47
India’s artificial intelligence startups are facing a growing challenge in accessing graphics processing units, or GPUs, the specialised chips used to train and run AI models.
Even startups with funding and cloud credits are finding it difficult to secure enough computing capacity. Delays in getting GPUs are slowing development, pushing up costs and, in some cases, delaying product launches.
According to founders and investors cited by Rediff, GPU deliveries can take three to four months. For startups working with tight product deadlines, such delays can leave engineering teams waiting and postpone planned launches.
The problem is increasingly about availability rather than just cost. Bengaluru-based payments technology company Juspay is facing a similar challenge. Its co-founder and chief operating officer, Sheetal Lalwani, said the company had received GPU delivery commitments stretching between three and six months.
Juspay runs around 95 per cent of its AI-assisted coding workloads on open-source models using GPUs under its control. Lalwani estimated that renting an eight-node GPU setup costs between Rs 25 lakh and Rs 35 lakh per month.
However, the report does not identify the specific GPU models involved, so the figure should not be considered a standard market rate.
For startups, renting GPUs can help avoid large upfront investments and the risk of hardware becoming outdated. Buying infrastructure can make more sense when workloads are predictable, but it also requires significant capital before a company knows how quickly its AI demand will grow.
The challenge goes beyond GPUs. High-bandwidth memory, data-centre capacity, electricity and cooling are also important factors in determining how much AI computing capacity is actually available.
The government says more than 38,000 GPUs have been onboarded under the IndiaAI Mission through empanelled providers. These computing resources are being made available to startups and academic institutions at subsidised rates.
However, shared computing resources may not meet every requirement. Robotics companies, for example, may need local processing to reduce delays, while some defence-related workloads could require dedicated infrastructure because of security considerations.
India’s AI ambitions will depend not only on investment and talent, but also on reliable access to computing infrastructure.
For startups developing AI products, the ability to secure computing capacity at the right time could determine how quickly a product moves from development to the commercial market.
Even startups with funding and cloud credits are finding it difficult to secure enough computing capacity. Delays in getting GPUs are slowing development, pushing up costs and, in some cases, delaying product launches.
According to founders and investors cited by Rediff, GPU deliveries can take three to four months. For startups working with tight product deadlines, such delays can leave engineering teams waiting and postpone planned launches.
The problem is increasingly about availability rather than just cost. Bengaluru-based payments technology company Juspay is facing a similar challenge. Its co-founder and chief operating officer, Sheetal Lalwani, said the company had received GPU delivery commitments stretching between three and six months.
Juspay runs around 95 per cent of its AI-assisted coding workloads on open-source models using GPUs under its control. Lalwani estimated that renting an eight-node GPU setup costs between Rs 25 lakh and Rs 35 lakh per month.
However, the report does not identify the specific GPU models involved, so the figure should not be considered a standard market rate.
For startups, renting GPUs can help avoid large upfront investments and the risk of hardware becoming outdated. Buying infrastructure can make more sense when workloads are predictable, but it also requires significant capital before a company knows how quickly its AI demand will grow.
The challenge goes beyond GPUs. High-bandwidth memory, data-centre capacity, electricity and cooling are also important factors in determining how much AI computing capacity is actually available.
The government says more than 38,000 GPUs have been onboarded under the IndiaAI Mission through empanelled providers. These computing resources are being made available to startups and academic institutions at subsidised rates.
However, shared computing resources may not meet every requirement. Robotics companies, for example, may need local processing to reduce delays, while some defence-related workloads could require dedicated infrastructure because of security considerations.
India’s AI ambitions will depend not only on investment and talent, but also on reliable access to computing infrastructure.
For startups developing AI products, the ability to secure computing capacity at the right time could determine how quickly a product moves from development to the commercial market.
