GPU firms step up their tech stack

Time of India

13-05-2025

Indian AI cloud and GPU-as-a-service companies such as Neysa, JarvisLabs and NeevCloud are altering strategies as AI market dynamics shift. While Neysa continues to invest in advanced GPUs to double down on inferencing, others such as JarvisLabs and NeevCloud are taking a step back and not buying any more GPUs. Instead, they are focusing on building the software stack.The reasons for this range from low utilisation of high-end GPUs, high capital investment required for purchasing them and better and cheaper access to such chips in recent times. In addition GPUs are capital intensive and unless startups have raised enough funds, it is a risky business. While training and fine-tuning of large language models is yet to pick up at scale, companies are also focusing on inferencing, a process in which trained AI models use the data to predict, reason, or solve problems.
The GPU issue
Vishnu Subramanian, founder of Jarvislabs AI, which offers
GPU rental
and
AI cloud services
, explained that one of the reasons they bought GPUs in 2019 was that back then no one else was buying them and costs were exorbitant. Jarvislabs offered the service at an economical price to startups and the student community in the country. But with more companies entering the field in India and globally, costs have plummeted.
'The rental prices and margins at which you can rent has significantly gone down. For example, on-demand H100 (an Nvidia GPU) used to cost $11-12 a year back. Now you can get them for less than $3 from a decent cloud service provider,' he said.
India's appetite for spending on high-end GPUs has declined. 'Even within Jarvislabs, we see a bigger chunk of revenue coming from the West,' Subramanian said. As a result, the company stopped buying GPUs a year ago and is partnering with global players to offer processing capacity.
The Tech stack focus
In addition, Jarvislabs is also focusing on building the orchestration layer. This refers to systems that manage multiple AI components by streamlining the process, scaling and bringing in efficiency. NeevCloud founder Narendra Sen said it began with a plan to build the CoreWeave of India. It started out renting GPUs and then entered orchestration and application layers.
'But we realised that GPUs are a commodity and you need to build a technology for consumer stickiness and provide value beyond the GPU such as improving chip performance,' he said. As a result, NeevCloud, instead of bulking up on GPUs, is taking a call to buy them based on demand. 'This is only 10-20%,' he added. The firm did not disclose the scale of GPU operations, but this had been a key strategy earlier.
Sharad Sanghi, co-founder of Neysa, an AI cloud platform, said some companies are not focusing on GPUs as they are capital intensive and, unless they have money, it is a risky business. Neysa has so far deployed 1,200 GPUs and is in talks to place further orders for advanced chips including Nvidia Blackwells, expected in India later this year. This backs up the company's focus on inferencing-as-a-service.
Sanghi said that with (Indic) foundational models such as Sarvam coming, there will be use cases for training and finetuning. While the firm was doing all three earlier—training, finetuning and inferencing, Sanghi said the enterprise space is going to be more of a market focused on the latter.