Alongside the current protests and heavy riots in some European capitals, another topic which is building up steam in Europe is the modest size of its compute. What is the meaning of the term "compute"? This is not merely processing power; it encompasses the entire ecosystem that enables AI algorithms to learn from data, identify patterns, and ultimately make intelligent decisions. In AI jargon, it refers to the hardware resources required to run artificial intelligence workloads, including training machine learning models and processing data. This is typically measured in terms of floating-point operations per second and is essential for the efficiency and effectiveness of AI models.
Therefore, the term "compute" refers to the allocation and utilisation of computational resources such as software, hardware, and infrastructure necessary to execute the complex mathematical operations required for training, validating, and deploying AI models. These operations range from basic matrix multiplications to intricate gradient calculations within deep neural networks.
Europe has produced very few models reaching the highest compute scales. The levels of compute contribute to weaker private investment, fewer frontier labs, and a risk of talent and capital flowing elsewhere. European officials and bodies, including ECB leadership, have flagged the infrastructure shortfall as a competitiveness issue. The consensus across recent policy analyses is that Europe's compute position is a real and pressing shortfall. Without faster private investment, grid upgrades, and permitting reform, the relative gap is unlikely to close meaningfully, leaving Europe more dependent and less competitive in the highest-value parts of the AI stack.
Europe is responding with public funding (for example, the €20 billion InvestAI/AI gigafactories initiative aiming to co-finance large facilities) and its efforts to streamline rules. This is a concern when discussing how AI's future measures up in Europe, as it also depends on having mammoth data centres. So it makes commercial sense that trillions of dollars have been invested in the US, producing dense, high-end AI clusters needed for frontier model training. These are heavily concentrated in the US, with China building aggressively from a lower base. Large-scale compute is the main bottleneck for training the most capable AI models. Most of the data-centre projects already underway were approved and started before the backlash reached its present fever pitch.
What is happening with the adverse sentiment against such US data centres? There is a torrent of angry concerns about the impact these centres have on their neighbourhoods and the environment. In America, there hovers a persistent myth that AI data centres consume vast quantities of water and energy. In the medium-term, existing European data centres and cloud access mean researchers and companies must use remote US compute. To its merit, Europe retains strengths in regulation/governance, research talent, and niche industrial capabilities (notably ASML, a Dutch firm specialising in chipmaking equipment). China's compute is also constrained by chip performance gaps under export controls, even as its capacity and power buildout grow quickly.
For instance, based on a white paper published last year, Anthropic argued that as much as 5GW of electrical energy would be required to train a single frontier model. Compute is simply the digital "horsepower" needed to train or run an AI model. To explain in layman's terms, if data is the fuel, algorithms are the engine, while compute is the raw power that makes everything actually move. A typical situation results when operators exclaim that they don't have enough compute to train a particular large model. In simple terms, this means the team lacks sufficient GPUs or processing power. Typically, it applies to huge GPU clusters needed for frontier models. In fact, one must properly balance the model size and dataset size to use compute efficiently.
Readers may wonder how we can gain a clearer understanding of the computing conundrum. A simple analogy is to imagine AI as a kitchen, where data = ingredients; algorithms = recipes; compute = the stove, oven, and appliances that actually cook the food. It stands to reason that without compute, nothing gets cooked.
Europe lacks an adequate drive in AI when compared to the US, but it is hoping to catch up. Compute is not merely processing power; it encompasses the entire ecosystem that enables AI algorithms to learn from data, identify patterns, and ultimately make intelligent decisions. Understanding its intricacies, from the hardware and software layers to the challenges and future trends, becomes crucially essential for anyone involved in developing, deploying, or leveraging AI. A core dependency for AI development, compute, along with data and skilled labour, is one of the three pillars that determine the capabilities of AI systems. Without sufficient compute, even advanced algorithms cannot be trained effectively. Larger language models (LLMs) require exponentially more floating-point operations. As AI continues to evolve, compute will remain at the forefront, enabling new capabilities and driving innovation across industries.
Europe should and must invest more to expand its share of compute. Currently, it hosts only 5% of global AI compute, whereas America has more than 75%. With ambitious reforms, European countries should aim to increase their share to 20% within five years. This means bringing tens of gigawatts of compute onto European soil. That will require dedicated economic zones with a streamlined pattern for data centres and a big increase in energy-generation capacity.
One of the imminent requirements for Europe is to radically reform its tight labour laws. The EU's strength has always rested on its growing market size. But rigid labour rules slow the adoption of new technology, exerting economic drag. Left unaddressed and underfunded, the widening gap, challenged by a more dynamic American economy, can potentially sap Europe's commercial power. In a nutshell, the extra funding may dilute the tax base that funds its vast welfare model.
However, simply throwing money at the problem is not enough to solve Europe's tech industry challenges. European companies face far more red tape when it comes to restructuring, with higher severance costs and longer delays hindering their ability to compete. Full parity with the US is unrealistic given differences in private capital, energy abundance, and regulatory speed; the practical goal is doubling Europe's share of global AI compute toward something more resilient while ensuring reliable local access for critical workloads. While the European system is slower, on the positive side, it excels in caring for its workers.
George M. Mangion is a senior partner at PKF Malta