India’s AI Challenge
Affordable ai for real applications
Bharat Gen is our only option AI is what ai does.
Technology, hardware, software etc are all just means to an end. There are survival clothes, fashion clothes and fancy dress competition clothes.
The field of ai has similar segments. What a lot of the world is chasing are fancy dress competition clothes. It is best left to the Global North. They will need a lot of these for the working to be able to afford four days of holidays every week and to pay pensions to the non working. We don’t have those issues.
For India, we have to focus on survival clothes. Which is using ai to improve our employability and productivit. to do so, we have to be self sufficient in Ai. Yet we have to be mindful that we are a capital starved nation with a lot of mouths to feed and the realities of a democracy.
We have to focus where value accrues across the AI stack. We have to use frugal innovation to lead in all the tiers of the stack that matter.
The first tier is the interconnect fabric. This is the electronics which connects the chips, storage and other components. The railroad analogy applies: it matters less who builds the locomotives if you control the rail gauge standard. This is Bharat Gen’s key deliverable.
The second tier is photonic interconnect components. Optical interconnects are becoming mandatory rather than optional for frontier AI training. Copper has distance limitations at high speed. Optical does not. we have to build high power processors which have photonics on board. This means diamond substrates for rapid heat removal and complex compound semiconductors for the photonic integration. The silicon itself can be a fpga which can be bought out. This is as yet no man’s land. China is not yet a contender here. In the U.S., it is the Indians mostly IIT Alumni startups like Poet that lead.
The third tier is nuclear power generation for AI. Nuclear is unique as an energy source. Waste heat can be used to produce refrigeration using supercritical co2 with geothermal pumps. 1 MW of nuclear equals 25 MW of solar or 10 MW of thermal. Nuclear wins hands down on being cheaper and greener and indigenous. The iit alumni fund got this right five years ago and has a massive head start on anyone else in the world. And this includes USA and China.
The fourth tier is accelerator compute. Custom silicon from hyperscalers will continue capturing share for specific inference workloads. But there is more value in the interconnections then there is in the compute. Intel isn’t recovering anytime soon. This is a commodity which we can buy.
The fifth tier is the software.. it is not foundational ai software which needs the man years. It is the fine tuning of the ai model and development of applications where we need resources. The fine tuning can be done using distributed resources.
It is the applications and the consumer apps that deliver those applications which will decide the winners.