NVIDIA: The technology that powers DeepL AI Labs

The new products and solutions that we’re creating in DeepL AI Labs are possible thanks, to our long-standing collaboration with NVIDIA. The latest milestone in that partnership involved DeepL deploying the first NVIDIA DGX SuperPOD with DGX GB200 systems in Europe earlier this year. This new supercomputer is unlocking new possibilities when it comes to training and deploying AI solutions.

We’ve called our new NVIDIA SuperPOD DeepL Arion. It uses NVIDIA’s Grace Blackwell architecture, which connects together ‘islands’ of 72 powerful Blackwell GPUs so that they can act as a single unit. The more GPUs that a SuperPOD connects in this way, the more powerful it becomes, and Arion is a lot more powerful even than our previous NVIDIA supercomputer, Mercury. It would have taken Mercury 193 days to translate the entire internet. If we gave Arion this hypothetical task it could manage it more than 10x faster, in just over 18 days!

Faster machines train larger models

Translating the worldwide web sounds impressive, but what Arion means for our ability to train Large Language Models (LLMs) is even more significant. Simply put, the faster GPUs can communicate, the bigger the models we can build with them. With Arion, we can use simple, scalable architectures to build much bigger LLMs.

We’re able to train these bigger LLMs with scaled-up training data using sophisticated techniques that we pioneered for training our Language AI models. It’s a proven approach to generating synthetic data that has enabled us to continually improve the quality of DeepL LLMs over time, and it will help us to leverage the full potential of the larger models we build.

Innovating with emergent capabilities

AI research shows that when you build larger models, and train them on larger amounts of high-quality data, those models can start to display previously unpredicted capabilities. 

This often takes the form of a model evolving very quickly from finding a task extremely difficult to finding it relatively easy. Such emergent capabilities reward researchers for pushing the boundaries of what AI can do, imagining new problems that can be solved and experimenting with new ways of solving them. We’ve created DeepL AI Labs to establish a pipeline of such experiments. And in doing so, we are making bold but intelligent bets on what DeepL’s models can do next. 

We do this by setting ambitious goals that we know will make a big difference to how people work and how productive they will be. We then experiment intensively, testing ideas for new features that can help us meet these goals. Emergent capabilities mean that even very ambitious experiments can prove surprisingly successful. Expanding the range of goals and solutions that we apply our models to helps new capabilities to emerge even faster.

Enabling more interactive, more valuable AI experiences

The impact of greater compute power and emergent capabilities is already shaping DeepL features, and the experience of people using our tools.

Clarify, the on-demand translation expert that knows when to reach out with intelligent questions to clarify meaning, is an early example of this. When models can understand ambiguity and detect assumptions in the way that Clarify does, they’re able to interact with users in more valuable, human-like ways. This helps to deliver an experience of working with AI that’s more responsive, and delivers far more relevant and impressive results than a model that attempts to reason by itself.

We first developed these interactive capabilities for translation tasks, but they are equally valuable for almost any application of AI. By developing models that can collaborate more naturally, more intelligently and more productively with people, we’re able to help enterprises and other organizations do much more with AI.

A pipeline of AI projects

The projects taking shape in DeepL AI Labs are exploring innovative new ways to fulfill that potential. In doing so, they’re benefitting from another result of our close work with NVIDIA on maximizing the potential of compute power: increased speed of inference. 

Generally speaking, larger AI models come with greater latency, which means that users have to wait longer for the results when asking AI to perform a task. Arion’s increased compute power, combined with NVIDIA’s enabling of FP8 training and FP4 inference, helps to change this. It means that our larger, more powerful models can still perform tasks extremely quickly. 

Arion helps AI agents to perform complex tasks at the speed users require. It also enables ambitious projects like our revolutionary approach to voice-to-voice translation, which depends on being able to intelligently predict what people are saying, and translate as they are saying it.

At DeepL, we’ve never been interested in innovation for innovation’s sake. Rather, we pursue innovative ideas that have real-world impact — specifically when it comes to people’s experience of work and life. DeepL Arion’s capabilities, and our collaboration with NVIDIA and EcoDataCenter that underpin them, are enabling us to test a far wider range of ideas for this type of real-world impact. It’s these rapidly expanding possibilities that make DeepL AI Labs such an exciting space.