The race to build AI infrastructure is hitting new heights. IDC estimates that global spending reached $318 billion in 2025, more than doubling the year before, and the firm has lifted its 2026 forecast to $497 billion. Should we automatically chase the trend? Not necessarily. “Today’s IT architectures were designed to host business applications, store data, and manage transactional flows. AI introduces new requirements in computing, data processing, governance, and model deployment,” notes Benoît Mangeard. Should you rewrite everything? Probably not… The challenge is first and foremost to evolve what already exists.
Data first, infrastructure second
In the face of the AI onslaught, companies feel drawn to starting with technology. A mistake. “Without qualified, governed, and accessible data, even the most powerful infrastructure won’t create value,” cautions the expert. Identifying, classifying, and stabilizing data is precisely the kind of nuanced knowledge of a company’s data capital that will determine subsequent performance, security, and sovereignity requirements. This is a particularly sensitive point because the hosting choice inherently shapes exposure to extraterritorial regulations. For strategic data, a sovereign cloud hosted in France or within Europe, or even on-premises infrastructure, helps preserve control over both data and the models that rely on it.
The challenge of AI industrialization
Many organizations have validated their initial use cases on external APIs. An approach that proves its limits when moving to scale. The consequences, already well known: costs rise inexorably as the volume of requests explodes, data confidentiality becomes harder to maintain, and reliance on suppliers grows. “As AI uses become more strategic and frequent, the question of infrastructure shifts from a purely technical matter toward one of economic control, sovereignty, and the durability of the AI initiative,” summarizes Mangeard.
Overprovisioning or underprovisioning: that is the question…
If an AI architecture is poorly tuned, you pay twice. “An overprovisioned infrastructure incurs unnecessary capital and operating costs, as well as excessive energy consumption. It can even tempt teams to force certain uses to justify the investments,” notes Mangeard. Conversely, an underprovisioned setup degrades the user experience with long response times, downtime, and an inability to absorb load. All of this creates a real risk: user adoption can falter, or even fail.
SLM and Edge AI, the frugal path
Is there a need for ever more GPUs? Not according to Sigma, which advocates frugal solutions built around specialized Small Language Models (SLMs) trained to meet targeted business needs and deployed close to users via Edge AI. The payoff: lower operating costs, faster response times, and a smaller energy footprint. “Where the market often seeks to find uses to monetize already deployed infrastructures, we believe the wiser approach is to invert the logic and start from business needs to build an infrastructure that’s proportionate to the actual usage.”
A matter of method
Sigma’s approach unfolds in four steps. First, cultivate a culture within teams around the different forms of AI, since GenAI, RAG, SLMs, AI agents, and Edge AI do not address identical needs. Next, identify and prioritize business use cases to begin from real requirements rather than chasing uses that would merely monetize an existing setup. Third, classify and map the data to derive the necessary privacy, sovereignty, and performance constraints. Finally, construct a road map that integrates technology, economic, and environmental dimensions, leaning (whenever possible and justifiable) on open source to reduce dependencies. In short, a sober architecture is not a constraint but a guiding principle of design.
Learn more
