Buy or build? One trajectory from ambition to pragmatism is worth a thousand AI strategies. Singapore’s SEA-LION AI model started by pretraining from scratch. By v2 and v3, the team shifted toward adapting open pretrained models with region-specific data and post-training. By v4, they deliberately scaled down from 70B to 27B, because deployment constraints and inference costs, not pride, were binding in Southeast Asia.
The paper behind that story, by Lu, Xu, Tjhi, Li, Bosselut, Koh, and Kankanhalli, lays out the acquisition pathways governments actually face. API access and managed services at one end. Hybrid approaches in the middle (RAG, sovereign cloud partnerships, fine-tuning). Full pretraining at the other end. It then compares the trade-offs across sovereignty, privacy and security risk, cost, domestic capability building, sustainability, and national context fit, with an explicit nod to how fast the cost-capability frontier is moving.
Two reasons I really like this paper. First, it is not written from the balcony. It draws on practitioner accounts from the teams behind SEA-LION and Switzerland’s Apertus, including what turned out to be harder than expected: data acquisition, infrastructure software, and talent retention. Most AI-for-government writing either stays at the level of slogans or disappears into architectures. This one connects strategic choices to operational reality.
Second, the idea of capability debt. When governments outsource everything, they do not just defer effort. They compound the eventual cost of catching up, because organizational learning accrues through iteration, not procurement. The paper is also clear on the other side of the argument: buying is not a lesser choice. For many use cases and economies, the most responsible path is to consume models as a managed service, while investing in data governance, integration, evaluation, and contractual controls. That is how states retain real control, through data rights, audit and safety requirements, portability, and credible exit options, even when they are not building the underlying foundation models themselves.
That last point aligns with what I keep seeing in work on sequencing AI deployment. The real gap will not be between countries that build models and countries that do not. It will be between the small set of governments that choose a deployment pattern they can sustain, and the larger set that reach for one they cannot maintain.