Data centre operator Yotta is planning to further expand its compute capacity to 32,000 graphic processor units (GPUs) from Nvidia by March 2025, months after placing orders for 16,000 high-powered GPUs from the US-based chip giant.
“We are having a soft commitment from Nvidia, and these chips will be delivered to us by 2025,” said Sunil Gupta, chief executive officer, Yotta Data Services.
The company has already invested around $1 billion to procure 16,000 H100 GPUs and develop infrastructure for servers, including associated network, storage, software, and additional layers. While Gupta did not give details of the funds required for the additional chips, he added that the capital requirements for the project will be backed by the promoters and they may also look for financing from banks.
The Hiranandani group company’s first GPU-based data centre, which is expected to go live by May 15, 2024, will provide compute capacity for Artificial Intelligence (AI) processing, powered by the first batch of 4096 GPUs it has received.
Gupta, who is bullish on the market demand for GPUs in India, said that the market for high-performance compute in India is expected to grow big, and these were still early days for the sector in India.
On the demand front for GPU servers, he said that Yotta’s first slot of more than 4,000 GPUs was already booked by enterprises and is expected to go live by May 15, 2024.
He also said that apart from serving India, the company was also looking to cater to the global demand.
“We can serve not only our respective country but also the nearby geographies. So our growth pattern in terms of sourcing GPUs possibly will not stop, simply because we are just starting,” he said.
Gupta further said that because of the global shortage of GPUs, there was a demand from European regions as well. “I can see the demand going up globally in the last four to five months, and a bigger part of my sales funnel today is actually requirements from Europe, the APAC region, and from the Middle East,” he added.
The company is aiming to compete with hyperscalers like Amazon and Meta by providing GPU servers at a competitive price of $2 – $2.5 per hour.
Contrasting it with the global rates of GPU servers, Gupta said that the company was looking at making money through economies of scale and right pricing.
“If you go to hyperscalers today, which many of the startups have done, you will get GPUs at somewhere between $9 to $12 per hour. Similarly, in the US, for specialised GPU providers like Corebeam and Lambda, the prices range between $3 to $5. My price point for the same thing is between $2 to $2.5,” Gupta said.
He added that depending on the contract, pricing can vary, where short-term GPU usage might cost more per hour or week, whereas committing to a longer contract with upfront payments could bring the price down to as low as $1.8 per hour of usage.
Gupta said that while traditional cloud and hosting services were typically taken by enterprises, large cloud operators like hyperscalers who wanted co-location services, these were not the market for GPU-servers, at least in the first phase of GPU subscription.
“For the next year or so the GPU subscription was being taken by entities such as startups, or research labs – who wanted to train AI models,” he added.
He also highlighted the need for ‘intermediaries’ in between the data servers and enterprises who can help the companies understand the use cases and changes that AI can bring to the organisation.
“We require companies like Accenture, Deloitte, and others to engage with enterprises, grasp their business situations and challenges, and demonstrate how AI could potentially boost their productivity. Subsequently, they can identify use cases, utilise pre-trained models, train them using enterprise data to tailor specific models for each enterprise, and integrate these models into their existing applications,” he explained.
First Published: Apr 07 2024 | 5:37 PM IST