Document Processing
DeepL needed to scale high-quality, context-aware translation for millions of daily users and hundreds of thousands of business customers without sacrificing latency, accuracy, or the ability to train larger models for new features like voice translation.
Reported impact
With Arion, DeepL cut translation processing roughly 10x, reducing a full-internet translation estimate from 193 days to about 18 days and dropping Oxford English Dictionary translation time from 39 seconds to 3.72 seconds while sustaining near-real-time multilingual inference for millions of users and nearly 200,000 businesses.
DeepL deployed its “Arion” AI infrastructure on an NVIDIA DGX SuperPOD built from DGX GB200 systems to remove the trade-offs between model size, training speed, and translation latency. The Blackwell-based cluster links 72 GPUs into one unit, then DeepL pairs that hardware with NVIDIA TensorRT-LLM and Transformer Engine so both training and inference can move from 16-bit to 8-bit floating-point formats while staying highly optimized down to the kernel level. DeepL also uses Mixture-of-Experts architectures, synthetic data generation that increases training volume by 1,000x, and EcoDataCenter’s liquid-cooled facility to sustain efficient large-scale operation. The result is a platform that can serve millions of users with near-real-time multilingual inference, power products like DeepL Voice and Clarify, and keep larger language models responsive enough for production deployment.
Translation requests reach Arion, GB200-based models process the multilingual context, TensorRT-LLM and Transformer Engine keep inference and training efficient on the DGX SuperPOD, and DeepL returns low-latency translations or voice outputs to end users and business workflows.
Arion can translate the entire Internet in approximately 18 days compared to 193 days with the previous system.
DeepL also developed sophisticated synthetic data generation techniques that boost data volume by a factor of 1,000.
DeepL's engineering team embraced floating-point optimization, moving from 16-bit to 8-bit floating-point training and inference using NVIDIA's Transformer Engine and TensorRT-LLM.
Oxford English Dictionary, which took 39 seconds previously, now takes just 3.72 second—a 10X speedup.
The company’s AI infrastructure, named “Arion”—powered by an NVIDIA DGX SuperPOD with DGX GB200 systems—enables millions of daily users across dozens of languages to receive accurate translations with near-zero latency.
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
Related
NVIDIA Sample Sep 22, 2026 Sluicebox Sluicebox needed to turn fragmented bills of materials, supplier records, PDFs, CAD files, and emissions inventories into timely product carbon assessments, but manual life-cycle-analysis work was slow, expensive, and… 51-80% lower inference costs Lucy · NVIDIA Nemotron 3 Ultra · carbon graph NVIDIA Sample Sep 1, 2026 Onetera Onetera needed to help more than 90 municipalities turn fragmented regulations, zoning maps, permit rules, and legacy documents into fast resident-facing services without imposing heavy IT overhead or expensive manual… 1.7x increase in data throughput GPT OSS 20B · Government Intelligence Agent · LangGraph NVIDIA Sample Jun 26, 2026 MISA MISA needed to move from expensive, externally hosted LLM APIs to a self-hosted Vietnamese AI stack that could protect sensitive data, reduce latency and operating cost, and support many concurrent public-service and… $200,000 in monthly cost savings MISA AVA · NVIDIA A40 GPUs · NVIDIA DGX H200 