I4U News / AI

AI · AI infrastructure

Daisytuner bets AI chip lock-in is a software problem

Darmstadt-based Daisytuner has raised €6m and launched Daisyflow, an inference stack built on its own compiler. The bet: if you can tune AI code for any chip, hardware loses its lock-in and data centres start buying chips on price rather than compatibility.

Researched and drafted with AI support, directed and edited by Manfred Lugmayr; facts checked against the sources. How we work

An engineer with a laptop connects an older black server rack to a new copper-coloured rack with a cable in a data centre (AI image)
AI-generated illustration

If you run an AI model in your own data centre, the most important decision was probably made before anyone asked you: the chip platform. Your code is tuned for it. So are your tools. Moving to different hardware means rewriting, adapting and retesting, which Daisytuner says takes months and costs millions.

A startup from Darmstadt, near Frankfurt, wants to make that move boring. It has just raised €6m to try.

What happened

Daisytuner GmbH closed a €6m seed round co-led by Elaia and LEA Ventures, with the Amadeus APEX Technology Fund and Seedcamp joining, according to the company's announcement on 16 November. Tech.eu reports the same figures. In 2025 the company raised a pre-seed round of around €1m led by LEA Partners, HiPEAC reported.

The founders are Lukas Trümper, whose research on AI-driven compilers at ETH Zurich led to the company, and Adrian Schmitz. The product shipped the same day: Daisyflow, an inference stack that serves language, vision, speech and recommendation models inside the customer's own infrastructure, behind OpenAI-compatible APIs. Underneath sits docc, an open-source compiler that takes PyTorch, NumPy and C/C++ and generates native code for CUDA, ROCm and newer chip architectures. Every optimisation is proven correct before use and stored in a performance database, the company says.

The one hard number is the company's own: on AMD Instinct GPUs, Alibaba's Qwen 2.5 runs up to 18% faster on Daisyflow than on vLLM, the widely used open-source inference engine. There is no independent benchmark yet, and Daisytuner names no customers.

The lock-in lives in software, not silicon

Here is what interests me. Europe's AI debate is about chips: who builds fabs, who gets allocation. But data centres rarely stay with a vendor because of the silicon. They stay because of the stack on top. CUDA is Nvidia's software platform, and every line of code tuned for it makes the next chip more expensive to adopt, whatever its spec sheet says.

To me that is a textbook value transfer. If your model only runs efficiently on one platform, you don't negotiate on price. You accept it. A compiler that makes the same code fast on AMD, on RISC-V accelerators or on Tenstorrent cards hands that leverage back to the operator. In January Daisytuner showed it on scientific software: OpenFOAM and the HPCCG benchmark ran on Tenstorrent's Wormhole and Blackhole cards, at roughly 60 watts versus 400 to 600 watts for server CPUs, per the company.

The second-order effect: if this holds, switching chips stops being a migration project and becomes a procurement decision. Cloud providers in Germany, Austria and Switzerland that sell sovereign compute could mix hardware on price, power and availability rather than compatibility. Providers with mixed hardware fleets are exactly the customers Daisytuner is targeting. And European chip startups would get a route into production without first building their own software ecosystem.

The catch

My rule of thumb: code is no longer a moat, distribution is. That cuts both ways here. If Daisytuner is right, part of Nvidia's lead shrinks. But the same logic applies to the startup. Chipmakers, hyperscalers and open-source projects like vLLM keep improving their own optimisation. An 18% gain on one model and one GPU family is a start, not proof.

The winner will be whoever signs the first paying data centres and hardware vendors and pulls their workloads into the performance database. In my view that database is the real asset, not the compiler, which is open source. €6m pays for a team and a few enterprise pilots. It won't last long in a sales fight against US platforms. The next round will show whether customers follow.

There is a pattern, though. As with Plural's new fund, the money is going into the layer beneath the applications, not the next chat app. More in our AI section; the round is now in the funding tracker.

What I would do now

  • If you run your own AI inference: → list your three most expensive models and check how many chip platforms they could run on in production today. If the answer is one, that is a procurement risk, not a technical detail.
  • If you invest in deep tech: → don't start with the benchmark. Ask about pilot customers with their own data centres, and who owns the performance data.
  • If you build chips: → negotiate compiler partnerships early. Without a software path, the best hardware never reaches a customer's production stack.

Bottom line: Europe's dependence on AI compute is a software problem too, and Darmstadt now has €6m riding on the bet that a compiler can shrink it.

The facts

  • Darmstadt-based Daisytuner raised a €6m seed round co-led by Elaia and LEA Ventures.
  • Amadeus APEX Technology Fund and Seedcamp also invested.
  • Daisyflow runs in the customer's own infrastructure; its open-source compiler docc targets CUDA, ROCm and new chips.
  • Per the company, Qwen 2.5 runs up to 18% faster than on vLLM on AMD Instinct GPUs; not independently verified.

Questions and answers

What does Daisytuner do?

Darmstadt-based Daisytuner builds a self-learning compiler (docc, open source) and the Daisyflow inference stack, designed to make AI models and other code run fast on different chips without rewriting, inside the customer's own data centre.

Who invested in Daisytuner?

Elaia and LEA Ventures co-led the €6m seed round in November 2026, joined by the Amadeus APEX Technology Fund and Seedcamp. A pre-seed round of around €1m led by LEA Partners came in 2025.

Is Daisyflow faster than vLLM?

According to the company's own benchmark, Qwen 2.5 runs up to 18% faster than on vLLM on AMD Instinct GPUs. No independent test has been published yet.

Sources

  1. Daisytuner (Unternehmen): Daisyflow launch and €6M seed round (16 Nov 2026) primary source
  2. Daisytuner (Unternehmen): Hardware lock-in and portability (31 Jan 2026) primary source
  3. Daisytuner (Impressum): Daisytuner GmbH, Darmstadt, HRB 108176 (17 Nov 2026) primary source
  4. Tech.eu: AI inference startup Daisytuner raises €6M seed (16 Nov 2026) independent source
  5. HiPEAC: daisytuner launches beta site following pre-seed funding round (19 May 2025) independent source

Briefing · every Monday

Robotics, AI and capital from DACH. One email a week.

The rounds, people and technology that mattered this week in Germany, Austria and Switzerland, with analysis instead of press-release copy. In English.

More from AI

Section →