This German Startup Just Raised $6.5M to Make AI Run Faster on Any Chip

·
Listen to this article~4 min

Daisytuner raises €6M to automate AI performance engineering, launching Daisyflow to run models faster on any chip. Could this fix the AI compute bottleneck?

What if the biggest bottleneck in AI isn't the chips themselves, but the software that runs on them? Daisytuner, a Darmstadt-based deep-tech startup, just raised €6 million (about $6.5 million) in seed funding to tackle exactly that. The round was co-led by Elaia and LEA Ventures, with Amadeus APEX Technology Fund and Seedcamp joining in. But the money is only part of the story. Alongside the raise, Daisytuner launched Daisyflow, a high-performance inference engine that runs AI models efficiently on any hardware—NVIDIA, AMD, or whatever new silicon comes next. And that could be a game-changer for an industry starved for compute. ### The problem: software is holding back your hardware Here's the thing: most AI models are trained on one vendor's software stack (ahem, NVIDIA's CUDA). That means when you want to run them on different hardware—say, AMD GPUs or a new accelerator—you often have to rewrite code or accept a big performance hit. As Sébastien Lefebvre, Partner at Elaia, puts it: "Compute is the scarcest resource in AI, and most of it currently sits behind a single vendor's software stack." Daisytuner's answer is a self-learning compiler that automates performance engineering. Their open-source compiler, DOCC, takes code from frameworks like PyTorch and translates it into native code for each target chip. Then, AI agents test optimizations around the clock on real hardware, storing every verified result in a performance database. That database gets smarter over time, so each new workload benefits from past optimizations. In plain English: they're replacing the slow, manual work of performance tuning with an automated process that learns as it goes. ### Daisyflow: the engine that makes it happen Daisyflow, the new inference engine, is built on open frameworks and serves language, vision, speech, and recommendation models through APIs that engineering teams already use. It can be deployed inside a customer's own infrastructure, giving companies full control over their data and costs—no need to build a huge in-house team from day one. And the early numbers are promising. Daisytuner says Daisyflow ran Qwen 2.5 up to 18% faster than the industry-standard vLLM on AMD Instinct GPUs. That's not a small margin in a world where every millisecond and dollar counts. > "Demand for AI compute is growing faster than any one chip generation can serve. The hardware to meet it is already out there, from established GPUs to a wave of new accelerators, but software keeps much of it on the sidelines." — Lukas Trümper, co-founder and CEO of Daisytuner ### Why this matters for the AI industry We're at an inflection point. For years, the race was about adding more capacity—more GPUs, more data centers. But now, companies are shifting focus to optimizing existing infrastructure. Why? Because compute is expensive and scarce. Getting more out of the hardware you already have is the new frontier. Daisytuner's timing is spot on. The team, founded in 2024 by Lukas Trümper and Adrian Schmitz, brings over a decade of research from places like ETH Zurich, TU Munich, RWTH Aachen, TU Darmstadt, and Lawrence Livermore National Laboratory. They're not just building a product; they're building on years of cutting-edge compiler technology. The startup plans to use the fresh capital to scale Daisyflow's reach across enterprises, grow its engineering team, and broaden its compiler's support for emerging chips. If they succeed, they could unlock a lot of latent compute power—and that's something the entire AI ecosystem needs. So next time you hear about AI compute shortages, remember: it's not always about making more chips. Sometimes it's about making the ones we have run at full speed. Daisytuner is betting on that—and investors are betting on them.