The emergence of advanced AI models like Anthropic’s Claude Mythos and Fable has spurred governments to action. The Reserve Bank of India earlier directed banks to disclose suspicious activities to CERT-In, concerned that the models may expose classified systems, critical infrastructure, and government networks.
Now, the central government has reportedly asked frontier AI startups Sarvam AI and BharatGen to develop cybersecurity models with capabilities similar to those of Anthropic’s Mythos, according to a recent ET report.
The concern is understandable. Governments increasingly want sensitive workloads to run on infrastructure they control instead of relying entirely on overseas providers. A domestically developed model could allow organisations to deploy AI inside isolated environments while maintaining greater control over data and inference
That said, it will not be easy for startups to build such a milestone model from scratch. The compute cost of sourcing tens of thousands of specialised GPUs is a big bottleneck, and so is developing massive proprietary datasets to train the models on. Such a monumental task would also involve advanced R&D to prevent AI from going rogue.
That’s why many startups are now proposing adapting existing open-source models instead of developing an entirely new one.
Moonshot?
“There is no chance of building a Mythos-like model from scratch,” an AI engineer at an Indian startup working on foundational models tells AIM on condition of anonymity.
The person says the more realistic approach would be to post-train an existing open foundation model, such as GLM, using extensive cybersecurity-specific reinforcement learning environments, adding that such an effort could be achieved for under $50 million.
Similarly, Pratik Desai, Founder and CEO of agritech startup KissanAI, questions the feasibility of building a Mythos-like model from scratch in India.
“Elon Musk with access to vast amounts of data, and some of the world’s top AI researchers, still can’t train a Mythos-like model and is probably nine months behind,” he remarks.
Meanwhile, Nikhil Malhotra, Chief Innovation Officer and Global Head of AI and Emerging Technologies at Tech Mahindra, believes that India does have the talent, research capability, and engineering expertise to build globally competitive AI systems. However, adapting existing foundation models can accelerate time-to-market for many cybersecurity applications.
According to him, the biggest challenge in building a Mythos-like model is bringing together high-quality domain data, AI infrastructure, and specialised talent. “Cybersecurity models require carefully curated datasets that capture real-world attack patterns, enterprise environments, and security workflows,” he explains.
With sustained investment, access to compute infrastructure, and a supportive policy environment, Malhotra expects a Mythos-like model can be built in 12–18 months and deployed at enterprise scale.
“Access to scalable compute remains a critical enabler, but success ultimately depends on combining robust infrastructure with deep cybersecurity knowledge and rigorous model evaluation,” he adds.
What IndiaAI is up to?
Access to large-scale AI infrastructure continues to remain a significant constraint for Indian startups. The IndiaAI Mission has committed around 62,000 GPUs to date, shared across the startups selected under the programme.
By comparison, SpaceX reported access to roughly 325,000 GPUs in its latest SEC filings. Despite those resources, its latest frontier model, Grok 4.5, is generally regarded as trailing Anthropic’s Mythos 5 on advanced cybersecurity capabilities.
India’s ambitions remain far more modest in terms of compute, but researchers argue they are enough to begin building frontier models. BharatGen has secured access to 1,400 GPUs to develop distributed frontier AI and plans to eventually build a 1 trillion parameter model.
“Our intent is to have these models before March 2027,” Ganesh Ramakrishnan, Founding Director of BharatGen and professor at IIT Bombay, told AIM in a recent interaction.
BharatGen’s roadmap begins with models in the 100 billion parameter range, followed by 500 billion parameter models, before eventually scaling to 1 trillion parameters as more compute becomes available.
Soket AI Labs is also pursuing a similar long-term vision. CEO Abhishek Upperwal recently said the company was not shying away from building a Fable-class model, with its next foundation model expected to launch in the coming months.
He argued that building a frontier sovereign model has become a strategic priority, as the AI race is increasingly being shaped by geopolitics, export controls, and national security concerns. “If oil doesn’t exist in your country, it doesn't exist. AI is different. India already has the data. We are the data factory of the world,” he earlier told AIM.
According to Upperwal, the missing ingredient is not money but a long-term commitment to research. “Hardcore R&D doesn't produce outcomes in 12 months.”
He also stressed that frontier model development demands massive, reliable compute infrastructure. “One thousand GPUs are not enough,” Upperwal said. “You need 10,000 GPUs working together.”
The economics of pricing frontier models also work against Indian AI ambitions.
Paras Chopra of LossFunk, a Bengaluru-based independent AI research lab, says the biggest hurdle for an Indian company developing a frontier model is that, right from the outset, it has to compete with the free plans from Anthropic and OpenAI, as well as Chinese open-weight models.
“If India is serious about the AI race, it needs to create a massive guaranteed market for domestic AI or else it’ll be economically irrational for any funder to underwrite a product in a highly competitive, already established market,” he wrote on X.
China Leads the Way
While India is still debating whether it can build a frontier cybersecurity model, Chinese AI companies are rapidly pushing the frontier with increasingly capable foundation models despite continued US export controls on advanced AI chips.
Moonshot AI recently unveiled Kimi K3, which it describes as the world’s largest open-weight AI model. Built using a mixture-of-experts (MoE) architecture and a 2.8 trillion parameter scale, K3 comprises 896 experts, with only 16 experts activated per token to improve inference efficiency.
The model supports a 1 million-token context window and introduces architectural innovations including Kimi Delta Attention to solve high memory and processing costs, Attention Residuals that replace fixed residual connections in transformers, an advanced routing architecture called Stable LatentMoE, Quantile Balancing to even out utilisation across hardware processors, and MiniTriton, Moonshot's custom compiler stack.
Kimi K3 outperforms OpenAI’s GPT-5.5 and Anthropic’s Opus 4.8 on several benchmarks but trails Claude Fable 5 and OpenAI’s GPT-5.6 Sol.
Meanwhile, Alibaba has also continued to scale its frontier AI efforts with Qwen3.8-Max-Preview, a 2.4 trillion parameter model released on its AI platforms ahead of a planned open-weight launch.
“We believe it’s one of the most powerful model available today, compatible to leading frontier AI models, second only to Fable 5 [sic],” Alibaba put in a post on X.
That said, US companies have also accused Chinese AI firms of using distillation, a technique wherein a smaller AI model learns by imitating the outputs of a larger, more capable model.
The White House recently accused Moonshot AI of distilling the Fable model to train Kimi K3. Anthropic also alleged last month that operators linked to Alibaba’s Qwen team carried out a much larger distillation campaign, involving approximately 28.8 million interactions with Claude through around 25,000 fraudulent accounts between April and June.
While building a Mythos-like Indian AI model seems to be a distant dream, for now, India’s real competitive edge may lie in voice-first interactions and deep vertical integration.
Your reaction