The AI infrastructure conversation is still dominated by compute. Faster GPUs, larger clusters and more sophisticated interconnects rightly attract attention because they determine how quickly models can be trained and how efficiently inference can be delivered. But as compute performance increases, another constraint becomes more visible: the data layer underneath it. That is particularly apparent in Language AI.
Training advanced language models already requires large and continuously evolving datasets. Add real-time voice, synthetic training data, larger model architectures and hybrid infrastructure environments, and the demands on storage change significantly. Performance still matters, but so do reliability, capacity efficiency, data mobility and the ability to support different access methods without continually rebuilding the surrounding architecture.
DeepL’s infrastructure evolution provides a useful example.
The company has invested heavily in accelerated computing. Its latest NVIDIA DGX SuperPOD, Arion, is based on NVIDIA Grace Blackwell architecture and represents a major step forward from its previous Mercury system. DeepL has said that a hypothetical translation of the entire internet would take Mercury 193 days, compared with just over 18 days on Arion. The purpose is not simply speed for its own sake. More compute enables DeepL researchers to train larger models, work with more data and explore new Language AI capabilities.
But faster compute changes the balance of the system.
As GPUs become capable of processing more data, the infrastructure feeding them has to scale accordingly. A high-performance accelerator that is waiting for data is still an underutilised accelerator. And if researchers need to manually copy datasets between locations, work around unreliable systems or maintain multiple storage approaches for different workloads, the bottleneck shifts away from compute and into the data architecture.
This is one of the reasons DeepL is deploying VAST Data’s AI Operating System as a new data foundation for its research and development environments. The objective is not simply to add more storage capacity. It is to create a more reliable, efficient and flexible data layer that can keep pace with increasingly diverse AI workloads.
One of the most tangible elements is capacity efficiency. Based on its current deployment plans, DeepL expects to achieve roughly three times the storage capacity from its existing hardware. Rather than scaling physical NVMe infrastructure in direct proportion to data growth, the company expects to extract substantially more usable capacity from the hardware it already operates.
That matters for more than storage cost. AI infrastructure increasingly has to be considered as a complete system. GPUs, networking, power, cooling and data infrastructure all represent significant investment. Improving utilisation in one layer can increase the efficiency of the whole platform. It is also important not to confuse capacity efficiency with a simple “more terabytes” story.
For research teams, additional usable capacity means more room for datasets, model checkpoints and experimentation. It can reduce the pressure to constantly decide what must be moved, duplicated or removed simply to create space for the next training cycle. DeepL’s own approach to model development makes this especially relevant. The company has described the use of synthetic data as a fundamental part of its training methodology and has said its techniques can increase training-data volume by a factor of 1,000. As compute capability grows, the ability to generate and work with larger quantities of useful training data becomes an advantage, but only if the data infrastructure can absorb it.
The second architectural issue is the relationship between file and object storage. AI environments rarely move cleanly from one to the other. Researchers often have established workflows built around file interfaces, tools and directory structures. At the same time, object storage has become increasingly important for scalable AI data architectures and cloud-oriented workflows.
Forcing a research organisation to choose one model and immediately abandon the other creates unnecessary disruption. DeepL will use VAST AI OS for both file and object storage, allowing its existing workflows to continue while providing a path toward greater use of object storage over time. The significance is that both access models can sit on a common underlying platform rather than requiring separate data silos and additional copies. This is particularly important because data movement itself is becoming one of AI’s hidden infrastructure costs.
DeepL initially plans to deploy VAST in its own data centres and, over time, extend its use into cloud environments. The goal is to give researchers more consistent access to data across infrastructure locations and reduce the manual copying of datasets that can otherwise slow experimentation.
The broader DeepL environment is already hybrid. The company has added AWS as an infrastructure provider alongside its own European data-centre infrastructure, while its large-scale AI systems include NVIDIA DGX SuperPOD deployments.
In that kind of environment, copying data each time compute moves is not a scalable long-term model. Large datasets take time to transfer, duplicates consume capacity, and every additional movement introduces another operational process researchers have to manage.
Migration is another practical part of the architecture. DeepL is using VAST SyncEngine to move data from its existing environment onto the VAST platform. This allows the transition itself to be treated as part of the data architecture rather than as a separate, manually managed migration project.
Then there is reliability. AI infrastructure is often benchmarked around maximum throughput, but research organisations also depend on predictable behaviour. Training jobs can run for long periods, datasets are accessed repeatedly by different teams, and the cost of infrastructure instability is not limited to an outage window.
When researchers stop trusting the platform underneath them, their behaviour changes. They create workarounds, duplicate data, avoid certain workflows or spend engineering time solving infrastructure problems rather than model problems.
That is why industrial-grade reliability becomes increasingly important as AI moves from isolated research projects into continuously operated products. DeepL Chief Scientist Stefan Andresen described the requirement clearly when the deployment was announced: DeepL needs both performance and reliability as its researchers push Language AI into more demanding areas. Voice AI makes the point particularly well. DeepL Voice has extended the company’s technology from written translation into live multilingual conversation.
Recent advances allow this technology to preserve characteristics of a speaker’s voice across languages. These workloads are considerably more data-intensive than text alone and create additional demands across training, data access and production infrastructure. This is where the architecture of the data layer starts to matter more than any single storage benchmark.
The platform has to serve researchers using familiar workflows, support increasingly object-oriented AI pipelines, provide consistent performance as datasets grow, make efficient use of existing NVMe capacity and ultimately allow data to be accessed across different infrastructure environments. And, it has to do those things while remaining sufficiently reliable that researchers do not need to think about it; that is the larger infrastructure lesson from DeepL’s deployment.
As AI compute continues to accelerate, organisations cannot afford to treat the data layer as a secondary consideration. Faster GPUs increase the value of fast, reliable access to data, and expose the cost when that access is missing.
For research organisations such as DeepL, the objective is straightforward: give scientists reliable access to the data they need, wherever they need it, without requiring them to redesign the infrastructure every time the model, modality or deployment environment changes.
AI infrastructure is becoming more about system balance and less about individual components. A faster accelerator does not solve a slower data path; more capacity does not solve fragmented access; and cloud flexibility does not help if every move requires another copy of the data.
The winners will be the architectures that remove those trade-offs rather than simply shifting the bottleneck somewhere else.
Dr. Jakob Jung is Editor-in-Chief of Security Storage and Channel Germany. He has been working in IT journalism for more than 20 years. His career includes Computer Reseller News, Heise Resale, Informationweek, Techtarget (storage and data center) and ChannelBiz. He also freelances for numerous IT publications, including Computerwoche, Channelpartner, IT-Business, Storage-Insider and ZDnet. His main topics are channel, storage, security, data center, ERP and CRM.
Contact via Mail: jakob.jung@security-storage-und-channel-germany.de