“We believe in the Zoho School of Thought, inspired by Sridhar Vembu, that you don’t need to build a company from a city like Bengaluru,” said Vishnu Subramanian, founder and chief of Jarvislabs.ai in an exclusive interview with AIM. Situated on the outskirts of Coimbatore, Jarvislabs.ai is silently playing a crucial role in providing AI infrastructure (GPUs on rent) for budding entrepreneurs and developers at an affordable price, alongside helping them advance AI research. 

Jarvislabs.ai is one of the few companies in India that provides GPUs on rent for training, fine-tuning, and deploying AI models. Moreover, utilising services from hyperscalers like Microsoft Azure, Google Cloud, and AWS can be a costly affair for those who simply want to experiment. “If you compare with a hyper scalar, I would say we are like 60 to 70% cheaper” shared Subramanian. 

“We wanted to make it easy and affordable for a small company in Chennai or Bengaluru to deploy a model natively, which can be expensive, especially when considering platforms like AWS or Google Cloud Platform. Even for older GPUs, they often charge around 6.7 cents, whereas we provide a modern GPU at a comparable price.” added Subramanian. 

Unlocking Affordable GPU Prices

Subramanian clarified that their strategy to keep GPU prices low involves not pursuing NVIDIA’s H100s. Instead, they provide alternatives that offer similar capabilities. “The reason we are able to keep the costs low is we are not going behind the fancy H100s which are super expensive today. But you can do a lot of really amazing stuff with the Quadro or the next generation GPUs.” said Subramanian. 

“We have A100s right now. We are not directly investing in H100; instead, we plan to partner with different hyperscalers or similar platforms to incorporate H100s into our platform. Currently, the GPUs we host include Quadro RTX 5000 and 6000. These are the latest generations and are as good as the older generation A100, but they come at a much cheaper price.” he added. 

“For example, RTX 6000 Ada comes at around $1.14, providing similar performance to A100, for which you would end up spending around $3,” he explained.

While Subramanian didn’t disclose the exact number of GPUs acquired by Jarvislabs.ai, he did shed light on the strategy behind securing them. “There is a national distributor named RP Tech in India. They are the sole company responsible for importing GPUs into India. Fortunately, we have partnered with them, and from the start, they have been generous enough to assist us in obtaining GPUs at a very competitive price compared to the open market,” he said, saying that the prices are usually higher on ecommerce websites like Amazon, compared to buying it directly from distributors. 

Subramanian also mentioned that they’ve barely spent a penny on marketing. Word of mouth has always helped them, as they have been offering a decent and good product. “People have been kind enough to introduce us to their friends, and word of mouth continues,” he said. 

This is quite true as we came across Jarvislabs.ai when AIM was talking to one of the local developers, where he, in passing, mentioned about this AI lab, and how it gives GPUs-on-rent for building AI models and advancing AI research. 

Current Customers

Subramanian said that their target customers are anyone ‘who is not a Fortune 500 company.’

Some of the notable customers of Jarvislabs.ai include Zoho and upGrad.”Some of the top companies in India, like Zoho, have been using LLMs for different workloads, probably to help students learn programming code, like a copilot or something similar. I’m not sure exactly what they’re doing, but I would assume that they could be fine-tuning some of the models.” said Subramanian. 

Universities like CMU (Carnegie Mellon University) from the US have been using Jarvislabs.ai’s services to train large language models. Additionally, Weights and Biases are also leveraging its services.

“Our primary emphasis is on the global market, and until at least mid-2023, there wasn’t much usage from Indian customers. It was predominantly outside India, in the US and Europe. However, we see a slow change now; Indian customers are also starting to adopt more and more AI,” said Subramanian. 

Not Just a Hardware Company 

Subramanian emphasised that Jarvislabs.ai is not merely a hardware company but rather a comprehensive solution provider saying “How to train your model and how to deploy it are much more challenging engineering problems to solve. We are packaging this entire thing for our customers.”

Additionally, as a bootstrapped company, Jarvislabs.ai cannot afford to offer substantial discounts to its customers, unlike VC-backed firms that often provide significant discounts as they aim to capture a larger market share. 

To compensate for this, Jarvislabs.ai is providing additional value to its customers at the same price. “We’re not just focusing on renting GPUs as raw material. We are trying to help. Let’s say, for example, you are a small company with 40-50 people. You may not have the DevOps team or the AI team to create an optimsed environment. So, we also simplify things for them. We have automated all these steps. With just a few clicks, all of this is done for you.”

He shared that, at the end of the day, as a company, it also has to make money, and it’s not sustainable to buy GPUs at a very high price. “NVIDIA has been increasing prices year-on-year with every release. For example, the RTX 6000 used to cost us around three lakhs. The current generation costs around six to 6.5 lakhs per GPU, but the cost at which we give it to the customer has not increased a lot,” he said.

Subrmanian said that he intends Jarvislabs.ai to be like an app store which will host multiple frameworks. “We are working on a new version of the product, and we hope to release it this month. This version will incorporate a lot more frameworks. When we started, there were only three popular frameworks: PyTorch, TensorFlow, and Fast.ai,” he concluded.