Recent incidents at Google and Anthropic have underscored the growing risks associated with using artificial intelligence as a public API, or application programming interface. These events have tested the resilience of companies that rely on hosted AI services, which are increasingly central to business operations. In one instance, Google's Gemini model was temporarily cut off, disrupting production lines overnight, despite no fault on the part of the customers. Shortly after, Anthropic's latest models, Fable 5 and Mythos 5, became inaccessible due to an emergency export restriction imposed by the U.S. government on non-American citizens. A similar situation occurred with OpenAI's GPT-5.6 series, which was reserved for a select group of American clients for political reasons and later withdrawn. For companies in France, these episodes echo concerns related to national laws like the Cloud Act and other extraterritorial regulations, which complicate data control and compliance. The availability of AI services has become unpredictable due to external factors, making resilience, control, and predictability critical for businesses. The real issue is not the AI technology itself but how well companies can withstand disruptions like geopolitical tensions, global outages, and opaque changes in product availability and pricing. These challenges directly impact business stability. The problem of costs often arises too late in AI projects. Initially, usage-based pricing models appear cost-effective, especially when AI adoption is limited. However, once AI becomes widespread, particularly with the use of AI agents, costs can spiral out of control. GitHub Copilot, for example, shifted to a per-token billing model, causing many developers to exceed their budgets with what was previously considered normal usage. Similarly, Uber exhausted its 2026 budget for development tools and generative AI tokens, such as Claude Code, within four months. These cases highlight how quickly opaque costs can become a burden for entire teams. Data security is another crucial concern when using hosted AI models. Issues such as data protection, the disclosure of sensitive company information, and intellectual property risks must be carefully managed. According to the Stanford AI Index Report 2025, the number of recorded AI incidents increased by 56.4% in a single year. Meanwhile, the use of public APIs makes it difficult to trace training and validation data, as well as the origin of the infrastructure, which is a requirement for "high-risk" AI systems under the AI Act and monitored under the General Data Protection Regulation (GDPR) by the CNIL in France. Contracts with service providers often overestimate their ability to prevent outages or access restrictions driven by geopolitical factors. This is why the SecNumCloud qualification aims to keep sensitive data out of the reach of non-European legislation. Private AI offers a more secure and controlled alternative. Unlike public AI models, private AI keeps all data within the company's sovereign domain, regardless of whether the infrastructure is hosted in a public cloud, sovereign cloud, on-premise, or in a private cloud within a data center. This approach strengthens security, protects intellectual property, and ensures data traceability and audibility, which are increasingly important in the era of digital sovereignty. Companies retain control over availability, governance, and costs, which are essential for long-term stability and profitability. While private AI may require a heavier initial investment, it can be more cost-effective in the long run, with reduced dependency on vendors and predictable expenses. Public AI models are a possible option, but not a necessity for companies aiming to deploy AI at scale. As the examples of Google, Anthropic, and OpenAI demonstrate, reliance on public models carries recurring risks such as outages, access restrictions, and geopolitical interference. Private AI, on the other hand, enables companies to operate autonomously, free from external disruptions, making it a more stable and secure foundation for long-term AI integration.