No, SSDs are not becoming VRAM
© NVIDIA I have seen several articles and social media posts claiming that NVIDIA is preparing a new tech that will unlock local SSDs as a some kind of replacement for low VRAM GPUs. All this is based on a newly open-sourced technology which has nothing to do with gaming.
NVIDIA has announced plans to open-source its cuFile APIs and the storage software stack built around them. The announcement was made during Future of Memory and Storage 2026 and has already led to reports suggesting NVIDIA is developing technology that would allow SSDs to replace or expand GPU memory. That is not what NVIDIA announced.
CuFile is part of NVIDIA GPUDirect Storage, a technology that has existed for several years. It provides a direct path between local or remote storage and GPU memory. Instead of copying data from an SSD into system memory before sending it to the GPU, GPUDirect Storage allows compatible storage devices to transfer data directly into GPU memory using DMA.
The new part is NVIDIA’s decision to open the cuFile interface and create a vendor-neutral storage stack around it. A new xio-sig project lists four components: cuFile, cuFileConformance, libxFile and xioLinux . The project is designed to support implementations from multiple hardware vendors rather than keeping the interface tied exclusively to NVIDIA’s existing stack. The code itself has not yet been published, with the project stating that it will appear after the founding members complete integration and validation.
© Github Storage is getting closer to memory
This does not mean an NVMe SSD becomes additional VRAM. GPU workloads still need their working data in actual GPU memory for processing. cuFile instead makes moving data between storage and that memory faster and removes the CPU and system-memory copy from the normal transfer path.
There is a similarity to what Microsoft DirectStorage and NVIDIA RTX IO do for games, where assets can travel from NVMe storage toward the GPU with less CPU involvement. However, cuFile is a different technology designed around CUDA and GPUDirect Storage.
NVIDIA is nevertheless working toward a model where the distinction between storage and memory becomes less visible to some AI applications. Its SCADA , or Scaled Accelerated Data Access, architecture allows GPU threads to initiate storage requests and is intended for datasets that exceed local memory capacity. This can make NVMe storage function as another data tier close to the GPU, but its latency and bandwidth remain different from HBM or GDDR memory.
What most should take from this post is that NVIDIA has announced no technology that converts conventional SSD capacity into replacement VRAM. The featured image posted by NVIDIA is literally titled ‘ai-storage’ if it wasn’t clear enough.
Sources: NVIDIA
