This article is the last of my articles on the 2026 Nvidia GTC. The show will feature booths from Lightbit Labs, ScaleFlux, FarmGPU inference environment, Western Digital, and Seagate, as well as announcements from Vast, Everpure, Penguin Solutions, Hammerspace, and HPE.
Lightbits Labs with ScaleFlux NVME SSD and FarmGPU Managed Inference Environment showed how their LightInferra software reliably persists, reuses, and streams KV cache data across inference sessions more efficiently, mitigating GPU stalls caused by repeated context recomputations and opening the door to more predictable and scalable performance and infrastructure efficiency.
The company accomplishes this by prefetching only critical data and delivering it to the GPU via high-speed remote direct memory access (RDMA) before it is needed to prevent GPU outages. This results in faster time to first token and more stable throughput under real-world workloads and high GPU utilization. According to Lightbits, this provides end-to-end security, including encryption of KV cache blocks, tenant isolation, and integration with key management systems (KMS) and trusted platform modules (TPMs) for shared inference environments.
I visited the Western Digital and Seagate booths at GTC. Seagate had a demonstration solution for context-driven AI using JBOF, an array of just a bunch of drives, which provides scalable, high-capacity data storage for long-term context affordability, combined with JBOF, an array of just flash that keeps immediate context data close to the compute.
I snapped a photo at the WD booth showing the company’s Ultra-SMR-based data center drive products in capacities up to 40 TB (see below). WD also showed flash arrays, presumably for contextual storage applications.
GTC Western Digital Data Center HDD
tom coughlin
At GTC, we met John Mao, Vast Data’s Vice President of Global Business Development. The company announced the VAST Foundation stack, an open source library that enhances and extends Nvidia AI blueprints into production-ready templates that run natively on the VAST AI operating system.
Everpure (formerly Pure Data) announced that its Evergreen//One program for AI, which enables pay-as-you-go consumption and on-demand scale, is now available on FlashBlade//EXA, introducing Everpure Data Stream.is said to eliminate the friction of manual data movement with a direct, automated pipeline from data ingestion to inference, and is expected to be released in late 2026.
Penguin Solutions made two memory-related announcements at GTC. The company announced the MemoryAI KV cache server with 3TB of DDR main memory and up to eight 1TB CXL add-in cards. We also announced enhancements to the OriginAI portfolio to address context size and concurrency and increase GPU memory to meet the low-latency demands of enterprise-scale AI inference.
Hammerspace announces launch of Ai data platformAIDP. AIDP helps you smoothly scale up your AI pilot program to production. This enables data access from anywhere across edge devices, data centers, clouds, and high-performance file and object storage.
According to HPE, Alletra Storage MP X10000 It became Nvidia’s first certified object-based storage platform. Additionally, HPE plans to support new features. NVIDIA STX Rack-scale reference architecture for developing new AI storage solutions leveraging NVIDIA Vera Rubin, BlueField-4, Spectrum-X networking, Connect-X NICs, and NVIDIA AI software.
Lightbit Labs, ScaleFlux, FarmGPU, Seagate, Western Digital, Vast, Everpure, Penguin Solutions, Hammerspace, and HPE announced storage products to accelerate AI workflows.
