India doesn’t need to build its own LLMs, but we need local context: NxtGen CEO

Can you share the split between your government and private sector clientele?

It is almost 50-50. The government segment grew rapidly post-COVID because we had an opportunity to work with the Ministry of Health. Our large enterprises share dipped during COVID.

Even with enterprises picking up, the government base is fairly large because whenever they do things, they do so at the population scale and the deployment is large. That is an advantage that does not come with private enterprise.  

What role are you playing in supporting the India AI mission, and what specific contributions are you making?

INDIAai is a unique approach. Concerning us, they wanted to ensure good GPUs are available at an affordable cost so AI adoption increases. We wanted to align with them and so, helped create the lowest price for NVIDIA H200 Tensor Core GPUs. This will augur well for their mission. We are among the three companies to provide GPUs, but we are focused on high-end GPUs.

Startups can apply in the INDIAai portal and register their projects, depending on which, some of their costs may be subsidised. At the minimum, they will give access to great pricing they would not get otherwise.

Some may get a 100 per cent subsidy, but many are getting 40 per cent or 20 per cent. Our pricing isn’t the same -- it is slightly higher than standard India AI rates, but still well below market levels. Access to this infrastructure is not limited to INDIAai but is available directly as well. However, INDIAai does receive certain exclusive benefits from us.

How will these collective efforts shape India’s AI journey? Once fully implemented, can India compete with global AI leaders?

India does not need to enter the race of building its own language models because most open-source models are familiar with our related content. The language issue is being solved. But with LLMs, you want to embed some cultural aspects into the model. For general commercial use cases, many open-source models are readily available to quickly leverage and build upon.

A few companies are working on building India-specific language models, with many building something tangible for India to consume.

In our case, we are doing two things -- building a model called Accura, and training on the Indian financial accounting system, GST, and other interpretations in the high court. From 1969, we took all court cases and distilled that data; we trained it on roughly 9 lakh pages. It understands and behaves like a chartered accountant now. This is an India-specific use case because it is Indian accounting laws, Indian corporate law, and Indian context. 

With Juris, we are doing the same thing with the law as well; We want to enable citizens to easily understand legal or tax processes—information they currently need a lawyer or chartered accountant to access. 

We wanted to solve this with a publicly facing model. Just like us, many others are building things specific to the country which cannot be solved by ChatGPT and others since they don’t have our context.

How have computing requirements evolved since 2014, when you set up your data center? Has AI and cloud impacted demand for data centers?

The data center is the foundation of the cloud. No customer accepts any outage irrespective of how big or small you are. So it is important to keep it running once switched on. When we built this facility, we thought we would run out of capacity in three to four years.

At the same time, compute density has been increasing. For example, when we started, each server had about 12 cores and typically ran 12 to 24 processes, depending on workload distribution. A server generally has a lifespan of five years.

We invest in the highest-grade GPUs hoping they’ll stay relevant for five years--something that worked well for stable, traditional workloads. But with AI evolving every few months, we have to be more cautious. We avoid heavy upfront investments and take a phased approach.

Ultimately, the customer bears the cost. If your infrastructure is inefficient and overpriced, and the value isn’t clear, the customer walks away and the business suffers. That’s why in AI, it’s crucial to move step by step.

What is the status of your announced $400 million investment in infrastructure?

We are in the final stages and have four interested parties. The process is being run by an investment banker called Avendus, based out of Mumbai, who is helping us get the right partner in place. For AI infrastructure, we may need $2-3 billion over the next two or three years since what we currently have is small.

The moment Indians have a computing device in their hand, they will use AI. Getting to that scale needs billions of investment. We are looking for a partner who can make that investment going forward.

Currently, we operate five data across Bengaluru, Ahmedabad, Faridabad, and Mumbai of which two operate dedicatedly for the government, with specific levels of protection due to sensitive information.

We will build more, but that depends on understanding the compute demands of future AI workloads. Once we have clarity, we will design and construct accordingly. We are also working with GPU providers.