Capgemini Unveils InsightGrid: A GPU-Powered AI Data Revolution
Capgemini has developed InsightGrid, a GPU-native data lakehouse designed for modern AI workloads. The platform, built with AWS and NVIDIA, will be showcased at NVIDIA GTC 2026 on March 16th at 2:00 pm. It aims to transform how enterprises handle structured and unstructured data in the Agentic AI era.
InsightGrid was created to address the growing demands of AI systems that rely on unstructured data, dynamic token flows, and real-time processing. Unlike traditional setups, it treats compute, storage, and networking as equally critical components, optimised for GPU performance. The platform integrates technologies like NVIDIA RAPIDS, Polars, and Apache Iceberg, running on Amazon S3 and FSx for Lustre to deliver high-speed analytics.
The system is built around four specialised grids. SentinelGrid ensures data trust at ingestion, while ConcordGrid merges structured and embedded data. SignalGrid handles ad hoc analytics, and PulseGrid monitors KPIs across user cohorts. A live demo of PulseGrid took place on December 3, 2025, at AWS re:Invent in Las Vegas. During the presentation, it processed 1.2 million tokens per second on eight NVIDIA H100 GPUs, supporting real-time inference for large language models under loads of up to 500 concurrent users.
Performance tests show InsightGrid delivers five to seven times faster processing than CPU-based systems. It also cuts infrastructure costs by 60 to 80 percent. The platform unifies records, events, tokens, vectors, and media within a single GPU-native architecture, making it a core component of Capgemini's AI Factory framework for enterprise transformation.
InsightGrid will be officially presented at NVIDIA GTC 2026, following its successful demo at AWS re:Invent. The platform's ability to handle high-speed, real-time AI workloads while reducing costs positions it as a key solution for businesses adopting advanced AI. Capgemini plans to integrate it into broader enterprise transformation programmes moving forward.
