Reliance Industries Limited has announced a large investment in AI compute in India. The interesting part, for me, is not just the scale of the infrastructure.
It is what happens to the economics of AI when compute becomes more affordable and locally available.
The first 120 MW phase, expected to be commissioned by the end of 2026, will start with an initial fleet of NVIDIA GB300 GPUs equivalent to more than 75,000 H100 GPUs on an inference basis, and can scale to over 200,000 H100-equivalent GPUs as the phase ramps up.
I have said this before: AI works across three layers - infrastructure, models and applications.
Infrastructure may become more accessible and economically competitive as capacity expands, but it will remain an essential layer. And when the cost of that layer falls, the economics change for everyone building on top of it.
I see this with many mid-market companies. They already have AI use cases identified. But when the economics are worked out, some of those use cases are still difficult to justify at today's compute prices.
That could change.
📍 Use cases that are expensive today could become economically viable as local compute capacity increases. 📍 Lower compute costs will make experimentation easier, which means companies can run more experiments and learn faster. 📍 But waiting for compute to become cheap is not a strategy. Companies should be building their data, skills and understanding of use cases now. 📍 The companies that benefit most when compute becomes cheaper will be the ones that already know what they want to do with it.
This is an important distinction.
Cheaper compute doesn't create AI capability. It makes existing capability easier to experiment with and more economical to scale.
Infrastructure providers may solve more of the compute constraint over time.
Our job as businesses is to understand our problems, prepare our data, experiment with the technology and identify where AI can create real value. By the time compute becomes cheap enough for everyone, you don't want to be starting your first experiment.
What would become viable in your business if compute stopped being the constraint?