India's AI Ambitions: Balancing Innovation, Accessibility, and Cultural Relevance

How India can leapfrog AI innovation & what DeepSeek's success can teach us

China's DeepSeek has showcased that AI innovation isn't reliant on massive resources but on smarter engineering. India stands at a pivotal point to leverage its ingenuity and talent to lead in AI, focusing on efficient models and cultural insights. Government and private sector collaboration is crucial. Marbella, Spain: Silicon Valley recently got a rude awakening - from China. Turns out, AI innovation isn't exclusive to overpriced lattes and hoodie-wearing billionaires. DeepSeek used older chips, less energy and homegrown talent, proving that AI breakthroughs don't require massive resources, just smarter engineering. Even Donald Trump called it a wake-up call for America, a sign that the West's dominance in AI is no longer assured. But this wake-up call isn't just for the US. It's also an opportunity for India. DeepSeek proved AI leadership isn't about data centres or billion-dollar war chests. It's about ingenuity and lean engineering. India has both of these in abundance. If it moves decisively, it can not only catch up, but also race ahead. For too long, AI has been framed as a battle between China and the US. Neither should define its future. China's AI is built for surveillance and military control. Chinese Communist Party ensures every breakthrough serves the state, whether for monitoring citizens or expanding influence. AI in China is a tool for oppression. Silicon Valley has turned AI into a gold rush-except the prospectors are VCs, and the only thing getting mined is our data. They'd sell their souls at the right valuation. Even OpenAI, once an advocate of transparency, is now a closed system prioritising Sam Altman's financial interests. China wants AI for control. The US wants AI for profit. India? It needs to build AI that actually improves lives, preferably without a dystopian side effect. DeepSeek's success is a model India can refine and surpass. China has quickly built an AI ecosystem with at least 10 labs on par with OpenAI, and dozens more closing the gap. Companies like Qwen (Alibaba), MiniMax, Kimi and DuoBao (ByteDance) are advancing AI through efficiency, not brute force, a strategy India can adopt and refine. AI isn't rocket science. India should know. It landed a craft on the moon for less than the budget of many Hollywood movies. In fact, it was India-origin Ashish Vaswani who, in 2017, co-developed the foundational Transformer model, the very architecture that powers today's AI revolution. His work proved that AI's future lies not in brute-force computing but in smarter, more efficient models. Many labs develop models with teams of fewer than 50 people. Nvidia's tools have simplified the process. The real challenge is training AI across thousands of computers efficiently, something DeepSeek mastered by optimising computing power and reducing waste. Unlike Silicon Valley, which kept throwing more computing power at AI problems, China has focused on making AI leaner and smarter. India can do this. Its engineers built Ola, Swiggy and Flipkart, platforms as good as anything in the Valley. The country produces top-tier software talent. But it lacks foundational AI models because it has no protected space to nurture and scale them. The US dominates because it has the world's biggest technology firms. China succeeds because its government and corporations treat AI as a national priority. Even weaker Chinese models get state and corporate support. DeepSeek did this in just two years, proving that AI success isn't about money or degrees, but about execution. If AI were a cricket match, India has the players. It just needs to step up to the crease. It needs institutional backing and the will to compete. DeepSeek also highlighted the importance of linguistic and cultural training in AI. The company hired humanities graduates to refine language models, ensuring a deep understanding of Chinese literature and philosophy. Western AI models struggle in these areas because they prioritise Western data sources. China preserved thousands of years of literary history, while the West lost much of its Roman, Greek and Egyptian texts. As a result, Chinese AI models now outperform Western models in understanding their own culture and philosophy. India's intellectual tradition is vast, spanning maths, astronomy, medicine and philosophy. Vedas, Upanishads and Arthashastra offer deep insights into governance, economics and ethics. AI trained on these texts could unlock new frontiers in reasoning, ethics and innovation. India's multilingualism offers a unique advantage. But it must act deliberately to ensure AI reflects its culture and knowledge. China trained its AI models on both Chinese and English sources, giving them a significant edge over Western counterparts. India can train its AI on its own literature, history and philosophy, and regain the global edge it once had. GoI has launched IndiaAI Mission with a $1.25 bn investment, and AI for India 2030 to promote ethical and inclusive AI. It has also encouraged AI adoption across sectors like healthcare, agriculture and education. But private sector must match this ambition. China's success wasn't just government-driven. It was fuelled by a mix of state backing and aggressive private investment. India's corporations must step up, take risks and drive AI forward, instead of waiting for government backing. A shining example of this is Bhavish Aggarwal's Krutrim, which announced the release of new open-source AI models earlier this week. With ₹2k cr investment and plans for India's first GB200 supercomputer, Krutrim is laying the groundwork for India's AI independence. Effectively, Aggarwal has shown that India doesn't need to rely on Silicon Valley. More Indian entrepreneurs and corporations must follow this lead, investing in AI research, infrastructure and talent. India now needs to build AI titans, claim its data, and go from the IT back office to the AI powerhouse.

AI in India: Silicon Valley's push and China's proof of feasibility

Silicon Valley is encouraging India to develop a complete AI infrastructure, with the potential for smaller models already accessible. Indian software developers are key in this process, supported by government policies to create demand, infrastructure, and regulate lightly to boost AI's commercial viability and economic productivity.Silicon Valley is nudging India to develop a full AI stack, even as the Chinese have demonstrated frontier technology is not prohibitively expensive. The latest exhortation comes from OpenAI's Sam Altman, who for DeepSeek's sake, has serious skin in the AI game. Like any technology, intelligence costs are declining, and Indian entrepreneurs will get over the entry hurdle at some point. Smaller models are already within reach, and there is enough local computing capacity to bring them to market. That's where things get fuzzy. Beyond the hype around AI, business is yet to see ways to make money from it. The technology is dispersing through enterprise-level application. Breakthrough in adoption will come when AI shows results at the product or service level. With current levels of data collection about producers and consumers, product and process innovation through AI is only a matter of time. GoI has a role here in creating demand for AI and setting up infrastructure around it, as tech entrepreneur Vivek Wadhwa in his article on this page today underlines. It has to pick up some of the cost of tech development. A proactive policy towards governance use-cases should speed up commercial development of large local models. The base layer of India stack must be a public one supporting commercial applications. The state's role in any dual-use technology is predetermined. In the case of AI, that extends to driving down computing and energy costs, and democratising access through skilling. Besides, regulation will need to be light to encourage commercial exploitation of the technology. India happens to be on the right side in the AI proliferation arena. It has access to technology and bandwidth to process it. Indian software developers are working on frontier AI both here and in the US. Generative AI can bring a fundamentally different approach to the country's development pathway. Rather than wait for AI to solve for India, the country must shape it to deliver on specific parameters, such as governance delivery and economic productivity.

India's 8-GPU gambit: Shivaay, a foundational AI model built against the odds

India's AI ecosystem is at a critical juncture, striving to balance its aspirations for technological self-reliance with the realities of limited resources. Amid this landscape, Rudransh Agnihotri, a 20-year-old entrepreneur and founder of FuturixAI and Quantum Works, has emerged as a prominent figure. His team's creation, Shivaay, a 4-billion-parameter foundational AI model, has sparked both excitement and controversy. While some hail it as a breakthrough in India's AI journey, others question its originality and scalability. "Building Shivaay wasn't just about creating another language model," Agnihotri explains. "It was about proving that even with limited resources, India can build something meaningful for its own people -- something that understands our languages, our culture, and our problems better than any foreign model ever could." Developing Shivaay was no small feat for Agnihotri and his team of college undergraduates. Without venture capital funding, they had to rely on programmes like NVIDIA Inception and AWS Activate to access GPUs. Also read: DeepSeek data breach: A grim warning for AI security "We trained Shivaay on just eight NVIDIA A100 GPUs," he reveals. "It was challenging because we didn't have the luxury of bleeding-edge hardware like OpenAI or Anthropic. But it was also an incredible learning experience that forced us to innovate." The team adopted techniques inspired by Meta's Lima paper, which emphasises using stylised prompts over massive datasets. "We realised that quality matters more than quantity," Agnihotri says. "By focusing on properly stylised prompts and synthetic datasets annotated by GPT-4, we were able to optimise the training process significantly." Also read: What is Distillation of AI Models: Explained in short Despite these constraints, Shivaay achieved impressive results. The model scored 91.04% on the ARC-Challenge benchmark for reasoning tasks and 87.41% on GSM8K for arithmetic reasoning. Shivaay's release was not without controversy. Following the publication of its benchmark results on Reddit, critics accused the model of being a fine-tuned derivative of open-source architectures like Llama or Mistral. This scepticism intensified after a system ROM (refers to system-level configuration or preset instructions embedded within the AI model) leak revealed references to other models in Shivaay's internal code. Also read: Deepseek to Qwen: Top AI models released in 2025 Agnihotri is candid about the origins of Shivaay's architecture. "We built the architecture from scratch but incorporated knowledge from multiple models," he says. "This wasn't about copying; it was about learning from what works while tailoring it for Indian contexts." He adds that transparency has been a priority: "We've published evaluation scripts and benchmarks so anyone can test Shivaay themselves." The backlash has also been a learning experience for the team. "The criticism made us realise the importance of implementing robust guardrails," Agnihotri admits. "We're now adopting advanced techniques like FP8 quantisation to improve efficiency and reduce memory usage." One of Shivaay's defining features is its ability to process multiple Indic languages, including Hindi, Tamil, and Telugu. This focus on linguistic diversity sets it apart from global models like GPT-4, which often struggle with cultural nuances. "No foreign model understands a farmer in Nashik or a weaver in Varanasi," Agnihotri argues passionately. "An Indic foundational model ensures digital sovereignty while addressing local challenges in healthcare, education, and governance. "This vision aligns with India's broader AI ambitions under the IndiaAI Mission, which aims to develop indigenous foundational models by 2026. The government has allocated ₹10,372 crore for AI infrastructure and announced plans to deploy 18,000 GPUs nationwide. However, Agnihotri believes more needs to be done to make high-performance computing accessible to startups. "Yotta offers GPUs at $3/hour -- that's not affordable for most Indian startups," he points out. "We need subsidised access if we want to level the playing field." Agnihotri draws inspiration from China's DeepSeek, which recently developed a reasoning model that rivals GPT-4 but at a fraction of the cost. DeepSeek achieved this by optimising its training pipelines and leveraging older GPUs with advanced engineering techniques. "DeepSeek showed us that innovation doesn't always require cutting-edge hardware," Agnihotri notes. "They used mixed-precision training and auxiliary-free loss balancing to maximise efficiency and we're exploring them for Shivaay v2." He also acknowledges the challenges posed by India's talent gap in AI research. A recent study revealed that fewer than 2,000 engineers in India specialise in building foundational models from scratch. "This is why collaboration between academia, industry, and government is so crucial," he emphasises. Looking ahead, Agnihotri envisions Shivaay evolving into an agentic AI capable of structured reasoning and task-specific applications. The team is already exploring partnerships with IITs to develop use cases in agriculture and healthcare. Also read: DeepSeek is call to action for Indian AI innovation, says Gartner "Imagine ASHA workers using Shivaay to detect early signs of tuberculosis or farmers predicting pest outbreaks via satellite data," he says enthusiastically. However, he remains pragmatic about the challenges ahead: "India should focus on developing domain-specific models rather than trying to replicate GPT's scale." Agnihotri also hopes that initiatives like MeitY's foundational model scheme will encourage more startups to enter the space. "If we can build smaller, compute-efficient models tailored for specific industries, there's no reason why VCs wouldn't invest," he argues.

As Sam Altman Returns to India, Will OpenAI Offer Region-Specific AI Pricing?As Sam Altman Returns to India, Will OpenAI Offer Region-Specific AI Pricing?

Artificial Intelligence still remains a relatively new phenomenon in the country, unlike the internet. OpenAI CEO Sam Altman is set to visit India for the second time this Wednesday. His Asia tour has already seen several product launches so far and his upcoming visit to India is expected to bring key announcements tailored to the country's unique needs. India, with its vast user base, is a prime candidate for simpler, more accessible ChatGPT interfaces tailored to local needs. In this light, OpenAI's announcement last December about adding a phone-calling feature to ChatGPT is particularly exciting. It removes the need for an internet connection or high-end devices, meaning even users with basic flip phones or rotary phones will have access to AI assistance. "1-800-ChatGPT might seem like a silly gimmick, but the underlying principle is critical to scaling AI adoption," wrote Google Deepmind's Logan Kilpatrick on X. If implemented well, a feature like this will be a huge win for accessibility in developing countries. "The next billion AI users will not be on the existing UX's, they will be using text, email, and voice," he added, suggesting whoever lands this experience is going to win in a huge way. A mail sent to OpenAI did not elicit any response about the specifics of the phone-calling feature's rollout in India. "One of the most significant advantages of such integration is its ability to empower users," added Osama Manzar, founder and director of the Digital Empowerment Foundation. By enabling people to create content, search for information, and share ideas, through voice commands, ChatGPT could make technology more accessible to those with limited digital literacy. This would be particularly impactful for rural and non-internet users, opening up new opportunities for learning, communication, and accessing essential services. "However, there are critical challenges to consider. The first concern is the contextual relevance of the AI's responses. ChatGPT's backend systems may not always adapt well to the diverse linguistic and cultural needs of local communities," he added, commenting that misalignment could reduce effectiveness and drive users away. Manzar also raised concerns about data privacy, as users may unknowingly share information without understanding how it's stored or used. From 2008 to 2024, active SIM cards in India surged over threefold, surpassing one billion in a population of 1.4 billion, underlining the immense potential for AI adoption at scale in the country. From a regulatory perspective, OpenAI can introduce such a feature as long as it complies with the relevant data privacy laws in India. Important to note that AI still remains a relatively new phenomenon in the country, unlike the internet. Beyond chatbots, intuitive design is key to making it accessible and human-centric. Developing AI solutions tailored to India's needs and cultural context is imperative. "The key lies in leveraging existing platforms that are already deeply engaging Indian masses like WhatsApp, Google, and Facebook," said Manzar. Despite the success of ChatGPT, it is interesting to observe that apps such as WhatsApp and YouTube continue to be the most popular in the country. "It is crucial to focus on how we can mine and utilise data generated on these platforms to create AI solutions that are not dominated by foreign tech giants, who might appropriate Indian content for their own benefit," he added. While AI accessibility is improving, true adoption in India hinges on both usability and affordability. Simplified interfaces help, but without cost-effective pricing, AI may remain out of reach for many. The implications of high-cost and premium pricing models for advanced tools are huge in a price-sensitive country like India. DeepSeek is slowly changing the game for everyone, forcing the AI labs to introduce cheap and open-source alternatives or at least rethink their strategy going forward. For instance, OpenAI now offers ChatGPT Pro at $200/month and is rumoured to introduce plans up to $2,000/month due to high compute costs of advanced models. While economics often sees costs decrease over time. Many people responded to Altman on X, highlighting that in the current realm, $200 per month is comparable to salaries and average incomes in many economies outside the US - suggesting that AI subscription pricing cannot be the same globally. In India, for instance, the average monthly income is around ₹20,000. This puts into perspective that AGI is accessible to the common man. essential part of daily life for many in India and worldwide. More so, as ChatGPT continues to become an essential part of many people's lives in the country and the world, at least in the urban parts. Interesting to note, Altman recently conceded on the future of AI ultimately being open-source. "I personally think we have been on the wrong side of history here and need to figure out a different open-source strategy," Altman said in a recent AMA session on Reddit. "Not everyone at OpenAI shares this view," he added. Closer home, the key question remains: Can India achieve its own free, open-source DeepSeek equivalent? More so, with a focus on Indic datasets. Aravind Srinivas, CEO of Perplexity, urged that building a foundational model in the country is as important as building on the application layer. Much is written about how the internet is divided on this question. Infosys co-founder Nandan Nilekani and former People + AI head Tanuj Bojwani have advocated to really solve Indian problems through use cases. Ola chief Bhavish Aggarwal has announced Krutrim AI Lab and the launch of several open-source AI models tailored to India's unique linguistic and cultural landscape. This announcement aligns with India's broader AI ambitions. The government has officially called for proposals to develop homegrown AI models, marking a decisive push toward sovereign AI that can compete globally. Through the IndiaAI Mission, startups, researchers, and entrepreneurs are invited to build large multimodal models, large language models, and small language models, ensuring AI that is deeply rooted in India's languages and culture. Per the website, the government expects the models must be trained on diverse Indian datasets, comply with Indian regulations, and serve both public and strategic interests

Will Uniform Pricing for AI Models Work in India?

Artificial Intelligence still remains a relatively new phenomenon in the country, unlike the internet. OpenAI CEO Sam Altman is set to visit India for the second time this Wednesday. His Asia tour has already seen several product launches so far and his upcoming visit to India is expected to bring key announcements tailored to the country's unique needs. India, with its vast user base, is a prime candidate for simpler, more accessible ChatGPT interfaces tailored to local needs. In this light, OpenAI's announcement last December about adding a phone-calling feature to ChatGPT is particularly exciting. It removes the need for an internet connection or high-end devices, meaning even users with basic flip phones or rotary phones will have access to AI assistance. "1-800-ChatGPT might seem like a silly gimmick, but the underlying principle is critical to scaling AI adoption," wrote Google Deepmind's Logan Kilpatrick on X. If implemented well, a feature like this will be a huge win for accessibility in developing countries. "The next billion AI users will not be on the existing UX's, they will be using text, email, and voice," he added, suggesting whoever lands this experience is going to win in a huge way. A mail sent to OpenAI did not elicit any response about the specifics of the phone-calling feature's rollout in India. "One of the most significant advantages of such integration is its ability to empower users," added Osama Manzar, founder and director of the Digital Empowerment Foundation. By enabling people to create content, search for information, and share ideas, through voice commands, ChatGPT could make technology more accessible to those with limited digital literacy. This would be particularly impactful for rural and non-internet users, opening up new opportunities for learning, communication, and accessing essential services. "However, there are critical challenges to consider. The first concern is the contextual relevance of the AI's responses. ChatGPT's backend systems may not always adapt well to the diverse linguistic and cultural needs of local communities," he added, commenting that misalignment could reduce effectiveness and drive users away. Manzar also raised concerns about data privacy, as users may unknowingly share information without understanding how it's stored or used. From 2008 to 2024, active SIM cards in India surged over threefold, surpassing one billion in a population of 1.4 billion, underlining the immense potential for AI adoption at scale in the country. From a regulatory perspective, OpenAI can introduce such a feature as long as it complies with the relevant data privacy laws in India. Important to note that AI still remains a relatively new phenomenon in the country, unlike the internet. Beyond chatbots, intuitive design is key to making it accessible and human-centric. Developing AI solutions tailored to India's needs and cultural context is imperative. "The key lies in leveraging existing platforms that are already deeply engaging Indian masses like WhatsApp, Google, and Facebook," said Manzar. Despite the success of ChatGPT, it is interesting to observe that apps such as WhatsApp and YouTube continue to be the most popular in the country. "It is crucial to focus on how we can mine and utilise data generated on these platforms to create AI solutions that are not dominated by foreign tech giants, who might appropriate Indian content for their own benefit," he added. While AI accessibility is improving, true adoption in India hinges on both usability and affordability. Simplified interfaces help, but without cost-effective pricing, AI may remain out of reach for many. The implications of high-cost and premium pricing models for advanced tools are huge in a price-sensitive country like India. DeepSeek is slowly changing the game for everyone, forcing the AI labs to introduce cheap and open-source alternatives or at least rethink their strategy going forward. For instance, OpenAI now offers ChatGPT Pro at $200/month and is rumoured to introduce plans up to $2,000/month due to high compute costs of advanced models. While economics often sees costs decrease over time. Many people responded to Altman on X, highlighting that in the current realm, $200 per month is comparable to salaries and average incomes in many economies outside the US - suggesting that AI subscription pricing cannot be the same globally. In India, for instance, the average monthly income is around ₹20,000. This puts into perspective that AGI is accessible to the common man. essential part of daily life for many in India and worldwide. More so, as ChatGPT continues to become an essential part of many people's lives in the country and the world, at least in the urban parts. Interesting to note, Altman recently conceded on the future of AI ultimately being open-source. "I personally think we have been on the wrong side of history here and need to figure out a different open-source strategy," Altman said in a recent AMA session on Reddit. "Not everyone at OpenAI shares this view," he added. Closer home, the key question remains: Can India achieve its own free, open-source DeepSeek equivalent? More so, with a focus on Indic datasets. Aravind Srinivas, CEO of Perplexity, urged that building a foundational model in the country is as important as building on the application layer. Much is written about how the internet is divided on this question. Infosys co-founder Nandan Nilekani and former People + AI head Tanuj Bojwani have advocated to really solve Indian problems through use cases. Ola chief Bhavish Aggarwal has announced Krutrim AI Lab and the launch of several open-source AI models tailored to India's unique linguistic and cultural landscape. This announcement aligns with India's broader AI ambitions. The government has officially called for proposals to develop homegrown AI models, marking a decisive push toward sovereign AI that can compete globally. Through the IndiaAI Mission, startups, researchers, and entrepreneurs are invited to build large multimodal models, large language models, and small language models, ensuring AI that is deeply rooted in India's languages and culture. Per the website, the government expects the models must be trained on diverse Indian datasets, comply with Indian regulations, and serve both public and strategic interests

Sops, DeepSeek drive local firms to work on foundation AI models

The government's subsidised common compute facility and opening for grants and funding under the IndiaAI Mission have also spurred energy among startups seeking capital.The DeepSeek challenge combined with government sops have turbocharged local startups into building India's own foundational and fine-tuned AI models. Fractal Analytics is planning to open source its 'Ramanujan' reasoning model. Scientist Pranav Mistry's startup TWO AI, funded by Reliance and based in Silicon Valley, will be releasing the 'SUTRA-R0' model this week using enhancements from DeepSeek and its own dual-transformer architecture, bringing down model cost to just $3.5 million. The government's subsidised common compute facility and opening for grants and funding under the IndiaAI Mission have also spurred energy among startups seeking capital. Socket AI Labs, the maker of the 'PRAGNA' models, is looking for an immediate funding of $10 million to start with, and founder Abhishek Upperwal says his company is keen to apply for government incentives which could bring down the project cost by one-third. FuturixAI founder Rurdransh Singh said he will now aim to scale up the 4-billion parameter 'Shivaay' model to 32-billion with government aid. Corover.ai shall be releasing its first foundational 1-billion parameter model within a couple of weeks which can be built at a cost of Rs 1 crore at government-subsidised compute cost, founder Ankush Sabharwal told ET. Also Read: ETtech Explainer: What is DeepSeek, China's competitor to OpenAI? Others like Ola Krutrim and SarvamAI are also in the fray to apply for government incentives. Founders like Mistry, Upperwal and Aravind Srinivas of Perplexity AI did not agree with Infosys cofounder and non-executive chairman Nandan Nilekani and Tata Consultancy Services CEO K Krithivasan's view that India doesn't need LLMs. "The idea that India should only focus on the application layer is, frankly, a defeatist narrative -- it is like saying 'grapes are sour'," said Mistry. "I humbly disagree with this opinion, coined by Nandan and a few others. With our immense talent pool, we are more than capable of building frontier AI models. Those who argue otherwise underestimate both AI and India's potential." Building AI foundation is critical not only for cultural and societal relevance but also strategic and national importance in areas like governance, healthcare and education, he said. "LLMs do more than process words; they encode societal structures, ethics and values. Depending solely on external AI risks compromising our digital sovereignty and narrative," he said in an emailed response to ET's queries. TWO AI's SUTRA V1 is a 75-billion parameter foundational model trained on 50 languages including English, Korean, Japanese, Arabic, Hindi, Marathi and Gujarati. "Until the DeepSeek moment, many doubted models like SUTRA and whether such models could be built on a budget. While SUTRA still lags in English, for Indian languages, it stands far ahead in performance, speed, and cost-efficiency," Mistry said. The company is looking to raise $35-$50 million as it brings next iterations of the model. AI startup Fractal Analytics, which is headed for a listing on stock exchanges this year, is planning to open source its 'Ramanujan' reasoning model which beats OpenAI's o1 and was awarded the Meta HackerCup recently, cofounder and CEO Srikanth Srikanth Velamakanni told ET. Fractal has also released AI models like Vaidya, Kalaido and Marshall. "We want to focus on building on top of foundational models because pre-training is something which everyone can do," he said, adding that India should focus on building models for the world. "Not just Indian languages, we need large general-purpose models like DeepSeek and GPT. AI doesn't know borders. Every nation should build AI for the world just like nuclear or space research," he said. Upperwal of Socket Labs said India can't directly jump on to a DeepSeek-size model (615 billion parameters) as the company took nearly two years of AI training to build this breakthrough. "It has to be an iterative process which is bootstrapped on previous versions. We previously had PRAGNA-1.25B in four Indian languages and now are aiming for a 7B-10B-size model. This would take nearly 1 million GPU hours of training on H100s," he told ET.

'DeepSeek shows there is still room for better AI maths'

Hiranandani Group leverages China's DeepSeek AI model, hosted in India, to offer free AI services to billions. This cost-effective approach challenges the notion that massive GPU investment is essential for AI leadership, offering a potential solution for India's AI ambitions. China's frugal development of DeepSeek is an outcome to US' sanctions on supply of AI chips and an answer to American hyperscalers, private equities and AI startups that hoarding monies and chips cannot give them the lead in AI, said Darshan Hiranandani, CEO of Hiranandani Group, leading its real estate, energy, warehousing and data centres business. "They (OpenAI) garnered crazy amounts of money and scared everybody else saying, if you don't have more GPUs, then you cannot compete with our maths. We are taking all the private equity money for foundational models. So we are only going to be superior." He said that the belief that OpenAI's math was the only math for AI reasoning in the world is now dismantled. "You don't shake up the US tech market for nothing." "DeepSeek showcased that there is still room for better maths. Suddenly with sensible resource allocation, they said, wait a minute, I can serve 10 times more population than ChatGPT with the same hardware," the executive said in conversation with ET. He said India needed exactly a moment like this, to scale AI inferencing at much less hardware and solve for sustainable consumption of energy for AI. It's also a wake-up call for India to now come out with a better math for our own foundational models. Allaying the worries that large GPU investments of companies like Yotta may be at risk with DeepSeek's architecture, the executive said India will definitely need lakhs of more GPUs to serve the large market. "Until now, a few million consumers may be using AI as a productivity tool. We want to take it to billions." Sunil Gupta, CEO of Hiranandani's data center arm Yotta Data Services, compared the future of AI to India's broadband market which unlocked the digital revolution even at 1/10th of global prices. "Anything and everything you are giving the greatest quality at lowest price points, our demographics, our population which is highly digitally enabled will take care of scaling it. Just imagine the number of AI applications that can now be created at Rs50 subscription per month, if you are bringing lower cost of training and inferencing for a model," Gupta said. Yotta is launching a B2C ChatGPT-like mobile app 'MyShakti' powered by DeepSeek, hosted in Yotta's data centre in India on just 128 H100 GPUs and offered to public for free. It can readily scale up to 1,024 GPUs if demand surges, the executives said. "The goal here is to host China's model in Indian servers so that no customer data/prompts leaves local infrastructure. And then offer this to billions of Indian users for free," Hiranandani said. When asked about the already growing popularity of DeepSeek app, which has surpassed ChatGPT's downloads in India, he said that 9 out of 10 queries on DeepSeek are unanswered because their servers are overburdened. "As for ChatGPT, there is only very limited availability of their free content. You need to use their Pro plans which cost $20 a month." If US' export controls on AI chips were a concern on India's AI ambitions? He answered in negative. The limits of 50,000 GPUs per year are fairly high, Hiranandani said. Global cloud providers are allowed to put 7% of their total compute in one country, which can take care of the rest. "There is also a very low likelihood of this Biden circular making it through the Trump administration."

Inside IndiaAI Mission's Plan To Build Indigenous AI Model

The proposal envisages offering both direct as well as equity-based funding for the selected entity(s) that will undertake this mammoth task The launch of its latest reasoning models, DeepSeek-R1 and DeepSeek-R1-Zero, by Chinese AI company DeepSeek, earlier this month, upended the AI ecosystem globally. It also put a question mark on OpenAI's arguable leadership in the space. Without relying heavily on high-end graphics processing units (GPUs), DeepSeek managed to surpass existing closed source models on all key benchmarks. While the Chinese company managed to achieve this feat in just $6 Mn, OpenAI spent $100 Mn over the years to build its AI model GPT-4. Back home in India, the emergence of DeepSeek raised curiosity, ignited hopes that India can also build such models, and led to a debate on the need to have domestic foundation models. Amid all these, Union IT minister Ashwini Vaishnaw said that India would host DeepSeek on local servers. Along with this, the minister also announced that India would build its own large language model (LLM). Not long after, Centre's flagship IndiaAI Mission floated a proposal inviting applications from Indian startups, researchers, and entrepreneurs to collaborate on building foundational AI models trained on Indian datasets. "This initiative aims to establish indigenous AI models that align with global standards while addressing unique challenges and opportunities within the Indian context," reads the proposal. However, what is interesting about the proposal is that the Centre will offer both direct as well as equity-based funding for the selected entity(s) that will undertake this mammoth task. At the outset, under the direct funding route, the IndiaAI Mission will offer "milestone-based disbursements" in the form of a direct grant and compute credits for AI Compute. Thereafter, IndiaAI Mission may also infuse further funding into the selected entity(s) by "take equity through mechanisms that will be finalised through mutual consent and agreement". That said, the proposal outlines several key objectives for the potential state-backed LLM: The intellectual property (IP) for the envisioned AI model will remain with the developer entity(s). However, there will be a provision of a perpetual licence for use by the government for public use. How Will The Project Be Evaluated? As per the proposal floated by the IndiaAI Mission, the submitted applications for building LLMs will be shortlisted on multiple factors, including innovativeness of the approach, scalability and sustainability, financial viability, ethical considerations, among others. Besides, the Centre will also select applicants based on the capability of the teams, feasibility and impact of the proposed applications. As per a blog post on IndiaAI website, the submitted proposals will be reviewed by a panel of experts and then the shortlisted applicants will be invited for a detailed presentation. This follows IT minister Vaishnaw saying that the government has also selected 10 companies to supply 18,693 GPUs. This AI compute is also expected to be available to the startups that set out to develop an indigenously-built LLM. While it remains to be seen whether the potential foundational AI model sees the light of the day, hopes hinge on the frugality, innovativeness and talent pool of India's startup ecosystem.

AI Mission Takes Off Chip & Model Plans on Board

India will offer the cheapest compute in the world at less than $1 per hour for high-end chips that power generative artificial intelligence (GenAI) as the government's ₹10,000 crore IndiaAI Mission comes into play from Thursday, said Ashwini Vaishnaw, minister for electronics and IT. The government will also incentivise the development of local language models built by academia and industry with investment capital and other support, Vaishnaw said. The move is aimed at building up Indian language foundational model muscle. Proposals for model development were invited on Thursday. At least six startups and developers that can do it within the next 10 months have been identified, he said. ET was first to report on January 23 that India will back indigenous foundational models. "The real value will come from two things -- algorithmic efficiency and the quality of training datasets," Vaishnaw said, adding that Chinese AI company DeepSeek has proven to the world that a cost-efficient model can be developed. "DeepSeek was trained on 2,000 GPUs (graphics processing units)," Vaishnaw said. "We have now 15,000 high-end GPUs. (OpenAI's) ChatGPT version 1 was trained on about 25,000 GPUs. So this gives us a huge compute facility, something which will really give a boost to our ecosystem." GPUs are high-capacity chips needed to run complex AI development tasks. Vaishnaw said India is not late to the AI party and will play a big role in the innovation taking place in the field globally. Since India has now incentivised compute, the models will follow, he said. The rates for common compute in India are "phenomenally competitive", he said. The government will provide a 40% subsidy to those accessing GPUs. It will bring down the cost to below $1 per hour for GPU access, one-third of the global average of $2.5-3. Under the IndiaAI Mission, commitments have been received for 18,693 GPUs, out of which 15,000 are high-end GPUs and 10,000 of those are available (with the bidders) as of today, said Vaishnaw. "We put huge thrust on getting the common compute facility developed," the minister said. "This gives us a huge advantage vis-a-vis many other countries. Because now we will have this compute facility available for so many innovative and new architectures that people would like to create and test." The frontrunners in GPU capacity empanelment include Jio Platforms, E2E Networks, NxtGen Datacenter and Cloud Technologies, Locuz Enterprise Solutions and CtrlS Datacenters out of a total 10 companies offering 18,693 GPUs. Locuz is a partner of hyperscaler Amazon Web Services (AWS). The average bid price during the commercial round was ₹115.85 ($1.34) per GPU hour. Of this, 40% cost will be subsidised by the government over the next three years with a total budget allocation of ₹4,563.36 crore. Further, ₹689 crore has been earmarked for AI application development. "The second big mission of the IndiaAI mission was to develop an AI model," Vaishnaw said. "We have now created the framework which will be launched today. We are calling for proposals to develop our own foundational models, where the Indian context, the Indian languages, the culture of our country (are considered), where the biases can be removed... where the datasets, for our country, for our citizens... that process will start today." The minister expressed confidence in India's ability to develop such models. "We believe that there are at least six major developers who can develop AI models in the next six to eight months on the outer limit, and four to six months on a more optimistic estimate," he said. Vaishnaw said DeepSeek's AI model will be hosted on Indian servers. "Good thing is that DeepSeek is an open-source model and we are very soon going to host DeepSeek on Indian servers, the way we have hosted (Meta's) Llama, so that data privacy parameters can be addressed," he said. "Already, our team has worked out the details of how much capacity, how many servers are required." Vaishnaw addressed past scepticism about India's capacity to develop foundational models. "The innovations that are happening in the world are humongous," he said.

Krutrim-2 - Can India's language-first AI outpace global benchmarks?

Ola's Krutrim-2 represents India's boldest attempt to create culturally resonant AI infrastructure, blending strategic investments with open-source collaboration. While achieving notable progress in Indic language processing, the model confronts systemic challenges in balancing local relevance with global performance parity. The model's transition to a 12-billion parameter architecture marks a strategic leap from Krutrim-1, prioritising India's linguistic diversity through expanded capabilities: Also read: India's 8-GPU gambit: Shivaay, a foundational AI model built against the odds Rudransh Agnihotri, FuturixAI CEO, notes: "The 71% parameter expansion hasn't closed the MMLU gap with Mistral-NeMo - tokenisation inefficiencies in scripts like Devanagari appear to offset gains from Indic optimisation". Krutrim-2's 128K-token context window facilitates complex vernacular dialogues across 22 scheduled languages, yet three critical hurdles emerge. Tokenisation struggles with Brahmic script complexities - conjunct consonants in Devanagari and vowel diacritics in Dravidian scripts require algorithmic overhauls to improve translation accuracy. Simultaneously, reliance on synthetic datasets introduces grammatical inconsistencies, particularly in low-resource languages like Bhojpuri, where 38% of outputs showed tense agreement errors during testing. Also read: DeepSeek data breach: A grim warning for AI security The prioritisation of BharatBench over global benchmarks creates a 7.4% performance gap against Mistral-NeMo, sparking debates about calibration methodologies for culturally grounded AI. Agnihotri elaborates: "Grammar correction trails 7B models in our trials - this isn't about scale but rethinking Indic training paradigms". Krutrim's decision to open-source aligns with global trends, yielding tangible ecosystem impacts. The Chitrarth-1 vision-language model now processes Tamil shopfront texts and Odia manuscripts with 89% accuracy, while Dhwani-1 enables Haryanvi dialect speech-to-text conversions for rural telemedicine platforms. Also read: What is Distillation of AI Models: Explained in short Over 150 startups leverage these tools - Vyakhyarth-1 embeddings power vernacular search in agritech apps, and Krutrim Translate handles 10 million daily conversions. However, Agnihotri cautions: "Community innovation addresses tokenisation flaws but demands NVIDIA-tier compute many lack.". Cost efficiencies emerge through DeepSeek model hosting at $0.003/token, 60% cheaper than GPT-4, though adoption remains constrained to 25,000 developers - just 3% of India's AI workforce potential. With ₹10,000 crore committed through 2026, Krutrim's roadmap focuses on three key frontiers. First, the operational NVIDIA GB200 supercomputer will process 2 trillion Indic tokens by Q3 2025, becoming India's largest AI infrastructure. Second, the "Bodhi" AI chip series - optimised for Bharatiya language processing - aims for 2026 deployment alongside 1GW data centres. Third, the Shivaay training framework, which compressed 200B tokens onto 8 GPUs, seeks to democratise access for vernacular AI startups. Aggarwal asserts: "We're redefining efficiency metrics for Indian AI", though matching DeepSeek's 275 tokens/second processing speed remains contingent on algorithmic breakthroughs. Krutrim-2 embodies India's aspiration to craft AI that resonates with its linguistic soul, yet its success hinges on resolving the triad of tokenisation complexity, data quality, and benchmark calibration. As Aggarwal concedes: "We're learning to walk before we run", the model's evolution will test whether cultural specificity and technical universality can coexist in India's AI future.

AI mission takes off: Chip & model plans on board

Proposals for model development will be invited soon. At least six startups and developers that can do it within the next 10 months have been identified, he said. ET was first to report on January 23 that India will back indigenous foundational models. India will offer the cheapest compute in the world at less than $1 per hour for high-end chips that power generative artificial intelligence (GenAI) as the government's Rs 10,000 crore IndiaAI Mission comes into play from Thursday, said Ashwini Vaishnaw, minister for electronics and IT. The government will also incentivise the development of local language models built by academia and industry with investment capital and other support, Vaishnaw said. The move is aimed at building up Indian language foundational model muscle. Proposals for model development will be invited soon. At least six startups and developers that can do it within the next 10 months have been identified, he said. ET was first to report on January 23 that India will back indigenous foundational models. "The real value will come from two things--algorithmic efficiency and the quality of training datasets," Vaishnaw said, adding that Chinese AI company DeepSeek has proven to the world that a cost-efficient model can be developed. "DeepSeek was trained on 2000 GPUs (graphics processing units)," Vaishnaw said. "We have now 15,000 high-end GPUs. (OpenAI's) ChatGPT version 1 was trained on about 25,000 GPUs. So this gives us a huge compute facility, something which will really give a boost to our ecosystem." GPUs are high-capacity chips needed to run complex AI development tasks. Vaishnaw said India is not late to the AI party and will play a big role in the innovation taking place in the field globally. Since India has now incentivised compute, the models will follow, he said. The rates for common compute in India are "phenomenally competitive," he said. The government will provide a 40% subsidy to those accessing GPUs. It will bring down the cost to below $1 per hour for GPU access, one-third of the global average of $2.5-3. Under the IndiaAI Mission, commitments have been received for 18,693 GPUs, out of which 15,000 are high-end GPUs and 10,000 of those are live as of today, said Vaishnaw. "We put huge thrust on getting the common compute facility developed," the minister said. "This gives us a huge advantage vis-a-vis many other countries. Because now we will have this compute facility available for so many innovative and new architectures that people would like to create and test." The frontrunners in GPU capacity empanelment include Jio Platforms, E2E Networks, NxtGen Datacenter and Cloud Technologies, Locuz Enterprise Solutions and CtrlS Datacenters out of a total 10 companies offering 18,693 GPUs. Locuz is a partner of hyperscaler Amazon Web Services (AWS). The average bid price during the commercial round was Rs 115.85 ($1.34) per GPU hour. Of this, 40% cost will be subsidised by the government over the next three years with a total budget allocation of Rs 4,563.36 crore. Further, Rs 689 crore has been earmarked for AI application development. "The second big mission of the IndiaAI mission was to develop an AI model," Vaishnaw said. "We have now created the framework which will be launched today. We are calling for proposals to develop our own foundational models, where the Indian context, the Indian languages, the culture of our country (are considered), where the biases can be removed... where the datasets, for our country, for our citizens... that process will start today." The minister expressed confidence in India's ability to develop such models. "We believe that there are at least six major developers who can develop AI models in the next six to eight months on the outer limit, and four to six months on a more optimistic estimate," he said. Vaishnaw said DeepSeek's AI model will be hosted on Indian servers. "Good thing is that DeepSeek is an open-source model and we are very soon going to host DeepSeek on Indian servers, the way we have hosted (Meta's) Llama, so that data privacy parameters can be addressed," he said. "Already, our team has worked out the details of how much capacity, how many servers are required." Vaishnaw addressed past skepticism about India's capacity to develop foundational models. "The innovations that are happening in the world are humongous," he said. "Our country will play a major role. We have great software capability, we have a good innovation ecosystem. AI models will be distributed, that's for sure. What would be important is how we use AI." Asked about US export controls on AI chips, Vaishnaw said India is well within the 50,000 GPUs per year limit. Besides, India is a trusted partner across the world and the IndiaAI Mission will flourish because of this, he said. The Ministry of Electronics and Information Technology (MeitY) on Thursday made public the L1 (lowest) bidders for various types of AI compute units in the IndiaAI Mission GPU tender. They include Jio, NxtGen, Locuz and E2E, which offer access to Nvidia, AMD, Intel Gaudi 2 and AWS chips. CtrlS emerged as the L1 bidder in one type of AI compute unit. Yotta Data Services proposed to offer the highest number of GPUs at 9,216 followed by E2E Networks at 1,353 and NxtGen at 1,088. Although Yotta has proposed to offer the highest number of GPUs, it has not emerged as the L1 bidder in any type of AI compute unit. AS Rajgopal, chief executive of NxtGen Datacenter and Cloud Technologies, said the mission will enable the company to deliver superior performance at the lowest price globally. "We will now be offering up to 70% savings to enterprises of all sizes for AI compute (GPUs) and services. Our focus now is on world-class execution," he told ET. IndiaAI will identify and approve the eligible end users from academia, micro, small and medium enterprises (MSMEs), startups, research community, government bodies, public sector units (PSUs) or any other entity. The L1 bidder in each AI compute instance category will be called the preferred service provider for that category. When awarding projects to empanelled vendors, IndiaAI will direct all requests to the L1 bidder until its capacity is exhausted. A portal will soon be launched for enabling such access. The IndiaAI mission also includes an Application Development Initiative, with 18 selected projects focused on areas such as agriculture, climate change, and learning disabilities. Underlining the importance of responsible AI development, Vaishnaw announced the establishment of an AI Safety Institute and eight projects under the Indigenous Governance of AI, aimed at developing tools, frameworks, and processes for AI safety.

Are GPUs Essential for India's Indigenous AI Models?

Disclaimer: This content generated by AI & may have errors or hallucinations. Edit before use. Read our Terms of use India is planning to launch its "indigenous" foundational AI models as a part of the next phase of the IndiaAI Mission, IT Minister Ashwini Vaishnaw announced. Inviting proposals for creating large language models (LLMs), Vaishnaw disclosed that they would be trained on Indian datasets, "ensuring linguistic, cultural, and contextual relevance". The government invited proposals from startups, researchers, and entrepreneurs and announced several evaluation criteria for this purpose: The government outlined two funding options: Direct Funding and Equity-Based Funding for this purpose. While the former includes grants and AI compute credits with milestone-based disbursements, the latter comprises additional funding through mutual agreements. As per the government-issued press release, companies can also use co-financing options to secure additional funding from venture capitalists and angel investors, among others. This development comes days after the IT Ministry's Additional Secretary, Abhishek Singh stated the government's intent to compete with foreign foundational AI models like ChatGPT and Gemini, the Hindu Business Line reported. Singh remarked that these models provide "inappropriate responses" as they are trained on Western datasets and not aligned with Indian languages and situations. During MediaNama's discussion on 'Governing the AI ecosystem', Co-Founder and CTO of Ozonetel Communications, C. Chaitanya explained that while applications like ChatGPT could produce stories in languages like Telugu or Kannada, they aren't nearly as perfect as the English outcome for the same. He claimed that the dearth of Indian language datasets deprives users of receiving AI content with crucial cultural context. Other participants echoed this sentiment, arguing that while the government's datasets are usually clean, they fail to reflect lived experiences and older language styles. Moving forward, Vaishnaw claimed that the IndiaAI Compute Capacity, originally envisaged as housing 10,000 Graphic Processing Units (GPUs), will add 8,693 more GPUs to benefit researchers, students, and developers. Additionally, the government-provided 40% subsidy aims to reduce AI model computation costs to Rs 100 per hour -- lower than the global average of $2.5 to $3 per hour. For context, the current funding allocated to this pillar is Rs 4,563.36 crore, approximately 44% of the total budget for the mission. Given its significant outlay, during a previous discussion, MediaNama Founder Nikhil Pahwa questioned whether this allocation was justified. Adding on to the discussion, Chaitanya suggested large compute and data were not the only possible pathways in developing AI models. Different algorithms and India's strong mathematical capabilities could be leveraged to build such models, Chaitanya explained. "Our goal is to build a language model without using GPUs to prove that we don't need that much compute", he stated. However, another attendee remarked whether the 44% allocation would be the bare minimum considering the rising importance of "test time" or "inference time" computation that occurs when users interact with the model in real-time. Speaking of the government's role in building computing capacity, Ajay Kumar of Trumvir Law added that the government must incentivise the private sector and invest in computing within the country instead of relying on global services. Besides compute capacity, participants advocated for a greater emphasis on datasets, arguing that the availability of diverse and high-quality datasets drives effective AI development. During the discussion, Pahwa raised a concern about the USA and China spearheading the AI development race and India lagging. Responding to his query, Umang Jaipuria, an engineer based in San Francisco argued that it wasn't too late for India and that the country could have its own frontier model and still utilise OpenAI's model or even Llama. Further, speaking about India's capacity to develop foundational models, Chaitanya argued that while the process depends on existing technologies like transformers, GPUs, and datasets, India shouldn't replicate pre-existing models but create different models or leverage open-source models like Llama to add value to them. Answering media queries about foreign AI models, Vaishnaw stated that DeepSeek and other foundational models can be hosted on Indian servers following security checks to enable coders, developers, and designers to benefit from its open source code. Interestingly, he also hailed the Chinese AI model's progress terming it a "very, very powerful model", TechCrunch reported. This statement comes at a time when countries like Italy and Australia have expressed concerns over the lack of clarity on the application's usage of personal data alongside companies like OpenAI and Microsoft probing it over allegations of extra filtrating data using OpenAI's API.

DeepSeek signals Indian startups' deep AI dive

As the battle between OpenAI and DeepSeek escalates, some Indian start-ups have started integrating DeepSeek with offerings. But the tide turned within a few days as OpenAI released its much-anticipated model, o3-mini, its most cost-efficient reasoning model.The launch of the Chinese reasoning model DeepSeek-R1 has sparked a fierce war in the generative artificial intelligence (GenAI) ecosystem. The challenger is being challenged, and the world is waiting to see who emerges the winner in this battle of David versus Goliath. As the battle between OpenAI and DeepSeek escalates, some Indian start-ups have started integrating DeepSeek with offerings. But the tide turned within a few days as OpenAI released its much-anticipated model, o3-mini, its most cost-efficient reasoning model. Largely prompted by the DeepSeek-R1's success, OpenAI slashed the price of the model significantly, making it an attractive proposition. Sam Altman, founder, OpenAI, in a recent discussion on Reddit, revealed that while it is not a priority, they might look at an open-source strategy. Will o3-mini be able to trump DeepSeek R1? o3-mini and DeepSeek Prasanna Krishnamoorthy, managing partner, Upekkha, an AI fund and accelerator, told ET that the restricted amount of o3 available for free is a result of DeepSeek. "But it does not change much for enterprises since they might find it better to run DeepSeek open source. It is good for individuals as well, if it performs better than o1-mini," he added. OpenAI launched its advanced reasoning model o3-mini on January 31, which beats o1-mini in areas such as math and coding and at a fraction of cost of o1 at $1.10 per million input tokens and $4.40 per million output tokens. "Now, for companies, it is about choosing a path - closed or open source," said Arko Chattopadhyay, co-founder & CEO, PipeShift, a model evaluation and training platform. "Earlier people thought that open source will never catch up to closed source and that perception is now broken," he added. Two of PipeShift's customers are moving some of their operations to DeepSeek. Seeking utility Chattopadhyay said that one of its clients in e-commerce customer support, which requires heavy reasoning involving translations, are using DeepSeek as a primary model, keeping Llama as a fallback. Another customer in the insurance-tech area was using OpenAI's o1 to pilot document processing of customer data. They now want to go ahead with DeepSeek instead of using OpenAI's APIs since it can be hosted on their premises, Chattopadhyay said. Conversational AI company Yellow.ai, which works on a bouquet of models from US, India, China, and UAE, found that DeepSeek's architecture could cut GPU bill by 30-40%. "DeepSeek wins by a big, big margin from some other open sources," said Rashid Khan, co-founder and chief product officer, Yellow.ai. "With this open source availability, a larger segment of industries such as governments, healthcare institutions, banking and insurance, who were limiting AI adoption because of concerns around sending data to an OpenAI or a Cloud, now have an on-premise option," he said. Meanwhile, OpenAI's o3-mini, "is demonstrating a high level of developer freedom and will be productive in code generation and complex reasoning," Khan added. Abhishek Upperwal, founder, Soket Labs said, they have shifted to DeepSeek for extracting more data to train and fine-tune its real time speech API for Indic languages since it performs better than Llama 70B. According to him, using DeepSeek is not just about the cost, which he said was "dirt cheap," but also the right marriage of cost and performance. Some users of JOHNAIC, a private AI server developed by Bengaluru startup Von Neumann Computers, are experimenting with DeepSeek. The server costs '2-3 lakh and can host small AI models with up to 15 billion parameters. This meant that its customers would not be able to host powerful models of OpenAI and Meta that were heavy on memory. With DeepSeek that is changing. Founder Sasank Chilamkurthy explained that DeepSeek has distilled versions of its large 671B model in the range of 70B, 30B and 14B which can still beat the powerful models in math and coding. Challenges But when it comes to accessing DeepSeek and other Chinese models through APIs, enterprises are hesitant. Several governments like Italy and Australia have already flagged security concerns. Taiwan has banned its government departments from using DeepSeek citing security concerns. Yellow's Khan said geopolitical sanctions, maybe from the US, could be a limiting factor in adoption. "That's why hyperscalers like Microsoft Azure are rushing to host DeepSeek on their platforms."

The Need for Building AI for Bharat

As the world grapples with the impact China's DeepSeek has created, Indian companies have included the model in their offerings, balancing the race and also driving innovation in the country. But the question remains: is this enough when it comes to building AI solutions that cater to the needs of the country's larger population? Speaking with AIM at the podcast What's the Point, Shekar Sivasubramanian, CEO of Wadhwani AI, explained the exact definition and the differentiation between AI for India and AI for Bharat. There have always been two distinct Indias within India. In the current context, we can interpret them as the urban, tech-savvy 'India' and the rural 'Bharat' with its first-time smartphone users. This divide presents unique challenges in bridging the gap between technology and its diverse users. "The tech that is built around the world is mostly focused on India," Sivasubramanian said. "It already assumes a certain amount of familiarity, knowledge, education, and understanding of technologies that have been used thus far." Delivering to this community is easy. However, the challenges become more complex when it comes to delivering AI solutions or any technology to someone using a smartphone for the first time. For this population, the experience is entirely new, and as a result, there is a lack of trust, knowledge, and contextualisation of what they get in their hands - in this case, an AI chatbot. "To deliver to a farmer, you need to know what their free time in the afternoon is and where they will use technology. You never think of that when you're delivering in India," Sivasubramanian said. According to him, this is also coupled with the fact that a section of the Indian population consists of people who speak thousands of dialects and languages and are not accustomed to English. Nikhil Malhotra, chief innovation officer at Tech Mahindra, told AIM that his definition of 'AI for Bharat' aligns with Sivasubramanian's. While building Project Indus for Tech Mahindra, Malhotra and his team also visited different cities and towns in India to collect data for training models in Indic languages. "This is important because for anyone in India, their first thought process triggers...in their own language - that's AI for Bharat for me," Malhotra explained, emphasising that this goes beyond people who think natively in English. He added that this includes not just the rural population but also many homemakers in the country who are first-time smartphone users. For Ankush Sabharwal, founder and CEO of CoRover and BharatGPT, AI for Bharat means not just introducing AI to Bharat but also using AI to educate the Indian population about government policies, changes, and schemes. "We are what we are," Sabharwal said, emphasising that while educating the population in English and other skills is valuable, language should never be a barrier to access and opportunity. "Make everything natural and accessible [through AI]." This is where voice-enabled AI plays a crucial role. Malhotra explained that since most people naturally prefer speaking over typing or writing, building speech models tailored for the Bharatiya population becomes important. Similarly, Malhotra said that India's 5,000-year-old civilisation holds a wealth of valuable information that can be used to build better AI solutions for the population. He cited an example of how his team built an algorithm for Panchang, which not only works in India for accurately predicting the weather but also abroad, particularly in Sydney. This is also why teaching AI in schools becomes extremely important. Realising this, the CBSE has also introduced AI as an optional subject in schools. Sivasubramanian noted that the curriculum is so advanced that even experts in the AI field right now will find it challenging to complete. However, building public trust in AI is easier said than done. Sivasubramanian explained this with an example: if farmers are asked to click pictures of pesticides, they often capture not just the pesticide but also pictures of the surrounding area, including the sky, trees, and even their friends. "But you cannot restrict or constrain that. That is how they will learn and investigate the technology. We should help them instead of criticising them," Sivasubramanian said, adding that the most important thing that researchers should focus on is building trust with people. There is no better measure of the success of a technology than its acceptance by the people it aims to serve. "Keep everything else aside," Sivasubramanian said. He added that the most important way to do this is to talk to people with empathy and vulnerability. "We don't talk down to them just because we are technologists. We understand that their dignity and work are as important as ours. We sit down one-on-one with them as equals and chat and all the truths and problems come out." Sivasubramanian cited another example from when a researcher from Wadhwani AI was trying to explain an AI chatbot to a farmer who was not very interested in learning about it. After listening for a few minutes to the researcher, the farmer just said, "I do not understand whatever you are saying, but I understand that you care about me." The researcher thought he failed to explain the solution to the farmer. However, Sivasubramanian saw it differently. He said this was actually the researcher's greatest success because he had built a bridge of understanding. "The opportunity to serve somebody is the only gift that is long-lasting. Not fame, not money, not designation, not power," Sivasubramanian said. He believes the goal should not be earning billions of dollars but, instead, helping somebody and making their lives better.

India to launch generative AI model in 2025 amid DeepSeek frenzy

India is set to join the global artificial intelligence race and release a generative AI model sometime in 2025, Union IT Minister Ashwini Vaishnaw told reporters gathered at the Utkarsh Odisha Conclave. According to the Economic Times of India, the country has acquired 18,693 GPUs, including 12,896 Nvidia H100s, and is also looking at $20 billion in foreign investment in data centers over the next three years. The minister provided a timeframe for India's homegrown generative AI model that will be custom-tailored for the country's many languages and cultures: "We believe that there are at least six major developers who can develop AI models in the next six to eight months on the outer limit, and four to six months on a more optimistic estimate." Vaishnaw's announcement comes on the heels of the release of DeepSeek R1, an open-source AI model that performs on par with leading models from OpenAI yet reportedly only needed a fraction of the cost to train. Related: Microsoft probing DeepSeek-tied group over OpenAI data gathering method: Report The AI race heats up The release of DeepSeek R1 upended many long-held assumptions about artificial intelligence, including that scaling was a linear process that requires huge amounts of computing power. In response to the DeepSeek reveal, US President Donald Trump is considering tightening export restrictions on high-performance AI chips produced by leading AI chip maker Nvidia. The US government has already placed three major export controls on Nvidia sales to China, including an embargo on the H100 AI processor in 2022 and a ban on semiconductor component sales in 2023. Modified AI chips that featured degraded performance to stay compliant with the initial US sanctions on AI component sales to China, like Nvidia's A800 and H800, were also banned under the expanded restrictions. Trump has vowed to make the US the AI capital of the world and continue the country's dominance in the semiconductor and high-performance computing sectors. The US president recently announced project "Stargate," a $500 billion initiative led by OpenAI, Oracle and SoftBank to develop AI infrastructure in the United States. However, critics say tighter controls over US companies will make the country less competitive on the global stage and will erode its leadership in AI as smaller and more nimble competitors enter the field.

DeepSeek Becomes the World's Leading App, Driven by India's Surge

Indian IT minister Ashwini Vaishnaw recently announced that India will host DeepSeek on its local servers. DeepSeek's AI assistant has rapidly become the most downloaded mobile app worldwide, with India leading the surge in new users. According to sources, India alone has accounted for 15.6% of the total downloads since its launch earlier in January. DeepSeek is also the top app on Google's Play Store in the US, having held the lead since January 28. In just 18 days, it reached 16 million downloads -- almost twice the 9 million downloads OpenAI's ChatGPT had at launch. The app's rapid growth has impressed many but also raised concerns. Its low-cost model challenges leading AI companies, shaking stock markets and raising questions about the need for huge infrastructure investments in AI dominance. Krutrim, Ola's AI platform, is integrating DeepSeek models into its cloud infrastructure, as announced by founder Bhavish Aggarwal. "India can't be left behind in AI. Krutrim has accelerated efforts to develop world-class AI. As a first step, our cloud now has DeepSeek models live, hosted on Indian servers. Pricing lowest in the world (sic)," he posted on X. Indian IT minister Ashwini Vaishnaw recently announced that India will host DeepSeek on its local servers, boosting the demand for compute capability in the country. He also stated that India is set to offer the world's most affordable compute power, providing high-end AI chips for under $1 per hour. "DeepSeek was trained on 2000 GPUs," Vaishnaw said. "We now have 15,000 high-end GPUs. ChatGPT was trained on about 25,000 GPUs. So this gives us a huge compute facility, something which will really give a boost to our ecosystem." Microsoft and OpenAI are reportedly investigating whether Chinese AI startup DeepSeek improperly accessed and utilised data from OpenAI's models to develop its own AI system. This investigation centres on the technique known as "distillation", where a smaller model is trained using the outputs of a larger, more advanced model.

ETtech Explainer: How OpenAI is moving the needle with new 'deep research' tool

OpenAI launched "deep research" in ChatGPT for complex online research and o3-mini, a cost-efficient reasoning model. These releases follow the emergence of DeepSeek R1, signaling a race for cheaper, efficient AI. OpenAI on Sunday added 'deep research' agentic capabilities to ChatGPT and last Friday released o3-mini, the most cost-efficient of its reasoning agents, that "advances the boundaries of what small models can achieve", the company said. ET explains the two announcements that come on the heels of the stir caused by Chinese upstart DeepSeek. What is 'deep research' on ChatGPT? With deep research, ChatGPT can now carry out complex and multi-step research by parsing large amounts of online data. "Give it a prompt and ChatGPT will find, analyse & synthesise hundreds of online sources to create a comprehensive report in tens of minutes vs what would take a human many hours," OpenAI said on X. Some online users tested the feature for tasks ranging from analysing literary classics to producing research reports sometimes running over 10,000 words. It provides a "peek into the future of human-AI collaboration for knowledge work", said one user, while another said it is like having an almost-PhD-level researcher by your side. Also Read: Sam Altman on DeepSeek: Invigorating to have a new competitor Some found downsides too, including sources not always being cited and having to start over to make the model stop answering a query. Deep research is available to Pro users. Plus and Team users will be able to use it soon. What does OpenAI's new o3-mini model do? Compared to OpenAI's first reasoning model o1 for broader general knowledge reasoning, o3-mini provides a specialised alternative for technical domains while being faster and more efficient, OpenAI said in a blog. The model is available as API (application programming interface) for developers, as well as in ChatGPT for Plus, Team, and Pro users. Free users can also try it. The model is available in low, medium and high reasoning effort tiers for developers to choose from based on their requirements. How do these fare in terms of cost and accuracy? OpenAI says the o3-mini model is part of its efforts to drive down the costs of intelligence -- now 95% lower since GPT-4 -- while maintaining high-end reasoning capabilities. Compared to its predecessor o1 which cost $15 per million input tokens, o3 mini costs $1.10, discounted further at $0.55 per million input tokens if API requests are submitted as a batch. This is comparable to DeepSeek's R1 which costs $0.55 per million input tokens. Evaluating accuracy on the 'Humanity's Last Exam' benchmark, OpenAI's deep research model scored a new high - 26% - while DeepSeekR1 scored 9.4%. DeepSeekR1's score is better than o3-mini's predecessor o1 which scored 9.1%, but not as good as o3-mini that scored 10.5% and 13% in its medium and high versions, respectively. Unlike DeepSeek, OpenAI's releases are not open source for developers to freely build on or modify, however. What does this indicate about the impact of DeepSeek? The releases were spurred by the launch of open-source Chinese LLM DeepseekR1, which caused a stir in the market as it was built on much less compute power, using much fewer resources. "We will obviously deliver much better models and also it's legit invigorating to have a new competitor! We will pull up some releases," OpenAI CEO Sam Altman had said on X. This suggests a heating up of the competition for cheaper and more efficient models and the global AI race. With developers seeking such alternatives, OpenAI's closest partner, Microsoft, even included the R1 in its cloud platform Azure and code-hosting platform GitHub. Tech giants like Nvidia and AWS as well as Indian startup Krutrim are hosting the Chinese model on their AI platforms. Also Read: Sops, DeepSeek drive local firms to work on foundation AI models