Summary: At GTC 2026 on March 16, IBM and NVIDIA expanded collaboration into GPU-native analytics, document processing, storage, sovereign deployment, and cloud compute. A Nestlé benchmark cited by IBM cut an order-to-cash data mart refresh from 15 CPU minutes to 3 GPU minutes, with 83% claimed cost savings. The companies also plan Presto with cuDF, Docling with Nemotron, Storage Scale storage, and early-Q2 Blackwell Ultra capacity on IBM Cloud.
GTC opened on March 16, 2026, and IBM used the day to put out a press release from Armonk: the partnership with NVIDIA is being expanded again, past the “GPU supplier” role and toward “data-layer integration.” Five areas make up the expanded scope — GPU-native data analytics, document processing, storage, sovereign deployment, and cloud compute — all aimed at the four blockers IBM says keep enterprise AI stuck at the pilot stage: fragmented data, infrastructure gaps, regulatory and compliance barriers, and missing expertise.
The framing came from the very top. IBM chief executive Arvind Krishna’s words: “In the next wave of enterprise AI, the model layer will rely on the data, infrastructure, and orchestration layers.” NVIDIA chief executive Jensen Huang’s definition: “Data is the ground truth that gives AI context and meaning.” For most of keynote week the talk centered on models; this partnership wagers on the layer underneath them.
A Verifiable Customer Number: Nestlé’s Refresh, 15 Minutes to 3
There was more than architecture on offer — IBM also brought a customer number that can be checked. Nestlé’s order-to-cash data mart spans 186 countries and terabytes of data organized into 44 tables. One refresh took 15 minutes on CPU; on GPU-accelerated watsonx.data the same job finishes in 3 minutes. The economics IBM cites: 83 percent cost savings and a 30x lift in price-performance. Chris Wright, Nestlé’s chief information and digital officer, explained what rests on it: at a company serving billions of people, data underpins every decision across global operations.
For data engineering teams, a field number like this beats any synthetic benchmark. What usually slows enterprise AI down is not the model — it is the pipeline that prepares the data the model consumes.
Two Open-Source Tracks: Presto Plus cuDF, Docling Plus Nemotron
The technically richest pieces are two open-source integrations. Track one pairs the Presto SQL engine in watsonx.data with NVIDIA cuDF, so queries over large datasets execute directly on GPU. NVIDIA’s developer blog sketches the next step: cuDF is being integrated into the Velox execution engine, pulling Presto and Apache-ecosystem queries along toward GPU-native execution. Track two pairs Docling, IBM’s open-source document parsing project, with NVIDIA’s Nemotron open models to turn complex documents into AI-ready formats while keeping source-level traceability intact — audit and compliance work cannot do without it.
Both tracks carry the same signal: the GPU’s remit is growing past training and inference, and data preparation is the next job it takes over.
Storage, Compute, and Sovereignty: From Storage Scale System to Blackwell Ultra
The infrastructure news is concrete too. NVIDIA chose the Storage Scale System 6000 to put 10 petabytes of high-performance storage behind its GPU-native analytics engines, and the system is certified and validated on NVIDIA DGX platforms. For compute, IBM Cloud is set to bring Blackwell Ultra GPUs online in early Q2 2026 for large-scale training and inference, and enterprises reach the same combination through Red Hat AI Factory with NVIDIA and IBM Consulting Advantage.
Sovereignty is still exploratory: the two companies are looking at integrating IBM Sovereign Core with NVIDIA infrastructure and Nemotron models so that GPU-intensive workloads stay entirely within their jurisdiction — a thread that extends the sovereign-software push IBM began in January.
Three Takeaways for Developers and Data Teams
One: the wall between SQL and AI is coming down — once Presto queries run on GPU, feature engineering and data prep stop moving back and forth between CPU clusters and GPU clusters. Two: the package is assembled almost entirely from open-source parts (Presto, Velox, cuDF, Docling), giving teams unwilling to be tied to one commercial stack a clear route to follow. Three: the 83 percent figure rests on one customer’s benchmark, so your query shapes and migration costs deserve their own evaluation.
Set the announcement inside GTC week and the positioning stands out: where stacks like Dell’s AI Factory lead with the machine, IBM entered through the data layer. The enterprise AI contest is turning from “who supplies the machines” into “who wires the data into the machines” — and this partnership is the clearest statement yet on that second front.