Keep pulling the thread on Jarek Kutylowski.
In 2017, DeepL had to build its own data centers because it was unable to acquire sufficient GPU compute from third-party providers.
DeepL considers owning and operating the latest large-scale GPU hardware, such as NVIDIA's DGX and Blackwell systems, as essential to maintaining its competitive edge.
DeepL has observed that its specialized translation models are less prone to hallucination compared to general-purpose generative AI models when used for translation tasks.
DeepL's future strategy involves moving beyond simple sentence translation to address entire enterprise workflows, including review cycles and versioning.
For enterprise customers, any incremental improvement in machine translation quality yields a significant return on investment by reducing the time required for human review.
Yarek Kotelowski believes that training a separate AI model for each customer is generally not a good business strategy, except in highly specialized situations with a clear ROI.
Within enterprises, translation services are becoming a decentralized, self-serve function where individual departments like legal and marketing directly procure solutions from providers like DeepL.
DeepL's primary business focus is on enterprise customers.
DeepL is generating significant revenue from its specialized generative AI use case in translation.
Many traditional human translation companies are shrinking or facing financial trouble due to the rise of generative AI.
The shift to neural machine translation around 2017 created an opportunity for startups like DeepL because incumbent companies had to abandon their existing technologies.
DeepL found that neural network architectures other than the standard Transformer model can be better suited specifically for translation tasks.