The long deployment timeframe, slow allocation process and financial constraints of end-user companies have resulted in poor utilisation of GPUs (Graphics Processing Units) that are allocated under the IndiaAI Mission.
Only a fifth of the GPUs, a critical piece of hardware powering artificial intelligence that are made available under IndiaAI Compute Portal, are being utilised.
In an aim to make compute infrastructure accessible for AI researchers and startups, the IndiaAI Mission, the centre’s AI initiative, launched its flagship IndiaAI Compute Portal in early 2025. Under the programme, the applicants are provided access to GPUs at subsidised rates. This has brought the effective per hour GPU usage cost below $1, the lowest in the world.
However, the utilisation rate remains underwhelming. In a communication on Monday, the IndiaAI Mission has conveyed to the empanelled cloud service providers that out of the 33,099 GPUs committed by them, only 12,638 (38%) have been assigned AI workloads by IndiaAI. Of the total, only 7,418 (22%) have been utilised by end users.
The IndiaAI Mission has onboarded a host of cloud service providers who in turn provide GPUs or compute power to end-user startups and researchers. While each cloud service provider (CSP) commits a certain number of GPUs, IndiaAI directs allocations to specific developers, be it app developers, large language model (LLM) developers, or others.
In this model, 40% of the utilisation cost is paid by IndiaAI and 60% is paid by the respective customer. But 5,317 GPUs or 16% of the total committed units of this critical hardware remain unused, shows this letter.
"Please confirm the latest GPU utilisation for the allocated GPUs," the IndiaAI Mission told companies in its email.
Also Read: IndiaAI Mission draws 120 more proposals for sovereign AI models
Slow allocation
Companies told ET fear of sanctions and supply chain disruptions have already led H100 GPU prices to go up. The slow allocation rate by IndiaAI has led companies to offer their GPUs outside for better prices.
"As they are not allocating and we are getting better prices outside, we gave it outside," said one company executive requesting anonymity. "They are slow, but demand is there, so we were able to sell outside," he said.
Another empanelled cloud service provider NxtGen Datacenter & Cloud Technologies told ET allocated GPUs are orders placed by IndiaAI, while utilised are what end customers are actually using. IndiaAI only pays for utilisation. There is a gap between committed and utilised GPUs, as the vendors have not yet provided them to IndiaAI, said AS Rajagopal, chief executive of NxtGen Datacenter & Cloud Technologies.
"We are at near 100% utilisation for our existing GPUs. Our 4,096 B200 cluster is under deployment," he said.
Anuj Bairathi, chief executive of Cyfuture India, another empanelled cloud service provider, told ET, "They (IndiaAI) identified certain requirements and asked us to provide GPU facilities to specific customers or prospects. The actual execution of that work happens at our end."
Also Read: Eight firms selected for incentives to develop foundational AI models
End-user constraints
The delay on the part of end-user companies is a key factor. "In one case, after receiving the request, we approached the customer and asked them to issue their 60% purchase order. However, since it is a government customer, they are taking their own time to process it," he said.
"We are fully ready with our inventory and infrastructure. As soon as we receive confirmation from the customer, we will immediately deliver," he added.
In the case of Yotta, another cloud service provider, they have committed 8,192 GPUs, as per IndiaAI's email. Out of this, IndiaAI has allocated 4,864 GPUs to specific developers although a source from Yotta said it has offered more than 5,700 GPUs. But the actual number being utilised is 2,024 GPUs.
This happens because developers may receive an allocation but choose to use only part of it for now, planning to use the rest later. In other cases, some customers face budgetary constraints to pay their share of the cost.
This explains why, against an allocation of 12,638 GPUs by IndiaAI, the actual utilisation stands at 7,418 — which is significantly lower.
Also Read: AI Mission CEO pegs funding window for models at Rs 1500 crore
Concerns about AI development
According to Bairathi of Cyfuture, "In my view, this is concerning. For a country like India, with such a vast developer ecosystem and at a time when nations are heading toward AI-driven competition, 7,400 GPUs in use is far from adequate. The number should have been much higher."
"It is also surprising that out of a total budgetary allocation of Rs 2,000 crore to the IndiaAI Mission for FY25–26, only Rs 800 crore has been utilised so far. For FY26–27, the allocation is just Rs 1,000 crore. In my opinion, this is far too small for GPU infrastructure," he said.
At the same time, hardware costs have risen sharply. DDR4 RAM prices have increased nearly eightfold. GPUs rely heavily on high-memory components, and globally there are only three major memory manufacturers — SK Hynix, Samsung, and Micron.
"Their production capacities are fully booked until 2028, leading to a sharp increase in prices. As a result, GPUs purchased now will be significantly more expensive," Bairathi said.
"We currently have around 33,000–38,000 GPUs available. This is a major opportunity for India and should be utilised extensively. However, a Rs 1,000 crore allocation under IndiaAI, within a broader Rs 10,371.92 crore budget, is in my view insufficient to fully capitalise on this opportunity," he added.
The IndiaAI Mission, and CSPs like AWS, CtrlS, E2E, Google Cloud, Jio, NTT, Neysa, Tata Communications, and Yotta, did not respond to ET's request for a comment.
Also Read: Sarvam and three other AI firms in MeitY’s LLM build out first shortlist
Only a fifth of the GPUs, a critical piece of hardware powering artificial intelligence that are made available under IndiaAI Compute Portal, are being utilised.
In an aim to make compute infrastructure accessible for AI researchers and startups, the IndiaAI Mission, the centre’s AI initiative, launched its flagship IndiaAI Compute Portal in early 2025. Under the programme, the applicants are provided access to GPUs at subsidised rates. This has brought the effective per hour GPU usage cost below $1, the lowest in the world.
However, the utilisation rate remains underwhelming. In a communication on Monday, the IndiaAI Mission has conveyed to the empanelled cloud service providers that out of the 33,099 GPUs committed by them, only 12,638 (38%) have been assigned AI workloads by IndiaAI. Of the total, only 7,418 (22%) have been utilised by end users.
The IndiaAI Mission has onboarded a host of cloud service providers who in turn provide GPUs or compute power to end-user startups and researchers. While each cloud service provider (CSP) commits a certain number of GPUs, IndiaAI directs allocations to specific developers, be it app developers, large language model (LLM) developers, or others.
In this model, 40% of the utilisation cost is paid by IndiaAI and 60% is paid by the respective customer. But 5,317 GPUs or 16% of the total committed units of this critical hardware remain unused, shows this letter.
"Please confirm the latest GPU utilisation for the allocated GPUs," the IndiaAI Mission told companies in its email.
Also Read: IndiaAI Mission draws 120 more proposals for sovereign AI models
Slow allocation
Companies told ET fear of sanctions and supply chain disruptions have already led H100 GPU prices to go up. The slow allocation rate by IndiaAI has led companies to offer their GPUs outside for better prices.
"As they are not allocating and we are getting better prices outside, we gave it outside," said one company executive requesting anonymity. "They are slow, but demand is there, so we were able to sell outside," he said.
Another empanelled cloud service provider NxtGen Datacenter & Cloud Technologies told ET allocated GPUs are orders placed by IndiaAI, while utilised are what end customers are actually using. IndiaAI only pays for utilisation. There is a gap between committed and utilised GPUs, as the vendors have not yet provided them to IndiaAI, said AS Rajagopal, chief executive of NxtGen Datacenter & Cloud Technologies.
"We are at near 100% utilisation for our existing GPUs. Our 4,096 B200 cluster is under deployment," he said.
Anuj Bairathi, chief executive of Cyfuture India, another empanelled cloud service provider, told ET, "They (IndiaAI) identified certain requirements and asked us to provide GPU facilities to specific customers or prospects. The actual execution of that work happens at our end."
Also Read: Eight firms selected for incentives to develop foundational AI models
End-user constraints
The delay on the part of end-user companies is a key factor. "In one case, after receiving the request, we approached the customer and asked them to issue their 60% purchase order. However, since it is a government customer, they are taking their own time to process it," he said.
"We are fully ready with our inventory and infrastructure. As soon as we receive confirmation from the customer, we will immediately deliver," he added.
In the case of Yotta, another cloud service provider, they have committed 8,192 GPUs, as per IndiaAI's email. Out of this, IndiaAI has allocated 4,864 GPUs to specific developers although a source from Yotta said it has offered more than 5,700 GPUs. But the actual number being utilised is 2,024 GPUs.
This happens because developers may receive an allocation but choose to use only part of it for now, planning to use the rest later. In other cases, some customers face budgetary constraints to pay their share of the cost.
This explains why, against an allocation of 12,638 GPUs by IndiaAI, the actual utilisation stands at 7,418 — which is significantly lower.
Also Read: AI Mission CEO pegs funding window for models at Rs 1500 crore
Concerns about AI development
According to Bairathi of Cyfuture, "In my view, this is concerning. For a country like India, with such a vast developer ecosystem and at a time when nations are heading toward AI-driven competition, 7,400 GPUs in use is far from adequate. The number should have been much higher."
"It is also surprising that out of a total budgetary allocation of Rs 2,000 crore to the IndiaAI Mission for FY25–26, only Rs 800 crore has been utilised so far. For FY26–27, the allocation is just Rs 1,000 crore. In my opinion, this is far too small for GPU infrastructure," he said.
At the same time, hardware costs have risen sharply. DDR4 RAM prices have increased nearly eightfold. GPUs rely heavily on high-memory components, and globally there are only three major memory manufacturers — SK Hynix, Samsung, and Micron.
"Their production capacities are fully booked until 2028, leading to a sharp increase in prices. As a result, GPUs purchased now will be significantly more expensive," Bairathi said.
"We currently have around 33,000–38,000 GPUs available. This is a major opportunity for India and should be utilised extensively. However, a Rs 1,000 crore allocation under IndiaAI, within a broader Rs 10,371.92 crore budget, is in my view insufficient to fully capitalise on this opportunity," he added.
The IndiaAI Mission, and CSPs like AWS, CtrlS, E2E, Google Cloud, Jio, NTT, Neysa, Tata Communications, and Yotta, did not respond to ET's request for a comment.
Also Read: Sarvam and three other AI firms in MeitY’s LLM build out first shortlist
