ARMONK, N.Y.ㅡArtificial intelligence in large companies is moving into a less glamorous phase. The question is no longer simply whether a model works, but whether businesses can connect data, compute, governance and compliance tightly enough to use AI inside real operations rather than isolated trials.
That was the backdrop to IBM’s announcement on March 17 at NVIDIA’s GTC 2026 conference that it is expanding its collaboration with NVIDIA across several parts of the enterprise AI stack, including structured data analytics, document processing, on-premises and regulated infrastructure, cloud services and consulting. The aim, as described by IBM, is to help companies move AI systems from pilot projects into production environments.
One part of the announcement focused on structured data analytics. IBM said its watsonx.data SQL engine, Presto, is being accelerated with NVIDIA cuDF through an open-source integration designed to improve performance on large datasets. To demonstrate the setup, IBM said it applied the system to Nestlé’s order-to-cash data mart, which tracks orders, fulfilment, delivery and invoicing across 186 countries and processes terabytes of data across 44 tables.
IBM said Nestlé had previously refreshed that dataset on CPUs in about 15 minutes and only a few times a day. Using NVIDIA software and GPUs, IBM said query runtime fell to about three minutes, which it described as an 83 per cent cost saving and a 30-fold price-performance improvement in that proof of concept. Those figures were presented by IBM as case-specific results rather than a broad benchmark for all enterprise deployments.
The companies also used the announcement to address unstructured data, which remains a practical obstacle for many businesses because information is often scattered across documents, content systems and internal knowledge bases in formats that are difficult to standardise. IBM said it is combining Docling, its document-conversion tool, with NVIDIA Nemotron open models to turn multi-modal documents into AI-ready formats with source traceability.
On infrastructure, IBM said NVIDIA has selected IBM Storage Scale System 6000 to provide 10 petabytes of high-performance storage for GPU-native analytics workloads, and that the system has been certified and validated on NVIDIA DGX platforms. IBM also said the two groups are exploring how IBM Sovereign Core could be integrated with NVIDIA infrastructure and Nemotron models for customers that need AI workloads to stay within specific regional or regulatory boundaries.
The partnership is also being extended into cloud and consulting. IBM said it plans to offer NVIDIA Blackwell Ultra GPUs on IBM Cloud in the early second quarter of 2026 for training, inference and reasoning workloads, and to integrate the technology with Red Hat AI Factory with NVIDIA and virtual private cloud offerings that include compliance and data-residency controls. IBM Consulting also said it plans to use Red Hat AI Factory with NVIDIA through IBM Consulting Advantage as part of its enterprise AI services. IBM added that statements about future plans remain subject to change.
In practical terms, the announcement reflects a broader shift in enterprise AI. The harder problem is no longer access to models alone; it is the slower work of rebuilding data pipelines, storage, compliance controls and operating processes around them. IBM and NVIDIA are betting that demand will increasingly centre on that layer of the market.
