Daisytuner Raises €6M to Speed AI Code on Any Chip
Daisytuner has raised a €6M seed round from Elaia, LEA Ventures, Amadeus APEX and Seedcamp to automate AI model tuning across Nvidia, AMD and next-generation chips.
By Daniel Okafor
2 min read

What's News
- Daisytuner secured €6 million in seed funding co-led by Elaia and LEA Ventures
- Early results show Qwen 2.5 running up to 18% faster than vLLM on AMD Instinct GPUs
- Founded in 2024 by Lukas Trümper and Adrian Schmitz, based on research from ETH Zurich, TU Munich and Lawrence Livermore
- Amadeus APEX Technology Fund and Seedcamp also joined the round
- Funding will expand the Daisyflow platform, the team and compiler technology
German deep tech startup Daisytuner has closed a €6 million seed round to automate the tuning of AI models to the chips they run on. Elaia and LEA Ventures co-led the round. Amadeus APEX Technology Fund and Seedcamp also participated.
The company attacks a costly gap in AI infrastructure: software that leaves much of a chip's available performance unused. Its answer is a two-part system. A self-learning, open-source compiler converts code from frameworks such as PyTorch into native code for Nvidia, AMD and next-generation chips. A patented optimisation platform then lets AI agents test improvements on real hardware around the clock. Every verified improvement gets stored and reused, so each new workload speeds up the next.
What do the early results show?
Daisytuner's production platform, Daisyflow, runs inside customers' own infrastructure. Early benchmarks show Qwen 2.5 running up to 18% faster than vLLM on AMD Instinct GPUs. That number matters for buyers weighing alternatives to Nvidia's dominant stack, where porting work often demands manual rewrites and months of engineering time.
The comparison is concrete: same model, same class of hardware, one serving layer swapped for another. An 18% gain on inference speed translates directly into fewer GPUs needed for the same traffic.
Who built it, and why now?
CEO and co-founder Lukas Trümper argues the timing is structural. Compute demand, he said, is outpacing any single chip generation, yet software keeps much available hardware sidelined. He founded the company in 2024 with Adrian Schmitz. The technology draws on research from ETH Zurich, TU Munich and Lawrence Livermore National Laboratory.
That research pedigree sits at the core of the product. The compiler learns from each optimisation it verifies on physical hardware. Over time, the platform accumulates a library of proven speedups that apply to new workloads — a compounding asset rather than a one-off tuning service.
What did the investors say?
Sébastien Lefebvre of Elaia framed the value proposition in terms of portability: Daisytuner lets code run at full performance on any silicon without rewrites. Alexi Malikotsinas of LEA Ventures pointed to two drivers behind the firm's commitment — compute efficiency and sovereignty. Sovereignty has become a live concern for European enterprises weighing dependence on US chip and software stacks.
Where does the €6 million go?
The funding will expand Daisyflow, grow the team and extend the compiler technology. Three workstreams, each tied to the same bet: that chip-specific optimisation can be automated rather than hand-built.
The seed round signals investor appetite for infrastructure tools that squeeze more from existing hardware instead of waiting for faster chips. If Daisytuner's self-learning approach scales across Nvidia, AMD and next-generation silicon, its stored library of verified optimisations becomes harder for rivals to replicate — a moat built from accumulated hardware test data rather than code alone.
By Daniel Okafor
Source: theaiinsider.tech
More from Daniel Okafor
Show full bio
Correspondent covering money, funding and market shifts around entrepreneurship, startups and venture funding at Business Bearings.
2,538 articles