The Heavy Price of Technological Gigantism
Recent financial results from technology giants highlight growing tension within the artificial intelligence industry. Meta recently revealed a 14 percent year-over-year decline in net profit, accompanied by a spectacular 91 percent drop in free cash flow, according to data reported by Reuters and Bloomberg. This erosion of profitability is directly attributable to massive capital expenditures on the computing infrastructure required to train and run increasingly large language models.
To ease investor concerns, Meta CEO Mark Zuckerberg shifted his focus to the opportunities that artificial intelligence agents offer for businesses. His vision relies on deploying massive models capable of centralizing all of an organization's administrative and operational tasks. However, this centralized and monopolistic approach faces a double reality: prohibitive energy and financial costs, and an increased dependence of organizations on infrastructure located outside their borders.
The Alternative: Software Sobriety and Composability
Behind the technological arms race lies a fundamental question: does operational efficiency truly require disproportionately large models? Many analyses, including a report published by the investment bank Goldman Sachs, question the actual return on investment of these colossal expenditures. Similarly, work by the analytical firm Sequoia Capital highlights the growing gap between the revenues generated by generative AI and the investments required to maintain these infrastructures.
Computer science, however, offers a proven alternative: composability. Rather than designing a monolithic AI agent that attempts to do everything centrally (and often fails due to hallucinations or high processing costs), software sobriety advocates for the use of specialized microservices. In this architecture, artificial intelligence is not a universal, opaque operating system, but a conductor. It does not execute all tasks directly; it coordinates existing, lightweight, and targeted applications to achieve a complex goal. This approach drastically reduces the required computing power and, consequently, the technology's carbon and financial footprint.
The Quebec Approach: Orchestration Through Microservices
It is precisely this philosophy of sobriety and composability that structures Quebec's sovereign ecosystem. Unlike Meta's centralized vision, the ProductivIA platform demonstrates that high-performing agentic AI for businesses and institutions relies on the intelligent orchestration of specialized applications via a standardized protocol named assistant_services.
At the heart of this architecture, the ProductivIA Assistant application is not an omniscient agent that reinvents the wheel with every request. It acts as an intermediary capable of understanding user intent in natural language, then delegating actions to other applications on the platform. For example, to prepare a progress report, the Assistant does not try to generate content solely from its general knowledge. It calls the semantic search service of the Document Library (which uses RAG, or retrieval-augmented generation, to anchor responses in the organization's actual documents), extracts the relevant data, and then requests the Doc application to draft the document, before preparing a draft email via the Email application.
This approach offers three major advantages for corporate and institutional organizations:
- Reduced attack surface: By avoiding the integration of heavy frameworks and unmanaged third-party software dependencies, the platform limits vulnerabilities. Each application runs in isolation and communicates via secure gateways.
- No vendor lock-in: Thanks to the GoIA application and the AI Comparator, administrators can evaluate and switch from one language model to another (OpenAI, Anthropic, Mistral, or the sovereign Quebec model Matania) without modifying a single line of application code.
- Data sovereignty: For institutions subject to Law 25 or businesses handling highly sensitive data, the orchestrator can direct requests exclusively to the sovereign provider Matania, whose infrastructure is physically located in Quebec. Data never transits to foreign servers.
This application stack can run in any modern browser, including on refurbished computers thanks to the native Boréal-OS operating system. This hardware and software complementarity offers a concrete response to planned obsolescence while guaranteeing an entirely local digital infrastructure.
Toward a Pragmatic Enterprise AI
The confrontation between the centralized model of tech giants and the distributed model of the sovereign ecosystem poses a question about the future for decision-makers. Should organizations tie their operational destiny to increasingly expensive and unstable foreign infrastructures, or should they prioritize composable, transparent architectures tailored to their actual needs?
The evolution of enterprise AI will not be measured solely by the size of language models, but by the ability of organizations to orchestrate their tools in a secure, cost-effective manner that complies with local legislative requirements.