In today’s time, when just about everyone wants to get their hands on NVIDIA GPUs, Jarvislabs, a GPU cloud platform, has announced that it will provide H100s on demand. “We are partnering with companies across the globe, and we already have access to thousands of NVIDIA H100 GPUs,” said Vishnu Subramanian, the founder of Jarvislabs, in an exclusive interview with AIM.
He said that Jarvislabs, just like Hugging Face, isn’t looking at buying GPUs directly but wants to partner with established companies and hyperscalers. However, the company will continue to own GPUs.
Hugging Face recently committed $10 million in free shared GPUs to help developers create new AI technologies. The shared GPUs are accessible to multiple users or applications simultaneously, removing the need for each user or application to possess a dedicated GPU.
“We have been owning GPUs for a very long time. We still own them, and we still run them, but for scaling purposes, we want to partner with other players. There are many GPU providers,” he said.
Jarvislabs, as a company, is bootstrapped and situated on the outskirts of Coimbatore. “We started as a small team of just four members, all from humble backgrounds, and none from the IITs or IIMs. Currently, we are a team of 12+ employees,” said Subramanian.
“The reason why we started Jarvislabs four years back is when I was working at a Fortune 500 company at their data centre with GPUs, [I saw] them struggle for months to use them properly,” said Subramanian.
He said that developers and artists from across the world now use Jarvislabs’ services, which include some prominent names like ZOHO, Weights and Biases, UNC, UpGrad, and many more.
The platform faces competition from local players such as Yotta, E2E Networks, Ola Krutrim, and Johnaic. However, Subramanian doesn’t perceive them as competitors. “We don’t see them as competitors, instead we see them as potential partners. If they can provide raw compute power, we view it as an advantage,” he said.
More Than Just GPUs
Subramanian highlighted that while several companies in India provide compute as a service, the real value lies in enhancing software to benefit customers. He said that they provide software along with the GPUs or bare metals, where users don’t need to learn coding. Even teams without data engineers can easily run LLMs and develop new products.
“If you go to a typical hardware provider you get the VMs (Virtual Machines), but you still need to have the know-how to train LLM foundational models or anything else, like Stable Diffusion, or any part of AI or ML yourself, which can be quite costly,” he explained.
Subramanian said they are following the Hugging Face route. “They don’t say that we sell H100s. They sell API endpoints. They help you with model training. That’s exactly what we do,” he said.
“We are inspired by Hugging Face. I think it’s one of the companies that has done a tremendous job in the AI space. The amount of energy both the CEO and CTO carry is very inspiring,” he said.
“I think their target is more towards enterprise customers, and we are targeting the SMBs of the world. In a way, I would say that they are the Salesforce, and we are the HubSpots, Fishboxes of the world,” added Subramanian.
Jarvislabs also supports Axolotl which is a no-code tool for model training. For instance you can train open-source models like Llama 2 or 3 with your data.
He further said that they also recommend the right GPUs suitable for a particular project along with the API endpoint, which is secured by Cloudflare. “You make one click, you get an API endpoint, which is secured by Cloudflare so that if someone is bombarding you with attacks, Cloudflare is there to protect you.”
Jarvislabs has integrated ComfyUI into its services, offering users a convenient way to access and utilise this tool without the need for local setup or installation. ComfyUI is a web-based Stable Diffusion interface optimised for workflow customisation.
It supports models like SD1.x, SD2.x, SDXL, Stable Video Diffusion, Stable Cascade, SD3, and Stable Audio.
Jarvislabs’ new updates include lightning-fast model loading, over 100 pre-loaded checkpoint models, more than 60 custom node packs, the full ControlNet suite, and all IPAdapter models.
“What gets more interesting is now you can just connect this via an API from our mobile app or from your website and you have an AI product ready without really writing any piece of code, without working on deployment, and without worrying about GPU,” Subramanian concluded.
The post This Coimbatore-based AI Lab Offers NVIDIA H100 GPUs and No-Code LLM Training appeared first on Analytics India Magazine.