The local processing of data by artificial intelligence is redefining the architecture of consumer and professional devices in 2024. Processors now integrate neural processing units (NPU) as a priority component, and this shift towards on-device embedded AI profoundly changes the management of latency, privacy, and energy consumption across the entire hardware chain.
Post-quantum encryption: why companies are migrating before the threat
Post-quantum cryptography is no longer a laboratory topic. Companies in sensitive sectors (finance, defense, health) have begun to inventory their encryption algorithms to prepare for migration, even though no quantum computer is yet operational in traditional IT services.
The stakes are asymmetrical: data encrypted today can be decrypted tomorrow if intercepted and stored while waiting for quantum maturity. This scenario, often referred to as “harvest now, decrypt later,” drives technical leaders to act without waiting for the threat to materialize.
We observe that early feedback on Sarkostique’s tech page confirms this trend across several European industrial sectors. The main challenge remains the exhaustive identification of cryptographic dependencies in heterogeneous information systems, often built over several decades.
Industrial edge computing: predictive maintenance and real-time video analysis
Edge computing is not progressing uniformly. The most mature use cases in 2024 are predictive maintenance on industrial sites and real-time video analysis, two areas where network latency to the cloud is prohibitive.
On a production line, a vibration sensor must trigger an alert within milliseconds. Sending this data to a remote data center introduces a delay incompatible with the preventive shutdown of a machine. Local processing, directly on the edge node, eliminates this constraint.
Interoperability and attack surface
The dispersion of computing nodes creates two simultaneous problems. Interoperability between edge platforms remains a major challenge: each provider pushes its own execution environment, complicating integration in a multi-vendor landscape.
The multiplication of computing points mechanically increases the attack surface. Each node deployed at the network edge becomes a potential target, and cybersecurity teams must adapt their monitoring policies to a distributed architecture that is much more complex than a centralized cloud.
European AI Act: concrete obligations for generative model providers
The European regulation on artificial intelligence mandates a classification of systems by risk levels. For providers of generative AI models, the obligations become tangible:
- Production of detailed technical documentation describing the architecture and capabilities of the model
- Implementation of a copyright compliance policy applicable to training content
- Publication of a sufficiently detailed summary of the data used for training the model
These requirements apply progressively according to a staggered timeline, allowing companies a variable compliance timeframe depending on the risk category of their applications. General-purpose models (foundation models) are directly affected, regardless of the final application sector.
Impact on corporate adoption strategies
For IT departments, the AI Act transforms the choice of an AI provider into a legal as well as technical decision. A model deployed without compliant documentation exposes the user company, not just the provider.
We recommend integrating regulatory compliance verification early in the qualification phase of generative AI tools, rather than downstream of deployment. Companies that wait for full implementation risk discovering structural incompatibilities with their business processes.
NPU and on-device AI: what the processor architecture changes in 2024
Qualcomm and MediaTek have designed their 2024 processors as “NPU-first” architectures, where the neural processing unit is no longer a secondary coprocessor but the component around which the rest of the SoC is organized. Real-time translation, image generation, and interface personalization are executed locally.
Google has applied the same logic with Gemini Nano on the Pixel 8 Pro: drafting contextual responses, automatic summaries, and intent detection without systematic transit to the cloud. Apple announced “Apple Intelligence,” a set of generative AI functions combining on-device processing and occasional server calls.
Consequences on energy consumption
Local processing reduces the volume of data exchanged with data centers, which decreases network consumption. However, thermal dissipation on the device increases during heavy inferences. The trade-off between local performance and battery life becomes a central design parameter for smartphone and laptop manufacturers.
- Light tasks (text suggestions, photo sorting) run continuously on the NPU with marginal energy cost
- Heavy tasks (image generation, full language models) push the NPU to full load and significantly reduce battery life
- Some manufacturers implement dynamic routing that switches to the cloud when the battery falls below a critical threshold
The year 2024 marks a technical turning point more than a visible break for the end user. The technological building blocks, from post-quantum encryption to embedded AI, are deployed in layers of infrastructure that the general public does not directly perceive. Yet it is here that the competitiveness of companies in the sector will play out over the coming years.
