How DeepL Built a Translation Powerhouse with AI with CEO Jarek Kutylowski

July 8, 2025

42 min episode · 2 min read

Jarek Kutylowski

Episode

42 min

Read time

2 min

Topics

Leadership, Artificial Intelligence

AI-Generated Summary

Published Dec 21, 2025

Key Takeaways

✓Model Architecture: DeepL builds translation-specific architectures that balance accuracy with fluency, combining copying mechanisms with creative text generation for superior results over general-purpose models.
✓Data Center Strategy: Building proprietary GPU infrastructure since 2017 enabled DeepL to maintain competitive advantage when cloud compute was scarce, requiring significant upfront investment but ensuring control.
✓Context Integration: Translation quality improves dramatically when models receive document context and company-specific terminology rather than processing isolated sentences, unlocking enterprise use cases previously requiring humans.
✓Speech Translation Latency: Real-time speech translation requires optimizing for speed over perfect quality, with latency being more critical than tone preservation for maintaining conversation flow.

What It Covers

DeepL CEO Jarek Kutylowski explains how his translation company competes against Google and OpenAI through specialized models, proprietary data, and enterprise workflows.

Key Questions Answered

•Model Architecture: DeepL builds translation-specific architectures that balance accuracy with fluency, combining copying mechanisms with creative text generation for superior results over general-purpose models.
•Data Center Strategy: Building proprietary GPU infrastructure since 2017 enabled DeepL to maintain competitive advantage when cloud compute was scarce, requiring significant upfront investment but ensuring control.
•Context Integration: Translation quality improves dramatically when models receive document context and company-specific terminology rather than processing isolated sentences, unlocking enterprise use cases previously requiring humans.
•Speech Translation Latency: Real-time speech translation requires optimizing for speed over perfect quality, with latency being more critical than tone preservation for maintaining conversation flow.

Notable Moment

Kutylowski personally assembled DeepL's first GPU servers in 2017, building data centers when cloud providers couldn't supply sufficient compute for neural translation models.

Know someone who'd find this useful?