DeepL Expects 3x Storage Capacity From Existing Hardware With VAST Data

DeepL expects to squeeze roughly three times the storage capacity out of its existing NVMe hardware by deploying VAST Data's storage platform and AI OS, the company has confirmed. The move is the first detailed look at the backend storage behind a translation infrastructure that now runs on Europe's first commercial NVIDIA DGX GB200 SuperPOD.

DeepL, headquartered in Cologne, was founded in 2009 as the Linguee online dictionary by CEO Jarosław "Jarek" Kutylowski. The company reached a valuation of around $2 billion after a $300 million funding round in May 2024, led by Index Ventures, and counts more than 200,000 business customers. It has never previously disclosed what shared storage sits behind its training clusters.

What is DeepL running today?

The company operates two large NVIDIA DGX SuperPOD deployments at EcoDataCenter in Falun, Sweden — a liquid-cooled facility running on 100 percent renewable energy:

• DeepL Mercury (2023): a SuperPOD of 68 DGX H100 systems, each with 8 H100 GPUs. It ranked on the Top500 list with roughly 21.85 PFLOPS of Linpack performance. DELTA Computer Products supplied part of the system.
• Arion (2025): Europe's first commercial DGX GB200 SuperPOD, built with DGX GB200 NVL72 racks — 36 Grace CPUs and 72 Blackwell GPUs per rack, linked by NVLink and InfiniBand. DeepL says Arion cuts the time to translate the entire web to roughly 18 days, from about 194 days previously.

Arion is DeepL's third major SuperPOD deployment and a substantial upgrade over Mercury. The DGX systems carry local NVMe for operating systems, caching and temporary data, but the shared storage backend for training datasets and checkpoints stayed hidden until now.

How did the storage evolve?

The backend has changed shape four times since 2017:

• 2017, on-premises: a centralized bare-metal footprint with active-passive servers attached directly to GPU worker nodes, and local SSD arrays sized for the throughput and latency needs of early convolutional translation models. Kutylowski personally racked the first machines.
• Iceland: a 5.1-petaflop supercomputer at Verne Global in Keflavik, with a tailored local file system optimized for high IOPS to feed parallel training reads.
• Sweden: as large language model work grew, heavy HPC data moved to EcoDataCenter in Falun, leaning on NVMe-over-Fabrics reference architectures to feed the Blackwell-generation systems.
• Hybrid: earlier this year DeepL added AWS as a sub-processor for global scale and performance. Because all translation code is containerized, the port was straightforward. Training remains on proprietary hardware under a strict Zero-Trust framework, while AWS provides scalable cloud storage layers for instant global API distribution.

What does VAST actually change?

DeepL will run VAST AI OS for both file and object storage, and use VAST's SyncEngine technology to migrate data from the existing environment. The claimed threefold capacity gain implies VAST's software will take over the existing SSDs and use them more efficiently through data reduction, fewer copies and similar techniques — more storage room for researchers without buying more physical infrastructure.

Jeff Denworth, a VAST Data co-founder, said: "DeepL is a great example of how quickly the infrastructure requirements behind AI are changing. As AI moves into production, infrastructure reliability becomes just as important as raw performance. Written language, real-time voice and the next generation of AI experiences all place different demands on data, but researchers should not have to redesign the underlying infrastructure every time the model or product changes."

Stefan Andresen, DeepL's chief scientist, said the platform also removes a workflow bottleneck: "VAST gives us the performance and reliability we need while allowing us to make significantly more efficient use of the infrastructure we already operate. Just as importantly, it gives us a path to make data more easily accessible across our own data centres and, over time, cloud environments, reducing the manual data movement that can get in the way of research. That becomes particularly valuable as we continue investing in data-intensive areas such as DeepL Voice."

Why the product mix drives the storage bill

DeepL's portfolio now spans more than text. DeepL Translator handles documents in over 100 languages with formatting preserved. DeepL Write, a Pro add-on, polishes grammar and tone. DeepL Voice, launched in 2024, delivers real-time speech translation for Zoom, Microsoft Teams and Google Meet, and can now preserve a speaker's voice characteristics across languages. An API embeds translation into customer tools, and emerging capabilities include Translation Flow automation and a planned DeepL Agent.

Voice in particular multiplies data demands — a point Andresen singled out. The company positions itself against general-purpose rivals such as Google Translate by prioritizing natural output and context handling, citing contronyms like "oversight," which maps to seven distinct German words, as a case where its models choose correctly more often.

What comes next

If the projected threefold capacity gain holds in production, DeepL can scale model development without a matching build-out of physical storage at Falun — and VAST gains a reference customer running Blackwell-generation SuperPODs at commercial scale. The stated roadmap points toward unified data access across DeepL's own data centres and, over time, its AWS cloud estate, which would reduce the manual data movement that currently slows its research teams.