Share with your CTO
DDN is making a direct bet that the next enterprise AI bottleneck isn’t GPUs, it’s the storage layer feeding them. The company’s Infinia 2.4 release, announced at the RAISE Summit in Paris, adds multi-tenancy, identity management, quota enforcement, and initial POSIX support to its AI data platform. The update targets production inference environments, specifically RAG pipelines, agentic AI, and large-scale model serving, where GPU idle time caused by slow data access translates directly into wasted capital. Named customers include NVIDIA, xAI, Salesforce, and Mistral.
What this means for your business
The organizations most exposed to this story are running shared GPU clusters across multiple teams or business units and discovering that governance didn’t ship with the hardware. Multi-tenancy in storage, meaning the ability to isolate workloads, enforce quotas, and manage identities across different organizational units on a single physical infrastructure, is the capability that turns a pile of expensive accelerators into something a CTO can actually hand to the CFO and defend. If your AI infrastructure is still single-tenant, you’re paying for separation you don’t need or absorbing risk you can’t audit.
DDN’s framing around “cost per token” is worth taking seriously, even though the company, which sells into the very infrastructure gap it’s describing, has an obvious interest in making that metric feel urgent. The specific claim it’s building toward is that storage latency, not model architecture or chip count, is where inference efficiency degrades in production. That’s plausible for RAG-heavy workloads, where retrieval from external data stores happens on every query, but it’s overstated for batch inference jobs where data is preloaded. The KV cache acceleration feature, which speeds up the memory layer that stores intermediate computation results during inference, matters most at high concurrency, which is exactly the multi-tenant scenario Infinia 2.4 is designed for. The two capabilities reinforce each other, which is a coherent product story even if the benchmark headlines are missing.
The POSIX addition is the signal to watch for legacy integration. Object storage like S3 is native to cloud-born AI workloads, but a meaningful portion of enterprise AI pipelines still depend on file system interfaces that expect POSIX semantics, the standard rules governing how applications read and write files. Limiting the initial release to Red Hat Enterprise Linux and Ubuntu with “defined throughput commitments” rather than full parity is honest scoping, but it means shops running mixed or non-standard environments should treat this as a roadmap item, not a shipping feature. The renewal decision this reshapes is whether your current parallel file system vendor, the one holding the POSIX contract today, is still the right answer when DDN is demonstrably closing that gap from the object storage direction.
Concept deep-dive: KV cache
During inference, a large language model repeatedly references earlier parts of a conversation or document to generate each new word. The KV cache stores those intermediate attention calculations in fast memory so the model doesn’t recompute them from scratch on every step, analogous to keeping a running tab open rather than recalculating a restaurant bill after each course. When that cache is slow or unavailable, GPUs stall waiting for data. Accelerating it at the storage layer, not just in GPU memory, is DDN’s core performance claim for high-concurrency inference environments.
Based on reporting from DDN Infinia 2.4 Adds Multi-Tenancy and POSIX Support for Production AI Factories, originally published 2026-07-08 09:31:00.
