BNY: an on-premises DGX SuperPOD with DGX H100 for banking AI

BNY looks after close to 50 trillion dollars of client assets, according to the 2024 announcement. A firmwide review in the year before turned up more than 600 possible uses for AI, and its internal AI Hub already ran over 20 AI solutions in production for prediction, automation and anomaly detection.

Hundreds of its data scientists, solution architects and risk and compliance staff had used NVIDIA DGX machines for several years. Moving from a long list of ideas to many applications in production called for more on-premises compute than the bank had, and its CIO linked the new system to launching AI features that help manage, move and protect client assets.¹²

In March 2024 BNY announced a DGX SuperPOD made up of dozens of DGX H100 systems joined by NVIDIA InfiniBand and laid out according to NVIDIA's SuperPOD reference architecture. NVIDIA Professional Services worked with the bank on installation, which the announcement says finished faster than usual.

The announcement lists end-of-day cash balances, deposit forecasting, payment automation and predictive analytics on trades as workloads for the system. At the time, dozens of projects were in development with NeMo, Triton Inference Server (now called Dynamo-Triton) and Base Command, and the bank intended to adopt NVIDIA AI Enterprise for building and running AI applications and for managing the cluster. Those software steps were described as plans and work in progress, not finished deployments.

An NVIDIA post updated in June 2025 adds fraud detection to the purposes BNY plans for the system. BNY's current AI page still describes the bank as the first major bank with an NVIDIA-powered AI supercomputer and reports more than 125 AI solutions in production alongside Eliza, its enterprise AI platform; it does not say which of them run on the SuperPOD.¹²³⁴

What others can take from it: BNY treated owned compute as the next step after years of smaller DGX use and a firmwide inventory of AI opportunities, so the system arrived with a queue of named workloads and an existing governance process. Limits: no throughput, cost or business result has been published for the SuperPOD, and the bank's later count of more than 125 production solutions cannot be attributed to this system from the sources. The four named workloads are mostly forecasting and tabular analytics, which many banks run well on CPU clusters or rented cloud GPUs; an on-premises SuperPOD fits an institution with a large, steady pipeline of model training and strict data residency needs.