The Return Journey

Three in four employees worldwide are already using AI at work, and most of that access runs through someone else’s servers, a login, a subscription, and a black box that a business doesn’t control. That arrangement is starting to show real cracks. Confidential code has been pasted into public chatbots, AI budgets have overrun at more than two-thirds of companies, and India’s Digital Personal Data Protection Act now carries penalties of up to ₹250 crore for data that doesn’t stay on infrastructure the business itself runs. At the same time, open-weight models from Meta, Mistral, Alibaba, DeepSeek and Google have closed most of the performance gap with the biggest hosted models, giving businesses a genuine choice for the first time between renting AI indefinitely and owning the infrastructure it runs on. This paper traces that shift in full: the pressures pushing organisations to bring AI in-house, how the move from pilot to production actually unfolds, the real economics of self-hosted open-weight models against rented APIs, and what changes on the ground for engineering, legal, regulated operations, and field teams. It sets out a practical case for businesses to make AI infrastructure a deliberate ownership decision, on their own terms.