Analysis ·
AI in production: should you invest in dedicated capacity?
Original title, in French: “IA en production : investir dans une capacité dédiée ?”
- Publication
- Brief IA (France)
- Author
- Tom Levy
- Published on
- 29 September 2026 · read in the page structured data
Summary
Companies that are industrialising their use of AI must choose between pay-per-use and investing in dedicated capacity. The choice depends on the level of sustained usage, the predictability of demand over 12 to 18 months, and the organisation's ability to keep that capacity productive through governance and ongoing adoption. According to Deloitte, employee access to AI grew by 5% in 2025, and the share of companies with at least 40% of projects in production is expected to double within six months. The article details the diverging cost profiles of assistants, retrieval systems and agentic workflows, and stresses the need to model actual workloads rather than rely on generic benchmarks when sizing infrastructure.
Summary based on the full text of the article.
Read the article on Brief IA →Figures cited
-
5 %
Increase in employee access to AI in 2025
companies, 2025
« Deloitte observe une hausse de 5 % de l’accès des salariés à l’IA en 2025 »
Source: Deloitte
-
—
Expected doubling of the proportion of companies with at least 40% of projects in production within six months
companies, six-month horizon
« anticipe un doublement, d’ici six mois, de la proportion d’entreprises comptant au moins 40 % de projets en production »
Source: Deloitte
Classification
- Country
- Worldwide scope
- Company size
- All sizes
- Industries
- All industries, or industry not specified
- Companies
- Deloitte (study author)
- Keywords
- dedicated capacity, AI costs, production deployment, AI infrastructure, AI agents, AI governance
Entry written by Marsanne Studio from the cited article, with the help of an AI model (claude-sonnet-5). The rights to the article and its illustrations belong to Brief IA and its author. Entry translated from French by Claude Sonnet 5.5.