GTC: storage architecture + demand surge
What happened
GTC coverage highlighted NVIDIA unveiling a storage architecture for AI agent systems and framed a major GPU demand surge from startups and enterprise users reported. Analysts at the event are connecting storage, agent tooling and huge chip revenue projections into a single infrastructure demand narrative reported.
Why it matters
BlueField‑4 STX delivers up to 5x token throughput, up to 4x higher energy efficiency and 2x faster data ingestion compared with traditional storage, NVIDIA said. (nvidianews.nvidia.com) The first rack‑scale STX implementation includes the CMX context‑memory storage platform that expands GPU memory and claims up to 5x more tokens per second versus conventional storage for long‑context inference. (nvidianews.nvidia.com) NVIDIA named early adopters of STX that span cloud and emerging compute providers — CoreWeave, Crusoe, IREN, Lambda, Mistral AI, Nebius, Oracle Cloud Infrastructure and Vultr. (nvidianews.nvidia.com) NVIDIA also shipped an enterprise Agent Toolkit (NeMo Agent Toolkit) plus the OpenShell runtime and listed 17 major software partners — Adobe, Salesforce, SAP, Atlassian, CrowdStrike, Cisco, Red Hat and others — as initial adopters. (nvidianews.nvidia.com) The Vera Rubin platform stitches together seven co‑designed chips and systems — Vera CPUs, Rubin GPUs, ConnectX‑9 SuperNICs, BlueField‑4 DPUs, Spectrum switches and an integrated Groq 3 LPU — and NVIDIA says the platform is now in full production. (storagereview.com) Bloomberg reported NVIDIA projects at least $1 trillion in revenue from its Blackwell and Rubin chip families through the end of 2027, up from a prior $500 billion sales forecast for the end of 2026. (bloomberg.com) Hyperscalers and OEMs are lining up Rubin/STX deployments: Microsoft said Azure Foundry will be the first hyperscale cloud to power Vera Rubin NVL72 systems, and HPE announced Vera Rubin rack systems with availability slated for late 2026. (blogs.microsoft.com)
Key numbers
- BlueField‑4 STX delivers up to 5x token throughput, up to 4x higher energy efficiency and 2x faster data ingestion compared with traditional storage, NVIDIA said.
- (nvidianews.nvidia.com) The first rack‑scale STX implementation includes the CMX context‑memory storage platform that expands GPU memory and claims up to 5x more tokens per second versus conventional storage for long‑context inference.
- (nvidianews.nvidia.com) NVIDIA also shipped an enterprise Agent Toolkit (NeMo Agent Toolkit) plus the OpenShell runtime and listed 17 major software partners — Adobe, Salesforce, SAP, Atlassian, CrowdStrike, Cisco, Red Hat and others — as initial adopters.
- (storagereview.com) Bloomberg reported NVIDIA projects at least $1 trillion in revenue from its Blackwell and Rubin chip families through the end of 2027, up from a prior $500 billion sales forecast for the end of 2026.
What happens next
- (nvidianews.nvidia.com) The first rack‑scale STX implementation includes the CMX context‑memory storage platform that expands GPU memory and claims up to 5x more tokens per second versus conventional storage for long‑context inference.
- (bloomberg.com) Hyperscalers and OEMs are lining up Rubin/STX deployments: Microsoft said Azure Foundry will be the first hyperscale cloud to power Vera Rubin NVL72 systems, and HPE announced Vera Rubin rack systems with availability slated for late 2026.
Quick answers
What happened in GTC: storage architecture + demand surge?
GTC coverage highlighted NVIDIA unveiling a storage architecture for AI agent systems and framed a major GPU demand surge from startups and enterprise users reported. Analysts at the event are connecting storage, agent tooling and huge chip revenue projections into a single infrastructure demand narrative reported.
Why does GTC: storage architecture + demand surge matter?
BlueField‑4 STX delivers up to 5x token throughput, up to 4x higher energy efficiency and 2x faster data ingestion compared with traditional storage, NVIDIA said. (nvidianews.nvidia.com) The first rack‑scale STX implementation includes the CMX context‑memory storage platform that expands GPU memory and claims up to 5x more tokens per second versus conventional storage for long‑context inference. (nvidianews.nvidia.com) NVIDIA named early adopters of STX that span cloud and emerging compute providers — CoreWeave, Crusoe, IREN, Lambda, Mistral AI, Nebius, Oracle Cloud Infrastructure and Vultr. (nvidianews.nvidia.com) NVIDIA also shipped an enterprise Agent Toolkit (NeMo Agent Toolkit) plus the OpenShell runtime and listed 17 major software partners — Adobe, Salesforce, SAP, Atlassian, CrowdStrike, Cisco, Red Hat and others — as initial adopters. (nvidianews.nvidia.com) The Vera Rubin platform stitches together seven co‑designed chips and systems — Vera CPUs, Rubin GPUs, ConnectX‑9 SuperNICs, BlueField‑4 DPUs, Spectrum switches and an integrated Groq 3 LPU — and NVIDIA says the platform is now in full production. (storagereview.com) Bloomberg reported NVIDIA projects at least $1 trillion in revenue from its Blackwell and Rubin chip families through the end of 2027, up from a prior $500 billion sales forecast for the end of 2026. (bloomberg.com) Hyperscalers and OEMs are lining up Rubin/STX deployments: Microsoft said Azure Foundry will be the first hyperscale cloud to power Vera Rubin NVL72 systems, and HPE announced Vera Rubin rack systems with availability slated for late 2026. (blogs.microsoft.com)
