D7.1 Survey of customer needs and HPC/AI maturity

1 Survey of Customer Needs and HPC/AI Maturity September 2025 Conducted by: LUMI AI Factory, 2025 License: CC BY-SA 4.0, refer to licensor “LUMI AI Factory” Disclaimer: This report is for informative purposes only. Contact: [email protected] 2 Glossary of Terms Item Description AI Artificial Intelligence DevOps Development Operations DL Deep Learning GDPR General Data Protection Regulation GPU Graphic Processing Unit HEI Higher Education Institutes HW Hardware HPC High-Performance Computing Hugging Face AI model hub & library IoT Internet of Things IP Intellectual Property LUMI AIF LUMI AI Factory LLM Large Language Models ML Machine Learning NLP Natural Language Processing PyTorch Flexible deep learning framework RL Reinforcement Learning ROI Return on Investment SME Small and Medium-Sized Enterprises TensorFlow Scalable machine learning platform Typeform Tool for conduction online surveys VR Visual Recognition WP Work Package 3 Table of Contents 1. Executive Summary ...................................................................................................... 4 2. Responses .................................................................................................................... 5 2.1 SECTION I: Organisation Profile 5 2.2 SECTION II. Current Use and Maturity 8 2.3 SECTION III. Technical Needs and Data Management 14 2.4 SECTION IV. Skills and Support Needs 19 2.5 SECTION V. Collaboration and Outlook 22 3. Conclusion ................................................................................................................. 28 3.1 Start-Ups and SMEs 28 3.2 Large Enterprises and Research Institutions 29 3.3 Cross-Cutting Needs Highlighted in Open Responses 29 4 1. Executive Summary A customer needs and maturity survey was designed and launched to collect data on the needs of current, previous, and potential customers of the LUMI AI Factory (LUMI AIF) services. The data to be collected involved current state of High-Performance Computing (HPC) and Artificial Intelligence (AI) adoption, technical needs, barriers, and collaboration outlook. From mid-June to mid-August 2025, the survey was distributed among the LUMI AIF consortium countries Finland, Norway, Denmark, Estonia, Poland, and the Czech Republic, resulting in 40 responses, incl. responses from these and other European countries. The respondents represented a broad mix of start-ups (30.8%), SMEs (30.8%), higher education institutions (HEI) (15.4%), large enterprises (15.4%), and research institutes (7.7%), providing a snapshot of current practice and immediate operational requirements in the rapidly evolving HPC/AI landscape. The content is structured around survey design, respondents’ organizational and technical profiles, and their day-to-day experience with AI and HPC systems. Key outcomes include strong demand for scalable GPU access, unified infrastructure supporting hybrid and cloud workloads, technical onboarding, and specialist support for integrating new hardware or platforms. Access to curated European data resources, straightforward funding pathways, and ongoing workforce training emerged as distinct needs, particularly from organizations scaling beyond pilot projects. Key findings reveal that AI adoption among the participants is highly mature, with 92.3% of organizations already using AI or machine learning (ML) technologies, compared to 53.8% using HPC. Moreover, 28.2% of organizations plan to introduce HPC soon, demonstrating a growing interest. Organizations selfassess their maturity levels as predominantly advanced (50%) or intermediate (35%), while only 15% identify themselves as beginners. The main drivers of adoption include efficiency gains (72.5%), innovation (45%), and cost reduction. Several barriers constrain broader integration, most notably budget constraints (65%), followed by data privacy issues (25%) and unclear business value (20%). In terms of technical needs, GPU computing stands out as the most critical requirement (79.5%), alongside AI inference services, data cleaning, and fast storage. Computational tasks focus on AI model training (85%), simulations and digital twins (47.5%), and large-scale data analysis (37.5%). Scalability, fast results, and reliability are the leading performance requirements. Skills and workforce readiness show promising trends: 69.2% of organizations already employ dedicated AI/HPC staff, and 61.5% are actively investing in workforce upskilling. Nonetheless, gaps remain in HPC operations (57.1%), AI/ML expertise (40%), cloud infrastructure (40%), and data science (37.1%). Collaboration is emerging but uneven. While 53.8% of organizations partner with external actors, awareness and use of competence centres and EU-funded HPC programmes remain limited. Only 33.3% of respondents reported participation in publicly funded programmes, while 41% had not engaged in any. Organizations highlight top priorities in infrastructure access (cloud, GPUs, scalable HPC), large model training and fine-tuning, commercialization of AI-based products, and domain-specific applications (e.g., agriculture, manufacturing, translation). Strategic goals include workforce development, demonstrating ROI, and making societal contributions, such as addressing labour market challenges. The evidence highlights the importance of targeted infrastructure investments, community-focused training, and flexible, partnership-oriented approaches to ensure broad and sustainable AI/HPC adoption across Europe’s research and industrial sectors. 5 2. Responses The following section provides the responses of the survey participants. 2.1 SECTION I: Organisation Profile a. Name of Organisation Confidential b. What is the type of your organization? The 40 organizations represented in the survey were primarily start-ups (30.8%) and SMEs (30.8%), followed by higher education institutes (15.4%) and large enterprises (15.4%). A smaller share was represented by research institutes (7.7%); no respondents from public administration took part. 6 c. What is your organization’s main business sector? In terms of the organization’s main business sector, the domain of AI prevailed (42.1%), followed by research and development (13.2%) and manufacturing (10.5%). Other domains, such as software development, engineering, manufacturing, energy, life sciences, and space applications, were represented by either fewer or single respondents; other domains by none. Other: • Quantum computing • Competitive & market intelligence • Stainless steel hose production • Languages and audiovisual translation • Healthcare 7 d. How many employees does your organization have? As for the participants’ size and structure, just over half of the participants represented small organizations with fewer than 50 employees (55%), while large organizations with more than 1,000 employees accounted for 22.5%. Categories of organizations with 50-250 employees (12.5%) and 2511,000 employees (10%) achieved a lower participation. e. In which country is your organization located? The response rate per country is laid out in the following table. Finnish participants are represented with the highest response rate (35.0%), followed by Czechia (22.5%), Denmark (17.5%), and Poland 12.5%. One response came from Norway, and no response came from Estonia. Answers from non-consortium countries were received from Belgium, Turkey, Sweden, and Germany. Country Number of Participants Percentage Finland 14 35.0% Czechia 9 22.5% Denmark 7 17.5% Poland 5 12.5% Norway 1 2.5% Sweden 1 2.5% Germany 1 2.5% Turkey 1 2.5% Belgium 1 2.5% Estonia 0 0.0% Total 40 f. In which other countries does your organization operate? In 72.5 % of cases, participating organizations noted that they had operations abroad, with Germany, France, Spain, Poland, Slovakia, Hungary, Czechia, Switzerland, Sweden, Norway, Denmark, and the UK being the most common ones. Respondents often grouped these under EU or Europe, sometimes listing specific Central & Eastern European countries alongside Western Europe. Apart from that, North America and the Middle East were also mentioned, indicating a global reach of the companies’ activities. 8 2.2 SECTION II. Current Use and Maturity a. Does your organization currently use HPC resources? This question revealed that 53.8% of participating organizations have already implemented HPC into their operations, with an additional 28.2% planning to do so. Only 17.9% of respondents stated that they do not use HPC at the current state. These statistics show a certain level of experience among the HPC community members and willingness to explore this field. If your organization currently uses or plans to use HPC resources, please specify the use case. • Time series forecast using support vector regression • Pushing the limits of quantum supremacy • Medical image processing • We mainly utilizes High-Performance Computing (HPC) resources to support the development and validation of advanced technologies for vehicles (but not only), particularly in the context of softwaredefined vehicles like: Advanced Driver-Assistance Systems (ADAS) development, Software-defined vehicle development, AI and Machine Learning integration, Testing and Validation, Sensor Fusion, Scalability, and Interchangeability. • LLM inferencing • Development of marketing technology using AI • Building predictive models, using DL • We use HPC clusters (IT4Innovations, EuroHPC) for GPU-accelerated AI crop phenotyping, CFD of vertical farms, massive parametric LED/watering optimization, and LCA simulations; plans genomescale analysis next. • Planning in production • Training foundation models • heavy analysis and model building • LLM, AI agent, data analytics • Can't tell about it; it's a secret • Large Language Models • Various research projects mainly focused on LLMs and physical simulations 9 • Training AI & running AI models • Innovation for demanding computations for client cases and in-house study. • We operate HPC resources to provide our services to our customers. • AI training, evaluation, and development • We're looking for ways to support translators' expertise • We have our own IT department, and we run multiple applications for internal purposes • AI/ML model training • Training LLMs, Fine-tuning GenAI, GRPO training • Model tuning, model training, reinforcement learning • Currently, we are mostly finetuning GRPO, LLMs on a smaller number of our own GPU cards, as we keep scaling, our own GPUs will not suffice • Specialized model training, running open source LLM • CFD • HPC for research and education • For research • Computational chemistry; Material Science; Bioinformatics, etc. b. Does your organization use AI or machine learning technologies? The utilization of AI and machine learning reported by the participants revealed a whopping share of 92.3% of companies that already have engaged with AI or machine learning in their operations, and only 7.7 % that have not. These statistics confirm the maturity of AI as compared to HPC. If your organization currently uses or plans to use AI or machine learning technologies, please specify the use case. • Time series forecast using support vector regression • Image interpretation • Already answered the questions before, mainly for ADAS • Personalization and Learning from User Data. The app “learns” user preferences: -which pictograms are chosen most often, -the typical vocabulary of the family/caregiver, -the child’s specific communication habits. 16 c. What computational tasks are relevant to your organization? (select up to three) The large majority of organizations regarded the training of AI models as a relevant computational task (85.0%). Tasks, such as various simulations and digital twins (47.5%) and large-scale business data analysis (37.5%) or IoT and sensor data (35.0%) followed. Other: • Business problem optimization; search for classical limits on quantum supremacy • marketing use cases, real-time graphic generation based on consumer data • Voice-to-voice • Running LLM to achieve as much independence as on big LLMs 17 d. What are your performance requirements? (Select up to three) Performance requirements, according to the survey, are ranked in terms of importance as follows: scalability (65.0%), fast results (62.5%), reliability (52.5%), high throughput (42.5%), and data security (37.5%). Other: • Speed - response time • User-friendliness • Support for modern AI software stack (Pytorch 2.7, flash-attention2, triton) 2. ease/unified deployment (k8s jobs instead of slurm) 3. GPUs with 80+GB of VRAM to fit 32k context size or larger during model training (there is a lower bound to sharding the model layers). 4 Hardware FP8 support 18 e. How often do you expect to use HPC/AI resources? The survey concludes that the majority of organisations engaged with HPC and AI expect to use these resources and tools daily (66.7%), and a further 12.8% of organizations intend to do so weekly. f. What are your main concerns regarding data security? (Select up to three) The most common concern regarding data security was stated to be sensitive data protection (62.5%). Secure data storage and transfer was mentioned by 50 % of respondents and must be considered a relevant concern. Compliance risk (35.0%) and cloud service transparency (32.5%) can be ranked secondary concerns, while IP confidentiality is a tertiary concern. No specific concerns were stated by 5% of respondents. 19 2.4 SECTION IV. Skills and Support Needs a. Do you have dedicated AI/ML or HPC personnel? Almost three-quarters (69.2%) of the organizations already have a dedicated AI/ML or HPC personnel, and another 17.9% of organizations intend to integrate such personnel. That shows a surprisingly good level of readiness and maturity. More detailed analysis shows that all research institutions have dedicated personnel. All start-ups either have or are planning to integrate dedicated personnel. While SMEs show a strong dedication to integrating personnel, large enterprises show a lower commitment. 20 b. What skills are lacking most in your organization? Organizations have reported that several skills are missing, which may limit further HPC and AI adoption. The most serious gap is seen in the field of HPC operations (57.1%), followed by AI and ML expertise (40.0%), cloud infrastructure 40.0%, and data science (37.1%). Other: • When working with research institutions or corporate partners, the biggest skills gap we see is in deploying our hardware: partner IT teams often lack the network and security know-how to open the right ports or configure VPNs, the dev-ops experience to register applications and manage API keys or certificates, and the database skills. From experience, many other IoT SMEs have big trouble here. • DevOps engineering, web development, to bring a wider array of cloud services to our customers. • High-quality sales c. Are you investing in workforce upskilling for AI/HPC? Most organizations reported they were investing in workforce upskilling for AI and HPC (61.5%), illustrating a dedication to engage even further with AI and HPC. Only 17.9% of organizations stated they made no such investments, while the remaining 20.5% of organizations intend to invest. 21 d. What type of external support would be most valuable to your organisation? (Select up to three) Respondents claimed that funding support would be most valued (55.0%), followed by HPC infrastructure access (52.5%) and technical consulting (42.5%). Funding support was particularly mentioned by SMEs, start-ups, and HEIs, but not by large enterprises and research organisations. Large enterprises are particularly in need of technical consulting. For Research Institutes, access to HPC infrastructure is most valuable. Skills training and regulatory guidance were mentioned by just under one-third of the organizations. 22 2.5 SECTION V. Collaboration and Outlook a. Are you collaborating with external partners on AI/HPC? Over half of the organizations (53.8%) reported collaborating with external partners on AI and HPC, indicating that awareness of competence centres and other external actors supporting the integration of these technologies remains limited and should be further strengthened. If you selected 'Yes' please specify the mode of collaboration with external partners on AI/HPC. • Deucalion • Collaboration with universities • MS Azure • Getting resources (inluding GPU/AI) with L2/L3 support • Cyfronet AGH • IT Dev, building solutions for our needs • AI startups • HPC access • Involvement of various parties during research projects • Google for Startups Partner. • We tightly cooperate with our largest customers to help them be successful in training and running AI models at scale. • Projects • Open research • We are looking into this with a broadcasting company. We are their preferred translation provider. • Consulting, human resources • We're collaborating with Czech universities and their supercomputer centers (e.g., CEITEC) • European projects • Nvidia (partner, GPUs,...), CVUT (access to Ostrava cluster) • Research • Research consortia 23 b. What are your goals for adopting AI/HPC? (Select up to 3) The motivation for the AI and HPC adoption is clear; the majority of organizations seek an increase in efficiency (72.5%). Increasing efficiency is the most important goal for HEIs, SME’s and start-ups. The second most recurrent goal stated is innovation (45.0%), particularly driven by HEIs and Research organisations. At least one quarter of companies also highlighted cost reduction, the launch of new services, and customer service improvement. Large enterprises particularly stress the aim of cost reduction and an increase in customer service. Other: • Increase the quality of the translations • better conditions than buying GPUs or provisioning GPU IaaS on big tech clouds 24 c. What is your time horizon for your next AI/HPC project? The prevailing long-term horizon (75.0%) for the organizations’ next AI or HPC project signals that they adopt more of a strategic decision-making approach. Only 5.0 % state AI/HPC is a one-off project. d. Have you participated in any publicly funded AI/HPC programs? The share of organizations that have not participated in any publicly funded AI/HPC programmes (41.0%) is higher than that of the ones that have (33.3%). That confirms some lack of awareness of such programmes among the organizations potentially interested in AI and HPC deployment. The high proportion of respondents stating “not sure” may be reflected in difficulties understanding the question. If you selected 'Yes' please specify which publicly funded AI/HPC programs you have participated in. • EuroHPC Deucalion • Cyfronet grant for startups • PLGrid • EDIH 25 • LUMI test period • Danish national funded • Norway's Olivia pilot testing, NTNU's IDUN events • EuroHPC grant for Lumi-G • HORIZON • EDIH Ostrava • Research Council of Finland, EuroHPC • PRACE, EuroHPC, LUMI.... e. What are your top priorities for AI/HPC in the next 12–24 months? The collected statements cluster into six main themes: infrastructure and compute needs (cloud, GPUs, HPC integration, debugging issues), LLM and model training (scaling, fine-tuning, inferencing, project completion), and productization and commercialization (AI-based products, domain-specific systems, startups, distributor roles). Additional themes include research and workforce development (trained personnel, researcher involvement, scalable architectures), domain-specific applications (agriculture, manufacturing, translation, optimization), and strategic goals (planning, ROI, piloting, societal impact, such as labour market challenges). Together, these codes reveal a balance between technical requirements, applied innovation, and broader economic or societal objectives. • Launching new AI-based products & services, enhancing existing AI applications, workforce upskilling • Involvement of researchers and leveraging the full potential of HPC • Move more workloads to GPU resources, increase the adopting AI workloads to HPC • Model training, model fine-tuning, model specialization • Train LLM model, launch an online service based on your custom own LLM • To get an overall understanding of the possible benefits vs. costs vs. resources needed • Building our new martech products based on AI • Infrastructure competence • Scaling computations, model training/finetuning, ML operations • Help companies in Europe with our sovereign cloud services for AI • Integration of AI/HPC in public cloud • More GPUs • Building a scalable and cost-efficient pipeline. Due to using multiple pipelines for multiple use-cases, the overarching architecture is the most important topic. • Train and infer the high-throughput large vision model for my brand-new vision foundation model, which I am currently developing. • Use large resources for demanding computation and optimization tasks. This way, we can provide the best solutions for clients and enrich the innovation for the company's own products, which will bring long-term value. (However, we cannot really invest a lot in this other than time, and Business Finland only funds larger companies doing random useless "EU projects" that they, or nobody, needs, etc. Just my experience.) • Anything how to utilize MI250 machines best way possible. The AI libraries develop too fast to keep up with such old hardware. How to keep the latest libraries and programmes listed. How to debug runs in LUMI, while using sbatch scripts and Python, many error messages are not clear at all. • Solve Finland’s labour market problems