The year 2024 has reshuffled the cards on several technical fronts simultaneously. Between the implementation of binding regulatory frameworks for generative AI, the rise of sovereign cloud infrastructures, and architectural developments related to quantum computing, companies’ technological choices have become more complex.
European AI Act and concrete obligations for foundation models
The AI Act was formally adopted on March 13, 2024 by the European Parliament. This regulation introduces a specific regime for general-purpose AI systems, a category that includes foundation models and generative AIs like GPT or Mistral.
See also : Discover The Living Web: innovations and trends of today's web
The first obligations will gradually come into effect starting in 2025: technical documentation, transparency towards users, and management of systemic risks for the most powerful models. Specifically, any company deploying an AI system will have to inform the user that they are interacting with a machine.
We observe that this regulatory constraint is already modifying development pipelines. Product teams are now integrating compliance steps from the design phase, which lengthens cycles but reduces legal exposure. To keep up with Web Internet’s tech news on these regulatory topics, monitoring application timelines remains a reflex to adopt.
Related reading : Discover the latest fashion trends to enhance your everyday style
The risk classification (unacceptable, high, limited, minimal) also requires mapping use cases before deployment. An identical model can fall under different risk levels depending on its context of use, making case-by-case analysis essential.
Sovereign cloud and cybersecurity: reconfiguration of the European market
The rise of cloud offerings labeled “sovereign” is not just marketing talk. ANSSI issued its first SecNumCloud qualifications in 2024, establishing a precise technical framework for hosts dealing with sensitive data.
The issue goes beyond the geographical location of servers. The SecNumCloud qualification imposes requirements on immunity to extraterritorial laws (notably the U.S. Cloud Act), encryption, and the operational governance of platforms. Orange and Capgemini jointly announced their Bleu offering in June 2024, positioned in this niche.
Consequences for enterprise architectures
Companies operating in regulated sectors (healthcare, defense, finance) must now arbitrate between the functional richness of American hyperscalers and the compliance of sovereign offerings. The cost of migrating to a sovereign cloud remains the primary barrier, but the legal risk associated with extraterritoriality is pushing legal departments to demand alternatives.
This dynamic has a collateral effect on cybersecurity. Attacks targeting cloud infrastructures are becoming more sophisticated, and “zero trust” security models are becoming a prerequisite rather than an option. We recommend evaluating the certifications of each provider before any multi-year commitment.
Quantum computing: where do real use cases stand
Announcements about quantum computing saturate news feeds, but the operational reality is more nuanced. Current quantum processors operate in noisy conditions that limit the reliability of calculations beyond a small number of logical qubits.
The most advanced use cases in 2024 involve:
- Molecular simulation for pharmaceutical research, where quantum algorithms significantly reduce modeling time compared to classical supercomputers
- Large-scale logistical optimization, particularly for fleet routing in transportation and energy network management
- Post-quantum cryptography, a defensive field: preparing current systems to withstand future decryption capabilities of quantum computers
The quantum threat to current encryption justifies an immediate migration to post-quantum algorithms. NIST finalized its first standards in 2024, and organizations managing long-lived data (medical records, patents, banking data) should anticipate.
Generative AI in production: bottlenecks and real costs
The transition of generative AI from prototype to production has revealed constraints that public demonstrations often overlook. The inference cost of large language models remains the main budget item for companies deploying these technologies at scale.
Training a model is expensive, but it is the continuous inference (each user query) that weighs on operational budgets. RAG (Retrieval-Augmented Generation) architectures allow limiting the size of the model requested by providing it with targeted documentary context, thus reducing GPU resource consumption.
Fine-tuning versus prompt engineering
The choice between fine-tuning a model on business data or relying on advanced prompt engineering depends on the volume of queries and the specificity of the domain. For vertical use cases (legal, medical, industrial), fine-tuning produces more reliable results. For cross-functional uses (writing, summarization, customer support), a well-structured prompt on a generalist model is often sufficient.
Companies that attempted to fine-tune without sufficient training data have experienced performance degradation. The quality of the training dataset directly conditions the relevance of the adjusted model.
- A dataset that is too restricted causes overfitting and rigid responses
- Poorly labeled data introduces biases that the model amplifies in production
- The absence of a human validation pipeline in the training loop compromises long-term reliability
The technological innovations of 2024 are not just a list of promising technologies. They impose architectural, compliance, and budget choices that commit organizations for several years. The AI Act redefines the rules of the game for artificial intelligence in Europe, the sovereign cloud redistributes power dynamics among providers, and quantum computing necessitates rethinking data security today.