AI in the News – MINERVA Newsletter #3

Scaling AI in the Era of Scarcity: Overcoming the Global “Memory Wall”

As the global AI landscape evolves at breakneck speed, a new structural constraint has emerged that threatens to reshape the trajectory of research and deployment: the critical scarcity of memory and storage components. For the European AI community, navigating this “memory wall” is not merely a logistical challenge but a strategic imperative for maintaining technological sovereignty.

The Crisis at a Glance: 300% Price Surges

The current shortage, centred on High-Bandwidth Memory (HBM), differs fundamentally from past semiconductor disruptions. Driven by an “AI Demand Supercycle,” hyperscalers are consuming unprecedented volumes of HBM, effectively outbidding academic institutions and smaller enterprises. The impact is stark: contract prices for 16Gb DDR5 chips surged by 300% between September and December 2025.
With HBM production reported to be sold out through late 2027, research institutions are facing 6-12 month lead times for GPU clusters. This has created a “scientific digital divide,” where hardware now constitutes up to 67% of frontier model development costs, forcing many labs to choose between reducing research staff or delaying computationally intensive projects.

Europe’s Strategic Hurdle

The shortage hits Europe particularly hard due to deep-seated structural vulnerabilities. Currently, over 90% of AI accelerators deployed in European data centers are sourced from non-EU suppliers. While the European Chips Act and the deployment of AI (Giga)factories represent ambitious steps forward, these facilities face the same global supply constraints and high market prices.
European researchers often lack the privileged supplier relationships enjoyed by US hyperscalers or the geographic proximity to manufacturing hubs in Asia. This results in higher effective costs and longer lead times, potentially relegating European labs to secondary roles in frontier AI development if not addressed decisively.

From Scale to Efficiency: A New Research Paradigm

In response to these constraints, we are witnessing a pivot in research methodology. The era of “brute force” scaling is giving way to a focus on:

• Algorithmic Innovation: Developing memory-efficient algorithms, sparse models, and mixture-of-experts architectures.
• Hardware-Aware Design: Integrating hardware constraints as first-order considerations, emphasizing “performance per parameter”.
• Model Optimization: Utilizing quantization and compression techniques to maximize scientific output per compute unit.

The success of models like DeepSeek R1, which achieved competitive results through highly optimized compute usage, serves as a blueprint for this efficiency-focused approach.

How MINERVA Bridges the Gap

As the European Support Centre for Scalable AI, MINERVA is uniquely positioned to help the community translate these challenges into opportunities for innovation. Our services are specifically designed to mitigate the impact of the component shortage:

• Maximizing Infrastructure Uptake (L1 Support): We provide expert assistance in porting AI applications to existing EuroHPC systems like Leonardo and MareNostrum 5, ensuring that current hardware is utilized to its maximum efficiency.
• Training for the “Efficiency Era” (WP5): Through our “Train-the-Trainers” programs, we empower researchers with skills in model quantization and scaling workloads, bridging the HPC knowledge gap essential for resource-constrained research.
• Fostering Open-Source Sovereignty: By prioritizing the R&D of open-source foundation models, MINERVA encourages the reuse of high-quality models and datasets, reducing the need for every organization to train massive models from scratch.
• Operationalizing Multi-Client Access: We aggregate demand from SMEs and startups, ensuring that even as hardware prices rise, smaller innovators can access the large-scale resources needed to stay competitive.

Conclusion: A Resilient Path Forward

The hardware shortage is a test of resilience for the European AI ecosystem. By focusing on competence, ethical alignment, and software optimization, we can ensure that our strategic investments in AI Factories do not become “monuments in the desert,” but rather the foundations of a functional, globally competitive ecosystem. At MINERVA, we remain committed to ensuring that European researchers have the support and skills to lead, regardless of the height of the “memory wall”.

References

• Intuition Labs. (2026, January 29). RAM Shortage 2025: How AI Demand is Raising DRAM Prices. https://intuitionlabs.ai/articles/ram-shortage-2025-ai-demand
• BISI. (2026, January 31). Global RAM Shortage and Price Hikes: Causes, Consequences, and Market Outlook. https://bisi.org.uk/reports/global-ram-shortage-and-price-hikes-causes-consequences-and-market-outlook
• E-Spin Corporation. (2025, December 18). Global Memory-Chip Shortage: AI Demand and Industry Impact. https://www.e-spincorp.com/global-memory-chip-shortage/
• BBC News. (2026, January 1). Why everything from your phone to your PC may get pricier. https://www.bbc.com/news/articles/c1dzdndzlxqo
• TBS News. (2025, December 2). The AI frenzy is driving a memory chip supply crisis. https://www.tbsnews.net/world/global-economy/ai-frenzy-driving-memory-chip-supply-crisis-1300721
• European Strategy and Policy Analysis System. (2025, November 24). Semiconductors as Key Strategic Assets: Navigating Global and European Security Challenges. https://esthinktank.com/2025/11/25/semiconductors-as-key-strategic-assets-navigating-global-and-european-security-challenges/
• Patent PC. (2026, January 25). AI Model Training Costs: How Much Big Tech Spends on AI Development – Latest Stats. https://patentpc.com/blog/ai-model-training-costs-how-much-big-tech-spends-on-ai-development-latest-stats
• Enki AI. (2026, January 27). AI Power Wall 2026: Why Grid Access Halts Chip Demand. https://enkiai.com/ai-market-intelligence/ai-power-wall-2026-why-grid-access-halts-chip-demand
• Epoch AI. (2024, June 2). How much does it cost to train frontier AI models? https://epoch.ai/blog/how-much-does-it-cost-to-train-frontier-ai-models
• CNBC. (2025, January 31). DeepSeek’s hardware spend could be as high as $500 million, report says. https://www.cnbc.com/2025/01/31/deepseeks-hardware-spend-could-be-as-high-as-500-million-report.html
• European Commission Joint Research Centre. (2025). Infrastructures for AI in Science. https://euagenda.eu/publications/download/622781
• Defence24. (2025, October 26). Trapped in the Chip War: Europe’s AI ambitions at risk. https://defence24.com/technology/trapped-in-the-chip-war-europes-ai-ambitions-at-risk
• LinkedIn. (2025, December 2). Europe is falling behind: why AI, HPC and sustainability must redefine data center strategy. https://www.linkedin.com/pulse/europe-falling-behind-why-ai-hpc-sustainability-must-redefine-steman-4gizf
• Basis Point Insight. (2026, January 30). What the Union Budget 2026 Can—and Should Not—Do for AI. http://basispointinsight.com/Story/what-the-union-budget-2026-can-and-should-not-do-for-ai_c7530b975877.html
• Asian News Network. (2026, February 2). Korea’s memory fortress under pressure as China advances, but the lead still holds. https://asianews.network/koreas-memory-fortress-under-pressure-as-china-advances-but-the-lead-still-holds/
• CNBC. (2026, January 27). Micron to invest $24 billion in Singapore plant as AI boom drives memory shortages. https://www.cnbc.com/2026/01/27/micron-24-billion-singapore-nand-dram-expansion-ai-memory-shortages.html
• European Commission Joint Research Centre. (2014, October). The hidden costs of R&D collaboration. https://joint-research-centre.ec.europa.eu/system/files/2014-10/JRC90991.pdf
• WEKA. (2025, October 20). The Impact of Storage on the AI Lifecycle. https://www.weka.io/resources/analyst-report/the-impact-of-storage-on-the-ai-lifecycle/
• lakeFS. (2024, November 18). Reduce Storage Costs: Effective Cost Savings for Data Lakes. https://lakefs.io/blog/storage-costs/