At A Glance
- NVIDIA Vera Rubin raises the performance bar for the entire data path, from training through inference. The NVIDIA Vera Rubin platform is used to build a POD-scale AI supercomputer designed for agentic AI, unifying compute, networking, and software to power next-generation AI factories.
- More powerful GPUs increase the cost of waiting on data, making storage performance a compute economics issue.
- Agentic AI moves context and KV cache directly into the inference execution path.
- DDN is extending its NVIDIA collaboration across EXAScaler for training and Infinia for inference.
- The opportunity is one data foundation that delivers performance, isolation and control as AI workloads scale.
NVIDIA Vera Rubin represents more than another generation of accelerated compute. Its architecture brings compute, networking, infrastructure offload, security, power, cooling, and software into a more tightly integrated system. That changes what the rest of the AI infrastructure has to deliver.
More powerful GPUs put more pressure on the data path. Training has to feed them and checkpoint them. Inference has to deliver context at the speed agents can consume it. Shared AI infrastructure has to do both while maintaining predictable performance and isolation.
DDN believes that the next generation of AI infrastructure cannot be built by treating each of those requirements as a separate storage problem. The data layer must move with the compute architecture and serve the AI lifecycle as one system.
Vera Rubin Raises the Performance Bar for Training
Vera NVIDIA Rubin raises the performance bar long before inference begins.
As GPU capacity increases, the cost of leaving that compute waiting on data increases with it. Training infrastructure has to sustain the data path into the GPUs and keep checkpointing from interrupting the run. Faster compute only increases the consequences when storage cannot keep pace.
DDN is already working with NVIDIA on reference architecture alignment for EXAScaler and Vera Rubin, extending a collaboration built around keeping accelerated compute fed at scale. As new models grow and training infrastructure evolves, the objective remains the same. Keep GPUs working and keep checkpoints from becoming the constraint.
Inference Moves Context into the Performance Path
Agentic AI workloads operate across long context windows, multi-turn reasoning, tool invocation, and trillion-parameter mixture-of-experts models. These workloads generate massive key-value cache footprints that cannot remain entirely in GPU HBM.
NVIDIA’s CMX Context Memory Storage architecture, based on the NVIDIA STX storage foundation, establishes context memory as a distinct, AI-native tier between GPU memory and traditional shared storage. Powered by NVIDIA BlueField-4, NVIDIA Spectrum-X Ethernet, NVIDIA DOCA Memos, and NVIDIA Dynamo, it is designed to provide high-bandwidth, low-latency access to reusable KV cache across the AI pod.
This changes the role of the data layer. Context is no longer simply stored and retrieved behind inference. It becomes part of the execution path. Its performance directly determines how much useful work the compute can produce and how efficiently it can do it.
For DDN, that validates a direction we believe extends well beyond KV cache. As AI infrastructure becomes more tightly integrated, data services must move closer to where the work happens.
Bringing the Data Path Closer to the GPU
AI changes the relationship between compute and data. Data moves through training, checkpointing, retrieval, inference and continuous improvement, with different performance demands at each stage.
Inference with agents intensifies that change. Agents retrieve information, invoke tools, create intermediate artifacts, and maintain state across multistep workflows. Long-context models also generate and reuse substantial KV cache data, especially when supporting reasoning, research or complex conversations.
The data layer becomes part of the serving architecture, not merely a repository behind it. And our approach is to treat KV cache as a first-class data path.
Infinia integrates directly with NVIDIA’s inference software ecosystem, including NIXL, Dynamo, vLLM, and SGLang. Through the Infinia NIXL plugin, KV cache moves by DMA between GPU memory and Infinia for cache offload, model artifact retrieval, and inference-state management. CPU memory can also exchange data with Infinia for embedding retrieval, checkpoint loading, tool calls, and preprocessing. Both paths use RDMA-based transport for high-throughput, low-latency movement.
The collaboration also reduces deployment friction. Beginning with NIXL 1.3, the Infinia plugin is included in NVIDIA’s official NIXL distribution and the prebuilt NVIDIA Dynamo inference container. Customers can enable Infinia through configuration rather than maintaining a separate integration or changing application code.
This is where integration becomes architecture. The data layer is not simply certified alongside the compute platform. It becomes part of the software path moving context to and from the GPUs.
The objective is to remove distance between data and compute, keeping context moving at the speed inference demands from the first deployment through production scale.
Collapsing the Inference Data Path
The next step is to bring those data services closer still.
Building on DDN’s work with NVIDIA BlueField-3, Infinia can extend that architecture to NVIDIA BlueField-4, NVIDIA Spectrum-X Ethernet, and NVIDIA CMX, running data services natively on the infrastructure that serves GPUs.
That creates the opportunity to collapse layers in the inference stack and give GPUs more direct access to KV cache through RDMA and NVIDIA GPUDirect, while integrating KV cache services into NVIDIA’s inference software stack.
The architecture allows storage services, data movement, security, and isolation to be offloaded from host CPUs. The result is a more direct GPU-to-data path, while the DPU handles infrastructure functions independently of tenant workloads.
For mixed environments running training, fine-tuning, and inference on shared infrastructure, this separation is essential. Multi-tenancy requires predictable performance and isolation, not policies that exist only on paper.
Infinia is designed for that reality, with native cluster, tenant, and subtenant isolation, configurable tuning, asynchronous transfers, parallel batch execution, and direct memory movement.
As AI factories consolidate more workloads, models and agents onto shared infrastructure, that distinction becomes critical. Performance and isolation have to be properties of the data path itself.
Sovereign AI Raises the Stakes
The same architectural shift matters for Sovereign AI. Data still has to move at the speed production AI demands, but it also has to remain within required jurisdictional, security, and operational boundaries.
That becomes harder as training, fine-tuning, inference and agentic workloads converge on shared infrastructure. More workloads should not require more infrastructure silos. Performance and control have to coexist on the same data foundation.
This is where architecture matters. Isolation cannot come at the expense of performance, and performance cannot require giving up control of where data lives or how workloads are separated.
The Architecture is Changing. The Data Strategy Has to Change With It.
NVIDIA Vera Rubin makes a broader shift in AI infrastructure visible. Compute, networking, infrastructure services and data are moving closer together because the workload increasingly demands it.
For training, that means keeping increasingly powerful GPUs fed and checkpoints moving. For inference, it means putting context and KV cache directly into the performance path. For shared and sovereign infrastructure, it means maintaining performance without giving up isolation and control.
These are not separate storage problems. They are different demands on the same data architecture.
Building What Comes After NVIDIA Vera Rubin
Vera Rubin is important not simply because it advances compute; it makes clear where AI infrastructure is going. Training, inference and agentic AI are placing different demands on the same increasingly valuable compute infrastructure. The data architecture has to serve them all.
DDN has spent decades engineering the data path around what happens when compute scales. Vera Rubin pushes that problem forward again. More powerful GPUs increase the cost of waiting. Agentic inference moves context into the execution path. Shared AI infrastructure makes isolation and control inseparable from performance.
We believe these challenges should not produce another generation of infrastructure silos. Training and inference need different data services, but they do not need separate data foundations.
That is the opportunity we see in the Vera Rubin architecture. Build on performance leadership for training. Bring data services closer to the GPU for inference. Preserve control as workloads and tenants converge. Then reuse that foundation as the next AI workload arrives.
The future of accelerated computing will not be determined by compute alone. It will be determined by how much useful work that compute can produce.