German deep-tech startup Daisytuner has secured €6 million in seed financing to develop software that automatically optimizes artificial intelligence models for different processors, addressing a growing challenge for companies seeking alternatives to expensive and specialized computing infrastructure.
The funding announced November 16 was co-led by Elaia and LEA Ventures, with participation from Amadeus APEX Technology Fund and Seedcamp.
The Darmstadt-based company is developing technology that reduces the manual engineering work required to run AI models efficiently across Nvidia, AMD and emerging accelerator architectures.
Its approach reflects a broader change in Europe’s AI industry: startups are increasingly targeting the software infrastructure that determines how efficiently AI workloads run, rather than competing exclusively in model development.
A Compiler Designed to Work Across AI Chips
Founded in 2024 by Lukas Trümper and Adrian Schmitz, Daisytuner combines compiler technology with automated performance optimization.
Its open-source compiler, DOCC, converts code from frameworks such as PyTorch into instructions designed for specific processors.
According to the company’s technology announcement, AI agents repeatedly test different optimization strategies on real hardware.
The compiler checks whether an optimization preserves computational correctness before adding the verified result to a performance database.
This allows successful techniques to be reused when the system encounters similar workloads.
The commercial platform, Daisyflow, brings those capabilities into production environments, supporting language, vision, speech and recommendation models.
It also allows customers to keep workloads within their own infrastructure rather than transferring them to an external inference service.
Early Benchmarks Show Performance Improvements
Daisytuner reports that Daisyflow runs the Qwen 2.5 model up to 18% faster than vLLM on AMD Instinct GPUs.
The figure comes from the company’s initial performance testing and should not be interpreted as an independently verified improvement across all models or processors.
In an interview with Tech Funding News, Trümper explained that the company’s focus had shifted from moving computing tasks onto Nvidia GPUs toward enabling the same workloads to run efficiently across competing hardware.
The problem is increasingly relevant as businesses adopt open-weight models and consider alternative AI accelerators.
Different processors typically require different optimization techniques. Adapting software to new hardware can therefore involve substantial engineering effort even when the underlying models remain unchanged.
European Investors Target AI Infrastructure Independence
The funding will support enterprise adoption of Daisyflow, additional compiler development and recruitment.
Trümper previously conducted doctoral research at ETH Zurich, while the broader team brings experience from institutions including TU Munich, RWTH Aachen and TU Darmstadt.
Elaia partner Sébastien Lefebvre identified software dependence on a single hardware ecosystem as an important industry bottleneck.
LEA Ventures general partner Alexi Malikotsinas similarly connected the investment with computing efficiency and technological sovereignty.
The investors’ central argument is that companies should be able to select processors based on cost and performance without repeatedly rewriting applications.
However, the startup still needs to demonstrate that its automated optimizations deliver consistent improvements across different AI models, hardware generations and production environments.
For European founders and engineering teams, Daisytuner represents an emerging opportunity in AI infrastructure: competing through better software utilization of available computing resources rather than simply purchasing more powerful hardware.
Whether that strategy can reduce operating costs at scale will be an important test of its commercial potential.