Should India build its own Foundational Models ?
3 min read
There is a debate among Indian tech landscape on whether India needs to train its own foundational models or should it just rely on OpenAI or Gemini which are owned by Americans. This article throws more detailed light on this topic.
Firstly, training a foundational models is not merely a problem of money. It does take millions of dollars to make one but India being a fairly large country it is doable for India. The real hard part of building great model is the technology that goes into training such model. India right now has all the raw material such as talented people, large data-centers etc. to achieve this but it will take India at least a year to put together teams that can achieve this.
Secondly, the reasons to train such model inside the country. Models trained in USA will be trained for American sensibilities and will be governed by US government which is not exactly a friendly state for India and does not treat India as close ally. While it might let India use the models India is unlikely to be the primary market for such models. If India builds its own models, they will suite India’s need better, might meet Indian expectations better and might not be constrained by US regulations.
Thirdly, the cost of training a foundational model are coming down and I expect that in 5 years, it would be possible to train ChatGPT O1 level model for less than $1M. This will come from improved training mechanism, cheaper hardware and lower energy costs. AI models do not have any technological moat which means there is no secret sauce that allows you to maintain competitive edge. Eventually everyone is going to catchup. All the top companies can do is to make sure their model is say 3 months ahead of everyone else.
India can focus on more flexible infra rather than model itself.
India missed industrial revolution and paid the price. There is no excuse why India should miss AI revolution. To make it happen India must deregulate everything related to AI. It should be easier to import NVIDIA chips or data center equipment into India. Anyone who wants to setup shop in India should be able to set it up easily to make these chips and related hardware.
Then, there should be no restriction on using foreign models either. Use OpenAI or Gemini or whatever that allows Indian companies to build fantastic products that make money.
However, Indian companies must not get into exclusive contracts or marry themselves to any one company. The policymakers must encourage companies to build a “plug and play” style model where switching to a different model should be considered norm and not an exception.
Once this is in place, there will be a market for new companies to arise that might want to build models from scratch. At that point India will also have deeper tech ecosystem to train its own foundational model and also compete globally.
Trust and allow markets to function
Economics is the natural predator of foolish societies. India must let the free market principles operate and manage the AI landscape of the country. It generally leads to better results and Indians have always been very successful in areas where Adam Smith’s invisible hand has guided the direction of economy.
If it is profitable for India to create its foundational models, it will happen in due course of time.