Back to Blog
June 20, 2026

Europe’s AI Power Depends on Open Models and RISC-V Open Hardware

AI as the New Layer of Sovereignty

The European Union has a historic opportunity: to move from being mainly the world’s most influential AI regulator to becoming one of the world’s decisive builders of AI infrastructure. This will not happen through declarations about digital sovereignty alone. It will happen only if Europe develops and releases a fully open frontier AI model, competitive with the strongest closed systems from the United States and China, and combines it with a serious industrial strategy for RISC-V open hardware.

AI is no longer just a software product. It is a foundational layer of power. Whoever controls the models, training data, accelerators, data centres, APIs and inference stacks controls a growing part of future production, public administration, education, research, healthcare, culture and democratic communication. If Europe continues to depend mainly on closed commercial APIs and proprietary AI accelerators controlled outside Europe, its public administrations, universities, SMEs and local economies will remain customers of infrastructures they cannot fully inspect, adapt, reproduce or govern.

The alternative is a European public AI infrastructure, a digital CERN for the age of large models. This means shared compute, shared data governance, shared evaluation systems, fully open models, open standards, open software and, crucially, open hardware. Europe cannot build real autonomy at the software layer while remaining structurally dependent at the silicon layer. A sovereign AI strategy must connect models, software, chips, memory, interconnects and deployment infrastructure.

What Fully Open AI Really Means

A model is not fully open simply because its final weights are available. Full openness means access to the training code, architectural choices, data documentation or reproducible data pipelines, filtering procedures, intermediate checkpoints, evaluation methods, model cards, datasheets and licences that allow meaningful reuse. Without these elements, researchers and public authorities may be able to run a model, but they cannot fully understand how it was built or reproduce its behaviour under public scrutiny.

This distinction matters especially for Europe. Europe’s advantage is not to copy the American model of closed hyperscale platforms, nor the Chinese model of state-controlled technological systems. Its distinctive strength lies in treating AI as a public-interest infrastructure embedded in democratic institutions, open science, legal accountability, multilingualism and interoperability. A fully open European frontier model, aligned with the AI Act and designed for all European languages from the start, would turn these values into technical architecture.

Such a model should not be seen only as a very large central system. Its deeper value is that it can become the source for thousands of smaller, cheaper and more efficient local models. Through knowledge distillation, fine-tuning, synthetic data generation and domain adaptation, a European frontier model could support models adapted to municipalities, regions, universities, hospitals, schools, cooperatives, cultural institutions and SMEs. This is how AI becomes a productive tool across local economies, rather than a service rented from a few global platforms.

RISC-V as the Second Foundation of European Technological Leadership

Europe should not stop at open models. If it wants real technological leadership, it must connect open AI with RISC-V and open hardware. Dependence on proprietary AI processors, especially from Nvidia and AMD, creates a new form of infrastructure lock-in. Even when the model is open, its training and inference can remain tied to closed architectures, restricted supply chains, export controls and expensive software ecosystems.

RISC-V offers a different path. It is an open, modular instruction set architecture that allows universities, companies and public consortia to design specialised processors and accelerators without paying ISA royalties. This does not mean that RISC-V will replace high-end Nvidia GPUs in frontier-model training tomorrow. It means that RISC-V can become the AI-native platform for custom, energy-efficient accelerators, especially in inference, edge AI, automotive systems, industrial computing, public-sector infrastructure and sovereign data centres.

For this to happen, several conditions must be met. First, RISC-V Vector extensions and emerging Matrix extensions must mature and become widely adopted. Modern AI accelerators depend heavily on matrix multiplication, BF16, FP8 and FP4 operations. The strength of RISC-V is that it allows custom instructions, enabling European companies and research labs to implement AI-specific instructions directly in the ISA without proprietary licensing constraints. This can improve power, performance and area, which are decisive metrics for both edge devices and data centres.

Second, RISC-V needs a mature software stack. Nvidia’s strongest moat is not only hardware, but CUDA. If RISC-V is to become a serious alternative, it must offer a developer experience that works out of the box: LLVM, GCC, the Linux kernel, PyTorch, TensorFlow, ONNX Runtime, Hugging Face Transformers and ML compilers must run reliably and efficiently. Initiatives such as RISE are therefore strategic. They aim to turn RISC-V from an elegant open architecture into a commercially and scientifically usable software ecosystem.

Third, RISC-V must be connected to chiplets, UCIe, HBM3e and HBM4. Large AI workloads are not determined only by processor cores. They are determined by memory bandwidth, chiplet interconnects, node-to-node communication, advanced packaging and scale-out architecture. Europe must invest in open chiplet interconnects, high-bandwidth memory, advanced packaging and energy-efficient clusters. Only then can RISC-V move from edge devices and controllers into serious data-centre AI accelerators.

Inference First, Training Later

The most realistic path for RISC-V is not to compete immediately with Nvidia at the very top of AI training. That market is deeply shaped by mature GPUs, high-bandwidth memory, advanced packaging, CUDA and large-scale cluster interconnects. The more viable path is AI inference, where energy efficiency, total cost of ownership, customisation and deployment locality matter enormously.

Inference is where Europe’s public and local AI strategy lives. Municipal assistants, legal RAG systems, healthcare triage tools, educational tutors, industrial monitoring systems, agricultural advisory systems, cultural heritage tools, climate-risk dashboards and public-service copilots do not all require frontier-scale training infrastructure. They require secure, efficient, auditable inference in local or sovereign environments. RISC-V can become highly competitive in these settings, especially when combined with open models, quantisation, domain-specific accelerators and local data governance.

This is why the European AI strategy must be layered. Large frontier training can run on EuroHPC systems and AI Gigafactories. Mid-sized models can run in sovereign data centres. Smaller models can run on regional infrastructure, municipal clusters, university labs, hospitals and edge devices. RISC-V open hardware can strengthen this entire stack by reducing dependence on proprietary processors and enabling European industrial specialisation.

From Open AI to an Open Computing Ecosystem

The European plan must be unified: open models, open data, open software, open hardware and open procurement. Public money should not produce new private lock-in. It should create public knowledge, reusable code, documented data pipelines, shared compute capacity and industrial autonomy.

A European frontier model would provide the cognitive foundation. RISC-V would provide the hardware foundation. AI Factories, EuroHPC systems, universities, regional data centres, public administrations and SMEs could become the productive network. On top of this, Europe could build public APIs, local LLMs, educational tools, legal RAG systems, cultural models, agricultural systems, energy optimisation tools, industrial copilots and public-sector AI services.

This is the political core of the proposal. AI can concentrate power in a few closed systems, or it can become a shared infrastructure. The choice is institutional, not merely technical. Europe has the universities, research centres, companies, open-source communities and democratic framework to choose the second path. What it needs is a clear mandate: not scattered projects, but a European strategy for open AI and open hardware.

If Europe succeeds, it will not simply own another model or another chip design. It will have created a democratic computing stack for the twenty-first century.

Sources for the article:

  1. European Commission, AI Continent Action Plan: The Commission’s AI continent framework documents InvestAI, AI Factories and AI Gigafactories as core elements of Europe’s effort to scale AI capability and industrial competitiveness: https://commission.europa.eu/topics/competitiveness/ai-continent_en,
  2. European Commission, AI Gigafactories: The official page documents the mobilisation of EUR 20 billion for AI Gigafactories, designed as large-scale infrastructure for trustworthy AI development: https://commission.europa.eu/topics/competitiveness/competitiveness-coordination-tool-projects/ai-gigafactories_en,
  3. European Commission, EUROPA Frontier AI Grand Challenge: The announcement of the EUROPA consortium supports the argument that Europe is moving toward frontier-scale model development above 400 billion parameters: https://digital-strategy.ec.europa.eu/en/news/commission-selects-europa-consortium-winner-frontier-ai-grande-challenge-project-build-european,
  4. EuroHPC Joint Undertaking, JUPITER: The official EuroHPC announcement confirms JUPITER’s role in bringing Europe into the exascale era, providing a critical compute foundation for science and AI: https://www.eurohpc-ju.europa.eu/jupiter-officially-propels-europe-exascale-era-2025-11-17_en,
  5. OpenEuroLLM: The project aims to develop strong, transparent and multilingual European foundation models covering EU official languages and supporting AI Act-aligned evaluation: https://openeurollm.eu/,
  6. Allen Institute for AI, OLMo 2: Ai2’s documentation defines fully open models as those released with weights, training data, code and evaluation, making it a central reference for genuine openness in AI: https://allenai.org/blog/olmo2
  7. ETH Zurich, EPFL and CSCS, Apertus: Apertus is a fully open, transparent and multilingual large language model, offering a concrete example of sovereign and inspectable AI infrastructure: https://ethz.ch/en/news-and-events/eth-news/news/2025/09/press-release-apertus-a-fully-open-transparent-multilingual-language-model.html
  8. Mistral AI, Mistral 7B: Mistral 7B demonstrated that efficient open models released under Apache 2.0 can deliver high performance at comparatively lower compute cost: https://mistral.ai/news/announcing-mistral-7b/
  9. EuroHPC Joint Undertaking, Advancing European Sovereignty in HPC with RISC-V: The DARE initiative documents Europe’s plan to design a RISC-V general-purpose processor, a vector accelerator and an AI Processing Unit for HPC and AI workloads: https://www.eurohpc-ju.europa.eu/advancing-european-sovereignty-hpc-risc-v-2025-03-06_en,
  10. Linux Foundation Europe, RISE RISC-V Software Ecosystem: RISE is a key initiative for maturing the RISC-V software stack, supporting open-source tools such as LLVM and GCC so that RISC-V can become a commercially usable platform: https://linuxfoundation.eu/newsroom/rise-project-launches-to-accelerate-development-of-risc-v.