By Rajiv Nair, Public Sector Governance Adviser at the Commonwealth Secretariat

If you look at Artificial Intelligence (AI) news anywhere today, people seem to be transfixed by the incredible leaps made in AI. We are trying to keep up with the developments, from ChatGPT to Gemini, then Llama to Claude and Grok, and now DeepSeek and Qwen. We look at these models and focus on their ability to reason, to understand advanced physics, and analyse images. With every iteration, they are getting better and what you see today is the worst AI will ever be.
I believe that the underlying AI models we are all grappling to understand and apply productively will serve purely as a low-margin utility in the AI Age.
Telecommunication déjà vu: Infrastructure vs. applications
Consider the telecommunications revolution. Significant investments were needed to lay out the infrastructure - telecom towers, copper and later fibre cables, undersea cables, and so on. These investments were either by state-owned enterprises or by the private sector (under a licensing framework). Yes, these organisations bill us monthly for using their services, but it’s a utility - much like electricity in our homes. Yet that infrastructure has enabled applications like Google, Amazon, Facebook, and Uber to flourish.
AI models and all the infrastructure that supports them will be uncannily similar. Why do I say that? Let’s first understand what these large language models (LLMs) actually do. Trained on almost all publicly available information on the internet, they now understand and find meaning in language, images (and thus videos), and voice inputs.
These capabilities will only keep improving. But these models don’t understand the specific problem you’re facing - nor do they create a tailored solution. That requires a layer of applications on top of these LLMs that require human intervention to design and harness the analytical and creative prowess of AI to solve real-world problems. Think back to first having the internet and then seeing the dotcoms spring up.
Why chasing infrastructure is not the answer
Building AI infrastructure is prohibitively costly for the developing world - especially at pioneering stages. However, building applications on top of existing AI infrastructure is considerably cheaper and offers added value.
What about data biases? Won’t the AI models built elsewhere be biased? Possibly, but since these models already know how to interpret language and analyse images, you can fine-tune them to your country’s nuances at a fraction of the cost it would take to build a whole new model.
What about your data sovereignty? In my view, it is a patient gamble. As DeepSeek proved, if you wait long enough, you can build your own version of the AI infrastructure at a fraction of the cost borne by the companies (and countries) at the frontier of these technologies.
For countries that see AI as an opportunity, I encourage you to focus your energies on becoming an application powerhouse. The AI infrastructure race will eventually stabilise into a commodity, but the real value will lie in how you solve local and global problems – on top of that AI pipeline.