On Monday May 4, SAP did three things in a single press cycle. It announced the acquisition of Dremio, an open data lakehouse built on Iceberg. It announced a €1 billion investment over four years to acquire Prior Labs, an 18-month-old German AI lab whose TabPFN tabular foundation model has been downloaded over three million times. And it confirmed that SAP “prohibits” unauthorized AI agents from accessing its products through its API — only SAP-endorsed agent architectures, namely Joule Agents and NVIDIA’s NemoClaw, are permitted on the platform.
Three moves, one direction. Buy the AI capability SAP does not have. Close the agent perimeter SAP does not control. Use the cloud-revenue strength reported six weeks earlier — Q1 2026 cloud revenue +27% at constant currencies, current cloud backlog €21.9 billion — to pay for both moves while the migration cliff is still funding the income statement.
The first bet is on tables. CTO Philipp Herzig framed it directly: “Early on, SAP recognized that the greatest untapped opportunity in enterprise AI wasn’t large language models; it was AI built for the structured data that runs the world’s businesses.” Prior Labs builds tabular foundation models — models trained to make predictions from rows and columns rather than tokens of text. TabPFN-2.6 ranks first on the TabArena benchmark; the model’s row support expanded from 10,000 to 10 million in less than a year. SAP is paying the founders more than $500 million in cash up front, with €1 billion of further investment over four years, to capture an AI capability category SAP did not previously have in-house and could not build at speed. Dremio brings the data federation underneath — combining SAP and non-SAP data into an Iceberg-native lakehouse that feeds the SAP Knowledge Graph. The combined acquisitions are a single architectural argument: enterprise AI value will accrue at the layer that owns the structured data, not at the layer that generates the conversational interface, and SAP intends to own both the data and the model trained on it.
This is the same argument Palantir’s Q1 numbers gave empirical shape two days earlier. The mechanism is consistent: when the model layer commoditises — token costs down roughly a thousandfold since 2023, by Palantir’s CTO’s accounting on the same week — the value released by that commoditisation flows up the stack to whichever layer holds proprietary context. SAP’s structured-data inventory is among the largest in enterprise software: financial transactions, supply-chain ledgers, procurement records. The bet is that AI built specifically for that data shape will be more valuable than general-purpose models trained on text.
The second bet is on gates. SAP’s API now permits only “SAP-endorsed architectures” — Joule and NemoClaw — to interact with its products. NemoClaw is NVIDIA’s enterprise agent deployment method, announced in March as a partnership in which SAP’s Joule began supporting NVIDIA’s Agent Toolkit. The endorsement framing is doing the work in that sentence: every other major enterprise software vendor is moving in the opposite direction. Salesforce shipped Slack as an MCP client in early April routing to any enterprise agent that complies with the protocol. Atlassian’s Rovo explicitly markets compatibility with Google, Salesforce, and foundation-model agents accessing the Atlassian platform through the Teamwork Graph. SAP is choosing the inverse: a closed-perimeter agent runtime, with one approved partner (NVIDIA) and one approved internal product (Joule).
The two bets reinforce each other. If tabular foundation models prove decisive for enterprise AI, customers need access to SAP’s structured data to deploy them — and the gate ensures that access runs through SAP-approved agents. If the gate holds, third-party agents cannot disintermediate SAP’s tools by reading the system of record directly. The combined strategy is a vertical lock from data layer to agent runtime, paid for in advance by cloud-migration revenue while it still flows.
This complicates the hypothesis in a specific direction. SAP is the deepest-moat vendor in the cohort the hypothesis tracks. Six weeks ago its Q1 print delivered the strongest falsification evidence to date — cloud +27%, AI attached to two-thirds of Q4 cloud orders — while the CEO openly described subscription pricing as “foolish.” This week’s M&A double is the operational answer to that tension: SAP cannot replicate Atlassian’s playbook of self-funding a transition through internal restructuring, because SAP’s Joule attach rate in production is still around 3% and the company does not have the AI talent in-house to ship a credit SKU at the speed Atlassian shipped Rovo. So SAP buys outside what Atlassian built inside. The transition mechanism is the same — turning cloud-cycle revenue into AI capability — but the cost structure runs through M&A premiums rather than severance charges.
Three falsification risks remain visible on the deals. First, closed-perimeter agent strategies have a poor record in enterprise software. The history of Lotus Notes, BlackBerry, and pre-iPhone mobile platforms suggests customers will route around restrictions when an open alternative produces comparable results, and the hyperscalers’ agent platforms are converging on open protocols. Second, tabular foundation models as a production category are 18 months old. TabPFN’s benchmarks are real, but no enterprise has yet run a TabPFN-derived production system through a full deployment cycle at SAP scale. The architectural argument may be correct; the time to value may be longer than four years. Third, the €1 billion commitment is concentrated in a single team of three founders — Frank Hutter, Noah Hollmann, Sauraj Gambhir — whose retention is now a critical dependency. Hutter, the CEO, will run Prior Labs as an independent unit, and the M&A literature on AI labs absorbed into incumbent enterprises is not encouraging for retention or research velocity.
What to watch. HubSpot reports Q1 after the bell today; outcome-based pricing on Customer Agent ($0.50 per resolved conversation) hits its first earnings test. Snowflake reports later this month — a comparable question for the data-platform layer that Dremio now competes inside. Salesforce’s response on the agent-perimeter question matters most: if Salesforce or Atlassian closes their platforms in the next two quarters, the closed strategy is gaining ground; if they double down on open MCP and external-agent compatibility, SAP is the outlier and the bet is harder. And the SAP Q2 print, due in July, will show whether Joule production attach moves off 3% as Prior Labs and Dremio integrate, or whether customer adoption of the closed stack lags the architectural ambition.
The hypothesis tracker drifts toward complicate. Verifying evidence: SAP’s defensive M&A and platform-gating are both responses to the AI-driven commoditisation pressure the hypothesis predicts, undertaken by a vendor that publicly disowned its own pricing model six weeks earlier. Falsifying evidence: the architectural bet — that value lives at the structured-data layer — aligns with Palantir’s Q1 print and would, if it holds, allow the deepest-moat vendors to capture rather than lose value through the transition. The mechanism the hypothesis tracks (AI commoditises software) is correct. The shape of vendor responses is more varied than the original framing assumed: Atlassian self-funds a pivot, ServiceNow grows AI revenue from labour budgets, SAP buys outside and gates the perimeter. Different cost structures, same destination, with the financial weight of the bet now visible on SAP’s books.
Sources: SAP acquires Dremio, Prior Labs — Constellation Research; SAP bets $1.16B on 18-month-old German AI lab — TechCrunch; SAP agrees to acquire German AI startup Prior Labs — Sifted.