It’s been three months since the government announced Sarvam as the first startup selected under the IndiaAI Mission to build sovereign foundational models. A month later, the mission added three more startups. However, the GPUs promised are yet to be received by the firms.
Soket AI Labs, Gnani.ai, and Gan.ai were promised access to compute resources, specifically thousands of high-performance GPUs to train their models. However, to date, only Sarvam has received approximately 1,500 GPUs from the initial tranche, as it continues to await further deliveries.
According to IndiaAI’s official website, Sarvam was allocated a subsidy of ₹98.68 crore towards its total compute bill of ₹246.71 crore for 4,096 NVIDIA H100 GPUs over a six-month period. While that makes Sarvam the largest beneficiary under the IndiaAI Mission yet, insiders told AIM that the remaining GPUs were yet to be delivered.
Meanwhile, the rest of the startups—Gnani, Soket, and Gan—have received nothing.
Delay Eats into Timelines
These models were supposed to be built within 6–10 months. That window is closing fast. Soket AI Labs, led by CEO Abhishek Upperwal, is quietly building what could become one of India’s most ambitious AI projects: a 120-billion parameter language model trained on India-centric datasets from scratch under its Project EKA.
But the journey to this number is anything but linear, as the company is slowly making small models with the available GPUs at the moment. “We are doing whatever we can before we get the large-scale compute. We don’t want to waste any time on the compute cluster,” Upperwal told AIM.
The company plans to keep it open-source and optimise it for sectors such as defence, healthcare, and education. “It will take time,” Upperwal said earlier. “It won’t be a one-shot deal. We will scale it up to 120 billion parameters and will have to scale it little by little.”
Meanwhile, Gnani.ai is working on a 14-billion parameter voice AI model. Gan.ai is building AI-powered video creation tools. But without access to compute, these projects are still stuck on the deck.
Gnani, for one, is already building agentic AI systems. It recently launched Inya.ai, a no-code Agentic AI platform designed to help developers and enterprises deploy intelligent voice and chat agents across customer-facing workflows, without writing a single line of code.
But with no access to the subsidised GPUs, Gnani might also go through delays for their ambitious plans of building voice models for Indic languages.
And yet, MeitY is reportedly preparing to announce a second batch of startups soon—four more companies that will focus on building small language models (SLMs), people familiar with the matter told AIM.
So, What Are We Doing?
The IndiaAI Mission aims to build sovereign AI capabilities, investing in foundational and vertical-specific models tailored to Indian languages, culture, and industries. For startups, it is not an easy task since most of the innovation in Indic languages is very hard to scale and earn customers from.
For this, the government is building a 34,000-GPU compute backbone to help startups selected in this philanthropic journey, with 15,000 more GPUs recently added.
Companies like Cyfuture, Netmagic, Sify, and Yotta have been empanelled to source a range of NVIDIA, AMD, and Intel chips, including the highly sought-after H100s and MI300Xs.
We reached out to Yotta for clarification on GPU access for startups, but have yet to receive a response.
Last year, Yotta planned to scale up its GPU inventory to 32,768 units by the end of 2025. Last year, the company announced that it would import 24,000 GPUs, including NVIDIA H100s and L40S, in a phased manner.
In an earlier interview with Forbes, Yotta’s chief, Sunil Gupta, said that India could build five GPT-4 models simultaneously using Yotta’s existing infrastructure. “I have ordered 16,000 [GPUs], so if there are five customers each wanting to make a GPT-4, I can handle their load simultaneously,” Gupta said last year.
This was when it was very hard to find NVIDIA GPUs for AI workloads, which was then confirmed to AIM by Sarvam and Vizzhy.
Now that the access is promised, the scale, even now, remains meaningless without timely access since motivation alone won’t help the startups build AI for Bharat.
As of now, only Sarvam has made visible progress, releasing multiple models—Sarvam-M, Sarvam Translate, and Sarvam Samvaad—all available on Hugging Face.
The company also recently confirmed that it will open source its IndiaAI models under permissible licenses. “We are committed to building public interest use cases,” Sarvam co-founder Vivek Raghavan confirmed to The Economic Times.
This is in response to online criticism about whether taxpayer-funded models should be proprietary. Abhishek Singh, who leads the IndiaAI Mission, also confirmed this to ET later.
Union IT minister Ashwini Vaishnaw has been bullish. “Whichever sector they focus on, they must become among the top five in the world,” he said while announcing the first round of startup selections.
The vision is ambitious, but the execution, so far, is slow. The country has 34,000 GPUs on paper, and no one seems to know where they are.
The post IndiaAI, Where Are the GPUs? appeared first on Analytics India Magazine.
