On March 16, 2026, the opening day of GTC, IBM issued a press release from Armonk announcing an expanded collaboration with NVIDIA that pushes the relationship from “GPU supplier” to “data-layer integration.” The deal spans five areas: GPU-native data analytics, document processing, storage, sovereign deployment, and cloud compute. Its stated purpose is to attack the four bottlenecks IBM says keep enterprise AI stuck in pilot purgatory — fragmented data, infrastructure gaps, regulatory and compliance barriers, and missing expertise.
The two CEOs framed the angle themselves. IBM chief executive Arvind Krishna: “In the next wave of enterprise AI, the model layer will rely on the data, infrastructure, and orchestration layers.” NVIDIA chief executive Jensen Huang: “Data is the ground truth that gives AI context and meaning.” While most keynote-week announcements talked about models, this one bets on the layer underneath the models.
The Nestlé Benchmark: 15 Minutes Down to 3
IBM did not stop at architecture slides; it shipped a verifiable customer number. Nestlé’s order-to-cash data mart covers 186 countries and terabytes of data across 44 tables. Refreshing it took 15 minutes on CPU; on GPU-accelerated watsonx.data, that dropped to 3 minutes. IBM cites 83 percent cost savings and a 30x improvement in price-performance. Chris Wright, Nestlé’s chief information and digital officer, put it plainly: for a company serving billions of people, data underpins decision-making across global operations.
For data engineering teams, that number is more useful than any synthetic benchmark, because the bottleneck in enterprise AI is rarely the model — it is the pipeline that prepares data to feed the model.
GPUs Enter the Data Layer: Presto, cuDF, and Docling
The technically interesting part is a pair of open-source integrations. First, IBM is combining the Presto SQL engine inside watsonx.data with NVIDIA’s cuDF, so queries over large datasets execute directly on GPU; NVIDIA’s developer blog goes further, explaining that cuDF is being brought into the Velox execution engine to make Presto and Apache-ecosystem queries GPU-native. Second, IBM’s open-source document parsing project Docling will be paired with NVIDIA’s Nemotron open models to convert complex unstructured documents into AI-ready formats while preserving source-level traceability — a hard requirement for audit and compliance workloads.
The shared signal in both tracks: GPUs are no longer just for training and inference. They are starting to absorb data preparation itself.
Storage, Blackwell Ultra, and Sovereign Deployment
The infrastructure side has substance too. NVIDIA selected IBM’s Storage Scale System 6000 to supply 10 petabytes of high-performance storage for its GPU-native analytics engines, certified and validated on NVIDIA DGX platforms. IBM Cloud will bring Blackwell Ultra GPUs online in early Q2 2026 for large-scale training and inference, and the same stack reaches enterprises through Red Hat AI Factory with NVIDIA and IBM Consulting Advantage.
On sovereignty, the two companies are exploring an integration of IBM Sovereign Core with NVIDIA infrastructure and Nemotron models, so GPU-intensive workloads can run entirely within regional boundaries — extending the sovereign-software line IBM opened with Sovereign Core in January.
What It Means for Developers and Data Teams
Three observations. First, the boundary between SQL and AI is dissolving: when Presto queries run on GPU, feature engineering and data prep no longer shuttle between CPU clusters and GPU clusters. Second, almost the entire technical payload is open-source parts — Presto, Velox, cuDF, Docling — so teams that refuse to lock into a single commercial stack have a clear path to follow. Third, the 83 percent figure comes from one customer benchmark; your query shapes and migration costs will vary, so validate before you budget.
Place the announcement in its GTC-week context: compared with other enterprise AI stacks shown at the conference, such as Dell’s AI Factory, IBM chose to enter from the data layer rather than the server. The enterprise AI contest is shifting from “who supplies the machines” to “who wires the data into the machines” — and this collaboration is the most concrete statement of the latter so far.
Sources
- IBM and NVIDIA Announce Expanded Collaboration at GTC 2026 — IBM Newsroom
- IBM and NVIDIA Announce Expanded Partnership to Operationalize Enterprise AI — StorageReview
- Accelerating Large-Scale Data Analytics with GPU-Native Velox and NVIDIA cuDF — NVIDIA Developer Blog
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.