DeepL's core strategy is to build highly specialized models for translation, which they argue are more accurate and less prone to hallucination than general-purpose LLMs. This focus allows them to achieve state-of-the-art quality that unlocks high-value enterprise use cases.
DeepL made an early, strategic decision to build its own GPU data centers and develop custom scheduling software. This control over the full stack, from hardware to proprietary data, is considered essential to maintaining their technological edge and optimizing performance.
The conversation frames AI translation as a bellwether for how AI will disrupt other industries. Traditional translation agencies are shrinking as enterprises adopt decentralized, self-serve AI tools, leading to a massive increase in the overall volume of translated content.
While AI automates the bulk of translation, humans remain critical for high-compliance use cases (e.g., legal, life sciences) and for providing the feedback and curated data needed to train next-generation models. The human role is shifting from mass production to quality assurance, oversight, and data generation.
Keep pulling the thread on Jarek Kutylowski.