DeepL and NVIDIA – the story continues
Unlike previous generations of SaaS products, AI technology requires an incredible synergy of software and computing power. It requires businesses that can carefully coordinate their efforts, anticipate what’s required to keep AI moving forward, and invest in the future together.
This is exactly the type of relationship that DeepL has built with NVIDIA. It’s a relationship we’re celebrating at NVIDIA GTC 2025 this week, the GPU technology conference that is commonly considered the Woodstock of AI.
Several members of DeepL’s technology leadership team will be attending GTC this week, including our Chief Technology Officer (CTO) Sebastian Enderlein and VP Research Stefan Mesken. DeepL’s Research High-Performance Computing (HPC) Engineer Markus Schnös will be presenting the story of how our two businesses worked together to deliver the industry-leading training and inference performance of DeepL’s next-gen LLMs. It’s a story of what’s involved in pushing the boundaries of AI performance today; not just investing in the right hardware, but developing innovative ways of using that hardware to maximize its potential.
The next chapter in an innovation story
It's also a story that’s just getting started. DeepL will shortly become the first company in Europe to deploy the NVIDIA DGX SuperPOD with DGX GB200 systems. NVIDIA’s Chief Solutions Architect, Thomas Schoenemeyer describes this cluster of NVIDIA GB200 Grace Blackwell Superchips as the AI equivalent of a championship-winning Formula 1 car. In NVIDIA’s previous testing of the SuperPOD’s capabilities, it leads on every benchmark in the open-source MLPerf framework, which measures the performance of hardware for training and inference. It’s capable of handling trillion-parameters models, enabling DeepL researchers to test new ideas faster, train models more quickly, and deliver near real-time inference.
Having an AI Formula 1 car in the garage will enable us to deliver Language AI that is more capable, more accurate, and more adaptable, every year. It enables us to be bold and ambitious in the tasks we set for DeepL, taking on new use cases through solutions like DeepL Voice and reimagining the experience of working with AI through features like Clarify. And just as before, these gains in performance are a result not just of investing in the right machines, but in building a relationship around how we use and develop those machines.
Working together to push the boundaries of AI
Our collaboration with NVIDIA involves discussions of the types of training and inference capabilities that LLMs will require, and the configuration of machines such as the new DGX SuperPOD that can deliver them. It also involves working together on optimization algorithms for the software running supercomputers, which itself can unlock major gains in energy efficiency and enable us to do more with the extra capacity we have. Anyone who’s able to watch Markus’s presentation this week will get a great insight into what that means in practice.
Finally, and crucially, a close working relationship with NVIDIA helps to inform our collaboration with other suppliers, as well. EcoDataCenter, which houses DeepL’s existing supercomputers, and is one of the most advanced and sustainable data centers in the world, had anticipated that the next generation of machines would require liquid cooling to deliver their step-change in performance.
We worked with EcoDataCenter to put the infrastructure required for this liquid cooling in place, ready to deploy the most powerful machine NVIDIA has yet built. That’s why DeepL was ready to become the first company in Europe to leverage the new SuperPOD. It’s this ability to anticipate the needs of different elements in the AI ecosystem that helps us to keep pushing the boundaries of what this technology can do.
We are honored to be presenting at GTC this week, and we look forward to the exciting innovations that will come from our ongoing collaboration with NVIDIA.