VC & Startups/Nov 16 edition/Reporting & analysis
FundingM&AEcosystems

FundingRounds, leads & terms

Daisytuner raises €6M seed led by Elaia and LEA Ventures to run AI inference across non-Nvidia chips

Darmstadt-based Daisytuner has raised a €6M seed round co-led by Elaia and LEA Ventures. It has also launched Daisyflow, a self-hosted inference stack, alongside its open-source docc compiler. Both aim to run AI models efficiently on Nvidia, AMD and alternative accelerators.

From the Tech.eu article of November 16, 2026 and cited sources: Elaia and LEA Ventures co-led the €6M seed round, with Amadeus APEX Technology Fund and Seedcamp joining. [1] [2] [9] [11]

SOURCES[1] Tech.eu: AI inference startup Daisytuner raises €6M seed tech.eu[2] Amadeus Capital: Daisytuner raises €6M in seed funding amadeuscapital.com[9] Tech Funding News: Elaia and LEA Ventures back Daisytuner techfundingnews.com

AI-assisted, not human-edited. Methodology Corrections Who owns this site

KEY FACTS
What happened
Elaia and LEA Ventures co-led the €6M seed round, with Amadeus APEX Technology Fund and Seedcamp joining. [1] [2] [9] [11]
Original source
Tech.eu article, independent perspective
Brief by
AT Pulse research desk AI-assisted; publisher and contact on the desk page
Studies
arXiv 1902.10345 (Feb 2019) [8] date from arXiv ID
When
November 16, 2026 article published
Evidence
13 sources across 13 sites

Reach: not measured yet

Illustration from Tech.eu: Daisytuner raises €6M seed led by Elaia and LEA Ventures to run AI inference across non-Nvidia chips
Image: Tech.eu — Original article ↗
THE CORE IDEAS4 TAKEAWAYS
01

Elaia and LEA Ventures co-led the €6M seed round, with Amadeus APEX Technology Fund and Seedcamp joining. Daisytuner was founded in 2024. Its valuation was not disclosed. Most of the money will go to hiring, taking the team from 12 people to more than 20 by early 2027. The rest will buy chips for testing and for training its agents. [1] [2] [9] [11]

02

The open-source docc compiler converts C/C++, Python and PyTorch code into Stateful Dataflow Multigraphs, a representation that comes from ETH Zurich's DaCe framework. It optimises them and then generates code for OpenMP, CUDA or ROCm. AI agents test candidate optimisations on real hardware and store the ones that work in a performance database. Daisyflow handles serving, routing and batching, and customers run it on their own servers. [4] [7] [8] [9]

03

The only performance figure is a company-run benchmark: Qwen 2.5 ran up to 18% faster than vLLM on AMD Instinct GPUs. The test setup has not been published, and AMD keeps improving vLLM on ROCm, so the baseline is moving. [12] [13]

04

Qualcomm's agreed all-stock purchase of Modular, valued at about $3.9B, puts a price on software that runs AI across different hardware. Daisytuner has named no customers yet. It says it works with Tenstorrent and Q.ANT and is close to its first enterprise deals. [3] [6] [9]

WHY IT MATTERS

investors are funding AI software that runs across chip vendors, and Qualcomm's roughly $3.9B Modular deal priced a similar business.

Read the full assessment

Implication: teams using AMD or alternative accelerators may get more deployment options, but each workload still needs its own testing.

Sources are gathered; writing the dossier now.

Executive brief

Daisytuner, a two-year-old startup in Darmstadt, Germany, raised a €6M seed on 16 November 2026 to run AI inference efficiently on chips other than Nvidia's. Elaia and LEA Ventures co-led the round (Amadeus Capital; Tech.eu). The company's main competitor is Modular, which Qualcomm agreed to buy for about $3.9B (BGOV). That deal suggests what hardware-agnostic inference software can be worth. Daisytuner's only performance number is its own: Qwen 2.5 ran up to 18% faster than vLLM on AMD Instinct GPUs.

What changed and event timeline

  1. Academic tech-transfer recognition

    The company's website lists a HiPEAC Technology Transfer Award for Daisytuner. The company was founded in 2024 (;).

  2. Compiler research at CGO

    Co-founder Lukas Trümper co-authored a paper at CGO '25 on automatic loop scheduling, called "a priori loop nest normalization." This is the research behind the company's compiler ().

  3. Qualcomm agrees to buy Modular

    The all-stock deal was valued at about $3.9B and expected to close in the second half of 2026. It put a price on cross-hardware AI software (;).

  4. €6M seed and Daisyflow launch

    Elaia and LEA Ventures co-led, with Amadeus APEX Technology Fund and Seedcamp participating. The company says Daisyflow is now live in production (;).

Capabilities and access

  • Daisyflow: an inference stack that handles serving, routing and batching. Customers deploy it on their own servers (Daisytuner; Tech.eu).
  • docc compiler: open source under the BSD-3-Clause licence and installable with pip install docc-compiler (GitHub).
  • Hardware the company lists: Nvidia CUDA, AMD ROCm, Tenstorrent Metalium, the Q.ANT NPU Gen 2 (a photonic processor) and RISC-V-based TPUs (Daisytuner).
Read the full section
  • Daisyflow: an inference stack that handles serving, routing and batching. Customers deploy it on their own servers (Daisytuner; Tech.eu).
  • docc compiler: open source under the BSD-3-Clause licence and installable with pip install docc-compiler (GitHub).
  • Hardware the company lists: Nvidia CUDA, AMD ROCm, Tenstorrent Metalium, the Q.ANT NPU Gen 2 (a photonic processor) and RISC-V-based TPUs (Daisytuner).
  • Model version: the benchmark is described only as "Qwen 2.5". Parameter count and exact variant are not disclosed.

Technical analysis for researchers and developers

  • Architecture: docc turns C/C++ (via LLVM), Python and PyTorch code (as a torch.compile backend) into Stateful Dataflow Multigraphs (SDFGs).
  • Lineage: docc builds on the reference implementation of ETH Zurich's DaCe framework (SDFG paper).
  • Search loop: AI agents test candidate optimisations on real hardware and store the ones that work in a performance database (TFN).
Read the full section
  • Architecture: docc turns C/C++ (via LLVM), Python and PyTorch code (as a torch.compile backend) into Stateful Dataflow Multigraphs (SDFGs). It optimises them with data-centric and polyhedral analysis, then generates code for OpenMP, CUDA or ROCm. Metal support is in progress (GitHub).
  • Lineage: docc builds on the reference implementation of ETH Zurich's DaCe framework (SDFG paper).
  • Search loop: AI agents test candidate optimisations on real hardware and store the ones that work in a performance database (TFN). This resembles the founders' earlier "performance embeddings" research (arXiv).
  • Reproducibility: the company has not published the vLLM version, GPU SKU, batch sizes or metric behind the 18% figure.

Claims and evidence

  • Round size and investors: the company announced them, and Tech.eu, TFN and FinSMEs report the same figures (Tech.eu; TFN; FinSMEs). No filing confirms them.
  • Valuation: not disclosed (TFN).
  • "Up to 18% faster than vLLM": a single company-run benchmark. No independent replication exists (CASRAI).
Read the full section
  • Round size and investors: the company announced them, and Tech.eu, TFN and FinSMEs report the same figures (Tech.eu; TFN; FinSMEs). No filing confirms them.
  • Valuation: not disclosed (TFN).
  • "Up to 18% faster than vLLM": a single company-run benchmark. No independent replication exists (CASRAI).
  • Customers: none named. The company says it is "very close" to its first enterprise deals and works with Tenstorrent and Q.ANT. It says other chipmakers are under NDA (TFN).
  • Investor comments: Elaia partner Sébastien Lefebvre and LEA Ventures general partner Alexi Malikotsinas, both backers in this round, described the company in terms of compute scarcity and technology sovereignty (Amadeus Capital).

Context and prior work

  • SDFG/DaCe: the data-centric representation Daisytuner builds on was introduced by Ben-Nun et al. at SC 2019 (arXiv).
  • Founders' research: Trümper's CGO '25 paper reports beating a baseline C compiler by about 21× on average, and the Polly and Tiramisu auto-schedulers, on its benchmark set (CGO 2025).
  • The baseline is moving: AMD's own ROCm work keeps tuning vLLM, including up to about 17% higher request throughput for Qwen3 MoE models (AMD ROCm blog).
Read the full section
  • SDFG/DaCe: the data-centric representation Daisytuner builds on was introduced by Ben-Nun et al. at SC 2019 (arXiv).
  • Founders' research: Trümper's CGO '25 paper reports beating a baseline C compiler by about 21× on average, and the Polly and Tiramisu auto-schedulers, on its benchmark set (CGO 2025).
  • The baseline is moving: AMD's own ROCm work keeps tuning vLLM, including up to about 17% higher request throughput for Qwen3 MoE models (AMD ROCm blog).

Limitations, safety and contested findings

  • One configuration: the 18% figure covers one model family on one vendor's GPUs, and "up to" marks it as a best case.
  • Small open-source footprint: the docc repository has 27 stars (GitHub).
  • Inconsistent hardware lists: CASRAI lists Intel support, but Daisytuner's own site does not (CASRAI; Daisytuner).
Read the full section
  • One configuration: the 18% figure covers one model family on one vendor's GPUs, and "up to" marks it as a best case. CASRAI's editors say it should be treated as a vendor benchmark and call for workload-specific testing before anyone deploys it (CASRAI).
  • Small open-source footprint: the docc repository has 27 stars (GitHub).
  • Inconsistent hardware lists: CASRAI lists Intel support, but Daisytuner's own site does not (CASRAI; Daisytuner).

Business and practitioner implications

  • Who it suits: enterprises with AMD or alternative accelerators, and European buyers who prioritise sovereignty, can run Daisyflow on their own hardware.
  • How to evaluate it: benchmark it against the latest ROCm vLLM on your own traffic before switching.
  • Scale: the team is 12 people and plans to pass 20 by early 2027.
Read the full section
  • Who it suits: enterprises with AMD or alternative accelerators, and European buyers who prioritise sovereignty, can run Daisyflow on their own hardware.
  • How to evaluate it: benchmark it against the latest ROCm vLLM on your own traffic before switching.
  • Scale: the team is 12 people and plans to pass 20 by early 2027. Most of the money goes to hiring, and the rest buys chips for testing and agent training (TFN).
  • Market signal: Qualcomm's Modular deal shows chipmakers will pay for portability software, so Daisytuner could become an acquisition target.
FOLLOW THE EVIDENCE

The source trail.

Sources (13)
A LITTLE LESS NOISE. A LOT MORE CONTEXT.

Stay curious.
Follow the evidence.

Independent perspectives, the original sources, and room for the questions that don't have easy answers.

How we build the brief
Connect with us

Find us where you already read.