Skip to the stories
Microsoft 365 & SharePoint

DDN Moves EXAScaler 7 Lustre Onto Standard Servers, Starting With AI2200

  • Harold Fritts
  • 552 views

DDN's EXAScaler 7 will run as software on qualified servers, starting with the AI2200, with a limited beta planned for early 2027.

DDN Moves EXAScaler 7 Lustre Onto Standard Servers, Starting With AI2200

DDN is letting customers run EXAScaler as software on qualified industry-standard servers, and that changes who can deploy its Lustre-based parallel file system. Until now, EXAScaler has come only inside DDN's purpose-built appliances or through cloud partners. The new software-defined option arrives with EXAScaler 7 and starts on DDN's own AI2200 server, with qualified third-party appliances to follow. A limited beta is planned for early 2027.

DDN AI2200 1U server with a red mesh front bezel carrying the Infinia badge

Why Software-Only Lustre Matters Here

If you've run Lustre in production, you already know the operational weight sits in the hardware and the support contract as much as the software. Buying EXAScaler as an appliance means one vendor, one bill of materials and a known configuration. For a lot of AI and HPC teams, that's the point. The catch is that it also ties the file system to a specific hardware path, and that doesn't suit every data centre.

DDN describes EXAScaler as open Lustre with AI I/O optimization, automated deployment and enterprise support. The reason a parallel file system matters for AI and HPC is that it reads from many storage servers at once. That's how it copes with small files, metadata-heavy reads and the frequent checkpoints that training and simulation jobs produce. Putting the same software on hardware you choose, from a supplier list you can check against your own standards, is a different procurement conversation from taking an appliance as delivered.

DDN says it will keep building purpose-built systems. EXAScaler already sits inside appliances such as the AI400X3M that DDN launched at ISC 2026, and that model isn't going away. The software-defined path is an extra option alongside it, not a replacement.

What Runs Where, and What's Still Unknown

With EXAScaler 7, the same software runs on what DDN calls verified hardware from a growing set of suppliers, beginning with the AI2200. DDN is pitching the option at hyperscalers and cloud providers that already have storage infrastructure, and at AI platform teams that want to choose their own hardware. The company hasn't named the third-party hardware vendors. It also hasn't published minimum hardware requirements or said how the software will be licensed.

Those three gaps matter for planning. Without a licensing model, you can't model per-node or per-capacity cost against an appliance quote. Without minimum requirements, a hardware refresh is guesswork. And until the qualified-vendor list exists, a team can't tell whether its preferred server platform will be supported. If you're evaluating this, the practical step is to ask DDN for the hardware compatibility and licensing terms in writing before you put a server order against the option.

DDN EXAScaler 7 deployment options graphic: DDN purpose-built systems, cloud partners, and the new software-defined option on qualified servers starting with the AI2200

Omer Asad, Chief Product Officer at DDN, put the change in portfolio terms: “By making EXAScaler available as software, we're extending that flexibility across our portfolio.” DDN Infinia, which gained multi-tenancy and POSIX support in version 2.4, already has a software-defined format, so DDN's two data platforms will both offer that deployment model. EXAScaler also stays available through cloud partners, including Google Cloud Managed Lustre.

Tenant Controls, Recovery and Auditing in EXAScaler 7

Beyond the deployment model, EXAScaler 7 adds features aimed at shared environments. Multi-tenancy REST APIs will apply per-tenant quality-of-service limits and fair sharing, so several teams can use one system without their jobs competing as freely. That's the kind of control that usually matters most once a cluster stops being a single research group's project and becomes a shared service.

File Restore recovers accidentally deleted files, and an audit log supports investigations. On performance, DDN lists intelligent hybrid I/O for higher throughput and IOPS. DDN hasn't published performance figures for any of these features and says validated data will follow. Treat the throughput claim as unverified until those numbers appear. DDN is scheduling these additions for the first half of 2027, separate from the software-defined option's beta.

Shared Data for Simulation and AI Training

DDN's example for the software-defined option is a site that runs HPC simulation and AI training against the same data, without copying it between separate environments. In life sciences, researchers train AI surrogate models on molecular dynamics simulation outputs to predict protein behavior, which avoids rerunning every simulation. Teams in this position typically spend a lot of time and storage on copies and data movement between a simulation cluster and a training cluster. A single file system that both can read from removes that step, provided the hardware underneath can keep up.

DDN hasn't given a general availability date beyond the early 2027 beta, and it says all of its dates are targets that may change. Plan around a beta, not a shipping product.

AMPHIX HyperPOD Reports 98.8 Percent GPU Utilization on Infinia

DDN also reported results from the first DDN Enterprise AI HyperPOD installed at AMPHIX, the RAVEL and Strata Expanse deployment model that named RT-One as its first commercial customer. HyperPOD is a packaged system built on Supermicro servers with NVIDIA accelerated computing, and it uses DDN Infinia for storage. That means the result reflects DDN's other data platform, not EXAScaler.

In a 24-hour training test at the AMPHIX Center of Excellence in California, RAVEL recorded 98.8 percent GPU utilization, with zero thermal throttling or GPU errors and 37C of thermal headroom at peak load. DDN said the system had eight NVIDIA RTX PRO 6000 GPUs and 500TB of Infinia on Supermicro hardware. The job trained a GPT-2-architecture large language model on 70TB of raw training data, with all of it read from and written back to storage over S3 across the network. GPUs held 95 percent utilization or higher for 98 percent of the run, and hardware telemetry was sampled once per second.

Forklift carrying an equipment rack toward a containerized data module with an external cooling unit at the AMPHIX site, where the first DDN HyperPOD is installed

Those are strong numbers for a single-node-class result, but they come with limits. DDN describes this as an AMPHIX internal test, and the results haven't been independently audited. Eight GPUs usually fit in one server, so the run shows a small HyperPOD keeping its GPUs busy for a day while streaming training data over S3. It doesn't show behaviour at cluster scale. DDN hasn't published training throughput, the inference results RAVEL also collected, or the server and network configuration. AMPHIX is offering the system for customer proofs of concept at the California site, and Strata Expanse and RAVEL plan to connect more than 100 sites. If you're weighing a similar build, ask for the network layout and the throughput numbers before you trust the utilization figure as a sizing baseline.

What to Watch Next

For most administrators, the near-term action is to note the dates and stop there. The EXAScaler 7 software-defined beta is planned for early 2027, the multi-tenancy and data protection features arrive in the first half of 2027, and the qualified third-party hardware list and licensing terms are still missing. Once DDN publishes the vendor list and minimum requirements, that's the point to start a real evaluation. Until then, it's a useful signal that DDN wants EXAScaler to run on hardware you choose, and that's a meaningful shift for a file system that has mostly arrived as a box.

DDN EXAScaler

Source: Harold Fritts · www.storagereview.com

Neighbouring cells