triage anomalies in live telemetry, predict failures before they escalate, reconstruct missing sensor readings and generate mission-health reports without sending data back to Earth. SAP plans the deployment for early 2027, which the release states and the keynote’s drama obscures.
What SAP actually announced
The substance sits in four areas, and the general availability dates matter as much as the features. Joule Studio, the AI-first development environment announced at Sapphire in May, is now reaching general availability in a phased approach, with intent-based development that configures Model Context Protocol connections and authorisations from a description, agent extensions through n8n workflows, and deployment on an SAP-managed runtime that inherits the platform’s security and observability. SAP plans a tighter integration with SAP Business Data Cloud for the first quarter of 2027, letting agents reach trusted data products within IT’s existing controls.
The Enterprise AI Gateway is the governance centrepiece: one control plane for AI traffic across MCP servers, APIs, agent-to-agent calls and LLMs, consolidating authentication, policy enforcement, routing, monitoring, usage tracking and auditing. It builds on SAP Integration Suite, which customers already run, and it is planned for general availability in the first quarter of 2027. Alongside it, an AI Governance Assistant aims to help customers classify and document AI systems for the EU AI Act.
The ecosystem announcements deserve a line because they show where SAP thinks agents will touch the physical world. A collaboration with Intel closes the loop between enterprise workflows and autonomous robots, bridging information technology and operational technology through what SAP calls physical AI. Solution extensions with Docusign, Parloa and Redwood push autonomous capability into signature, voice and scheduling workflows. None of these carries a general availability date in the release, which places them among partnership announcements rather than products.
The tabular argument
The technical thesis worth taking seriously is that structured data needs a different kind of model from a large language model, which is the case for tabular AI. An LLM predicts the next token; a tabular foundation model predicts the next value in a row, which is what most business decisions actually require. SAP’s RPT family now has three members. SAP-RPT-TS handles time-series forecasting and is generally available, aimed at demand planning inside SAP IBP without custom machine learning. SAP-RPT-2, due in the first quarter of 2027, moves beyond classification and regression into anomaly and outlier detection, embeddings for analysis, and native multi-table processing. SAP-RPT-Decisions, due in the fourth quarter of 2026, targets approval routing, order decisions and supply chain exceptions.
The claim attached to the tabular AI family is striking: SAP’s own newsroom says SAP-RPT is about 50,000 times more energy-efficient than state-of-the-art LLMs, and that it requires no pre-training as conditions change. If either half of that holds at scale, it changes the economics of applying tabular AI to the operational data that dominates enterprise computing. It is also a vendor claim about a vendor’s model, and a reader should want the benchmark.
The space station, and what it proves
The ISS deployment deserves separating from the rest, because it is simultaneously the most impressive and the least verified item in the announcement. The constraints SAP cites are real and relevant: intermittent connectivity, severe energy limits, radiation interference and no tolerance for failure are the conditions of remote industrial operations on Earth too, from disaster response to distributed manufacturing to agriculture in places with poor connectivity. Running inference on the hardware rather than shipping telemetry to a data centre is the right architecture for all of them.
What the announcement does not yet establish is that the model has run there. SAP’s own newsroom describes the deployment as planned, by early 2027, and the release uses the future tense. The keynote showed the model reading live server telemetry and reconstructing missing readings, which demonstrates the capability on a testbed rather than in orbit. That is a meaningful distinction for a claim framed as validation, and it is one a reader can hold alongside the ambition.
The governance play and the EU AI Act
The governance layer is where SAP has the strongest argument and the least glamorous story. Its analyst material puts the case in numbers a buyer can repeat: 7.3 million data fields and more than 400 million semantic triples in the knowledge graph, 120 mission-critical processes covered, and more than 55,000 agents already managed in its AI Agent Hub across customers. The pitch is that context is what makes an agent accurate and governance is what makes it deployable, and that both come from fifty years of process encoding rather than from a better model.
The EU AI Act gives that pitch a deadline. Enterprises deploying agents into regulated processes need classification, documentation and audit trails, and an assistant that drafts the paperwork is a reasonable product for a compliance officer who has to sign it. Whether SAP’s controls satisfy a regulator is a question only a regulator answers, but the demand is real and the timing is right.
The land grab
One line in the release is easy to miss and commercially important. Joule Studio will run a time-bound promotion under which deploying and running custom agents on SAP’s managed runtime stays free for select customers through December 2026, with design-time tooling free under fair-use limits. Free compute for agent workloads is a land-grab tactic, and a sensible one: the switching cost in agent platforms comes from where agents run and what they can reach, not from the price of a builder. Expect the pricing conversation in January.
The commercial pressure behind all of it is the installed base. SAP’s keynote set out end-of-life dates for several generations of its own technology: ECC in 2027, PI/PO in 2027, Neo and NetWeaver Java in 2028, BW in 2034. Every one of those is a customer who has to move, and every one is an opportunity to move them onto a platform where tabular AI and agents run. That is the context in which a free runtime through December 2026 makes sense: the migration window and the agent window are the same window.
What the release does not say
Four things are absent. There is no customer outcome: no named enterprise reporting what an agent did, how much it saved, or what it got wrong. There is no benchmark behind the tabular AI energy-efficiency claim or the assertion that Joule Studio is more than five times faster than do-it-yourself work, which SAP attributes to Constellation Research without publishing the method. Moreover, here is no pricing beyond the promotion. And there is no failure data, which for agents that take actions in systems of record is the number an operations leader wants first.
What to watch
Three markers will show whether this is a platform shift or a conference. The first is the first-quarter 2027 cluster, when the Enterprise AI Gateway, the Business Data Cloud integration and SAP-RPT-2 all land; a slip there would be informative. The second is the ISS deployment itself, because a model running in orbit is a different claim from a model demonstrated on a testbed. The third is the agent count in the AI Agent Hub, which SAP already reports at more than 55,000: if that number compounds through 2027 while customer outcome stories appear alongside it, the autonomous enterprise argument will have evidence behind it rather than architecture. That is a higher bar than shipping a gateway, and a fairer one.
What it signals
The announcement marks a shift in how the largest enterprise software vendor argues for itself. For two years the question was whether SAP could build a credible AI platform at all; the answer now is that it has built one, and the argument has moved to whether its context advantage is worth a premium over assembling the same pieces from a hyperscaler and a model provider. SAP’s answer, in effect, is that the moat is the ontology and the audit trail, not the model, which is why it offers 70 of them and gives the runtime away for a quarter.
It is also a claim that will be tested by the customers SAP has spent five decades acquiring, at the moment many of them face a forced migration. That is a defensible position and an expensive one to hold, and the next four quarters will show which. The real test is not whether SAP can ship the gateway, but whether a customer can describe what an agent saved without reading from a vendor slide.