Startup and VC Ecosystem Updates | Issue# 12 [February 16, 2025]
The DeepSeek Revolution
Recently, Chinese AI startup DeepSeek disrupted the AI landscape with its open-source Large Language Model (LLM) DeepSeek-R1, proving that AI startups can build competitive models without relying on vast amounts of high-end GPUs. This efficiency has drawn global attention, highlighting that AI breakthroughs do not necessarily require unlimited hardware / computational resources. The impact? Built in under two months with a small team and limited technical and financial resources, DeepSeek’s AI assistant – powered by DeepSeek-V3, the foundation of DeepSeek-R1 – displaced OpenAI’s ChatGPT to become the #1 app on the US app stores (Google Play Store and Apple’s App Store). More here. Some of the salient features of the development below.
- Low-Cost AI Development: DeepSeek has demonstrated that advanced AI models can be developed effectively with fewer GPUs and at a significantly lower cost. While AI giant OpenAI’s GPT-4 was trained on tens of thousands of GPUs (25,000 NVIDIA A100 GPUs as per OpenAI, 100,000 NVIDIA H100 GPUs as per a report by Inc42), DeepSeek used just 2,000 NVIDIA H800 chips – designed specifically for the Chinese market – and kept training costs under $6 Mn. Therefore, DeepSeek is able to offer access to its model’s APIs at a cost 95% lower than OpenAI’s latest models (including OpenAI o1), and this has disrupted the perspective of AI investors globally. More about DeepSeek’s new LLM here.
- Open-Source and Circumventing Restrictions: DeepSeek’s open-source nature and utilization of less powerful GPUs (H800) allows it to bypass restrictions on chip exports, offering a potential solution for countries facing similar constraints.
- A “Garage AI” Moment: DeepSeek’s achievements exemplify efficient development, showing that AI innovation is not solely the domain of massive tech companies, and prompting a possible shift away from the perception that AI development requires vast resources.
- Stock Market Impact: DeepSeek’s demonstrated capabilities have had a negative impact on market leader NVIDIA’s (and other semiconductor and tech companies’) stock prices, arising out of the fear that their market dominance may be threatened. See more.
Note: For the uninitiated, GPUs are specialized chips crucial for building deep learning and AI models because they process vast amounts of data quickly. During training, which is when the models learn from the data, GPUs accelerate complex computations, particularly matrix operations, allowing the models to learn patterns from data, such as recognizing objects in images or understanding natural language. For inference, GPUs enable the trained model to make predictions or generate responses in real-time.
India’s AI Ambition and the IndiaAI Mission
OpenAI (with models like GPT, o1, o3-mini, and more), Google (Gemini), Anthropic (Claude), Meta (Llama), Mistral AI (Mistral), Microsoft (Phi), xAI (Grok), Alibaba (Qwen), and Amazon (Nova) are all established players in the LLM space, but India is yet to produce its first home-grown foundation LLM. The country is actively planning to develop its own domestic foundation model / LLM under the INR 10,037 Cr IndiaAI Mission, focusing on models trained on Indian datasets. The Indian government is encouraging collaboration with startups, researchers, and entrepreneurs, providing funding to strengthen model training for both Indian or Indic languages and global benchmarks.
India aims to follow DeepSeek’s cost-efficient approach to AI development. India claims that it has exceeded its target of procuring 10,000 GPUs through public-private deals and has identified 10 companies that will supply 18,693 GPUs for developing its first domestic LLM. Out of those, about 15,000 are high-performance GPUs including NVIDIA H100s and H200s, and AMD’s MI 325X and MI 300X. More here. The government plans to distribute the GPUs to companies like Reliance Industries, Adani Group, and others working towards building AI infrastructure (data center capabilities) to tap the AI opportunity.
Challenges in GPU Acquisition and Infrastructure
India faces significant challenges in acquiring GPUs and establishing the necessary infrastructure, including developing its own GPUs and data centers, to support the creation of foundation models and their applications. Let us take a look.
- US Export Restrictions: The US government has imposed strict controls on the export of high-performance GPUs to India, allowing up to 1,700 NVIDIA H100 GPUs without a license. Beyond this limit, data center businesses and countries requiring large-scale computing infrastructure must obtain a Validated End User (VEU) approval, raising concerns over potential disclosure of sensitive business data. Additionally, VEU rules restrict GPU transfers even within the same geography, including limits on the number of GPUs that US-headquartered companies (AWS, Microsoft Azure, Google Cloud Platform, etc.) can deploy outside the US This creates a significant advantage for US-based cloud providers, who face fewer barriers to GPU access. With Donald Trump’s return to the White House, concerns over a US-first, protectionist tech policy have intensified, with anticipated tariff hikes and stricter immigration policies further impacting non-US companies. Experts warn that these restrictions will hinder India’s data center growth, while US tech giants can easily secure National Verified End User (NVEU) licenses to deploy GPUs at scale.
- Data Center Readiness: There are concerns about India’s data center infrastructure and its readiness to support the massive computational requirements of training AI models. For instance, Indian data centers are not designed to handle energy requirements of 100 kW or more and traditional designs are commercially unviable in such cases. There’s a need for data centers optimized for AI with improved energy efficiency, cooling systems, and high-speed networking. The current generation of data centers is not equipped to support the high-density racks typically required for hosting LLMs and handling intensive AI workloads.
- GPU Demand and Availability: India’s demand for high-performance GPUs is rapidly outpacing supply, further exacerbated by US export restrictions. The US government has capped GPU exports to India at 50,000 units – not per year, but in total – a limit that applies only to Indian companies, while US hyperscalers face no such restriction. As AI operations and applications continue to scale, this cap is likely to fall short. In late 2023, the Indian government estimated a need for 24,500 high-performance GPUs, but experts warn this figure is already outdated due to the rapid expansion of AI workloads. Additionally, while Indian companies are competing for limited resources, US-based players have unrestricted access, making it harder for Indian AI firms to build cost-efficient, scalable models. As a result, many Indian AI startups may be forced to rely on global cloud providers at significantly higher costs, impacting their competitiveness.
- Power Constraints: Indian cloud providers, such as C-DAC, Yotta, Mesa Networks, and Tata Communications, collectively have around 5,000-6,000 GPUs. In contrast, large-scale projects like Reliance’s planned 3GW data center would require the power equivalent of 50,000 NVIDIA Blackwell GPUs. Therefore, beyond GPU shortages, power availability in metro cities poses an even greater challenge, as large-scale AI development will demand massive energy consumption.
Self-Reliance and Indigenous Development
Experts, such as AI search engine Perplexity founder and CEO Aravind Srinivas feel that India should focus on building foundation AI models from the ground up, rather than relying on fine-tuning existing open-source models or proprietary models from companies big tech majors like Google, Meta, or OpenAI to create applications. There is promise in leveraging open-source models strategically, much like DeepSeek’s success in frugal AI development. However, most Indian AI startups still depend on fine-tuning existing models. For example, startups like Sarvam AI and CoRover use synthetic data to train existing AI models. This is primarily due to challenges, such as limited computing resources, lack of specialized expertise, and inadequate patient capital. While initiatives like Ola’s Krutrim and BharatGen signal early progress, they remain nascent compared to global players.
India’s massive, diverse, and largely untapped datasets are a key advantage that could give it a competitive edge. A major focus is and should be on developing foundation LLMs tailored to Indian languages and use cases, leveraging open-source models and local data to strengthen India’s position in the AI landscape. Companies like Krutrim AI Labs and Sarvam AI have already started working on India-specific foundation models.
Additionally, to ensure long-term self-reliance, reduce dependence on NVIDIA, and to avoid potential future sanctions, India should consider developing its own GPUs. China showcased a strong intention to innovate, execute with speed, and thrive under constraints and restrictions. The US embargo on China has pushed Chinese companies to develop cost-effective alternatives to NVIDIA’s GPUs. Companies like Huawei, Biren Technology, Cambricon Technologies, and Moore Threads are advancing in AI hardware, with Huawei’s Ascend 910B chips now rivaling NVIDIA’s older Ampere series A100 chips. While NVIDIA has since released more advanced GPUs, such as the Hopper series H chips and the latest Blackwell GPUs (B series), China is rapidly closing the gap. This raises the question of whether US sanctions could give China an edge in restricted markets and if EU nations, often impacted by US policies, might look to China for their AI infrastructure.
Grassroot-Level Adoption
India’s AI development should be centered on practical applications that address its specific challenges at the grassroot level, such as improving agriculture, streamlining supply chains, and enhancing governance, rather than prioritizing chatbot-based and generative application development. As this article from YourStory aptly puts it, instead of pursuing Artificial General Intelligence (AGI), the focus should be on Artificial Grassroots Intelligence – AI solutions tailored to serve India’s vast and diverse population. AI companies and startups can create significant value by focusing on building AI systems that drive real impact, from helping farmers increase yields to bridging healthcare gaps and scaling governance across the country.
Additionally, Indian tech companies, startups, and developers developing AI and its applications should work together to form a close-knit ecosystem and encourage open-source AI adoption. Innovation and problem-solving becomes easier with communities. For example, cloud services provider NxtGen is offering free public access to DeepSeek-R1 and pricing its cloud offerings 80% lower than major cloud platforms.
Threats and Concerns
Let us look at the key challenges India must address to achieve self-reliance in AI development and adoption.
- US Protectionism: There are concerns about US-first protectionist global technology policies, tariffs, and immigration policy changes. See US Export Restrictions under section Challenges in GPU Acquisition and Infrastructure for more.
- Dependence on External Suppliers: Over-reliance on foreign companies for GPUs could leave India vulnerable in the long run, especially as AI becomes a strategic asset. While the impact of DeepSeek has been significant, a greater concern is China’s aggressive investment in AI (includes 50 Tier II labs and nearly 10 anthropic labs), signaling a shift toward a new unipolar technological landscape. In this evolving scenario, where nations must rely on their own capabilities, India needs to focus on achieving strategic autonomy in AI and computing infrastructure.
- Increasing Costs: The import embargo may increase AI development costs for Indian companies, giving global tech giants an advantage. See GPU Demand and Availability under section Challenges in GPU Acquisition and Infrastructure for more.
- Risk of Being Left Behind: While the US, China, and the EU have already made strides into the AI arms race, experts are warning that India needs to act quickly and not fall behind other countries in the AI arms race.
- Brain Drain: India faces a significant brain drain challenge, with many of its top AI researchers and engineers contributing to advancements in US labs rather than driving innovation at home. To build a competitive AI ecosystem, India must create an environment that not only nurtures but also retains its talent, ensuring that the country’s brightest minds develop cutting-edge AI solutions for India rather than global tech giants.
Thoughts
OpenAI CEO Sam Altman once famously remarked that India should not even attempt to build foundation LLMs given the compute- and cost-heavy nature of such a project. Unsurprisingly, due to the high investment required, most Indian AI companies fine-tune existing models to build applications rather than developing foundation models, aligning with investor preferences. However, India has long demonstrated its engineering prowess across top AI and non-AI firms globally and has even reached the moon at a fraction of NASA’s budget. India’s global leadership in digital innovation is evident in the widespread adoption of UPI, its mobile-first payments network, and Aadhaar, its large-scale biometric identification system. Additionally, foundation models trained in countries like the US or China will reflect their respective values. Building its own foundation model would allow India to develop AI that aligns with its own values and culture. As this article by YourStory rightly highlights, India’s rich and diverse data, quality talent pool, resourceful and frugal mindset, and unique country-specific challenges provide distinct advantages that can propel the nation onto the global AI stage through its own set of foundation models.
That said, restrictions on GPU imports limit India’s ability to scale AI infrastructure, making it difficult to compete with well-funded global counterparts. However, DeepSeek’s success highlights that mastering AI and foundation models is primarily an execution challenge, driven by talent and speed. This offers a glimmer of hope – suggesting that India could carve out its own AI path through a resource-efficient and innovation-driven approach. To accelerate AI progress, India must foster an open-source ecosystem of developers, researchers, entrepreneurs, and indigenous infrastructure providers while also focusing on building specialized, high-performance AI computing systems, rather than relying solely on monopolized growth led by well-funded corporations.
On a global scale, VC investments in AI reached $110 Bn last year, accounting for 33% of total VC funding, according to Dealroom. Generative AI (GenAI) funding in particular reached new heights in 2024, with companies raising $56 Bn across 885 deals. A growing trend among investors is to back startups that develop GenAI applications rather than those building foundation models from scratch. This shift is primarily driven by faster returns on investment, as foundation models are costly and time-intensive to develop and are subject to regulatory uncertainties across geographies.
For India, the optimal strategy involves a dual focus: while developing foundation models is important, a strong emphasis must be placed on fine-tuning and building applications on top of them to capitalize on AI’s rapid progress, maintain relevance, and drive revenue growth. By leveraging established foundation models, startups can focus on delivering unique, value-added services tailored to specific market needs, enhancing their competitive edge. GenAI applications, in particular, offer scalability and faster deployment, allowing startups to adapt quickly to market demands without the burden of building complex AI systems from scratch. With the cost of accessing state-of-the-art models declining, AI has become more accessible, opening up opportunities for Indian startups to build impactful GenAI applications.
Beyond applications, emerging AI domains such as agentic AI (an evolving subset of GenAI) and the curation of novel synthetic training data – especially in Indian languages – are expected to attract significant investor interest. Additionally, as AI model training costs decline, there may be a renewed focus on foundation models beyond LLMs, including Small Language Models (SLMs) optimized for edge devices (such as phones, cars, and medical equipment) and Vision-Language Models (VLMs) designed for multimodal applications.
Looking ahead, significant advancements in novel AI architectures and robotics-integrated AI applications are anticipated, as highlighted by Meta’s chief AI scientist, Yann LeCun. While most AI progress today remains confined to digital domains, the next wave of innovation will focus on creating physical models of the world, bridging the gap between AI and real-world automation. These breakthroughs will accelerate progress toward AGI and intelligent robotics-driven applications.
Ironically, Altman too found DeepSeek-R1’s performance and cost-effectiveness to be “impressive”. Renowned tech investor Marc Andreessen famously referred to the model as “AI’s Sputnik moment.” The DeepSeek development marks a breakthrough moment in the AI race and India does have some catching up to do. Now is the time to build. With its strong engineering talent, resourcefulness, frugal innovation, and growing government support, it is only a matter of time before India establishes itself as a formidable force in the global AI landscape.
If you are interested to learn more, feel free to check out these coverages by Inc42 [1 and 2] and YourStory [1 and 2]. Additionally, here are some valuable insights from Deeplearning.AI on the major DeepSeek breakthrough.
Thank you for reading through! I genuinely hope you found the content useful. Feel free to reach out to us at [email protected] and share your feedback and thoughts to help us make it better for you next time.
Acronyms used in the blog that have not been defined earlier: (a) Venture Capital (VC), (b) Artificial Intelligence (AI), (c) Graphics Processing Unit (GPU), (d) United States (US), (e) Million (Mn), (f) Application Programming Interface (API), (g) Crore (Cr), (h) Amazon Web Services (AWS), (i) Kilowatt (kW), (j) Centre for Development of Advanced Computing (C-DAC), (k) Chief Executive Officer (CEO), (l) European Union (EU), (m) National Aeronautics and Space Administration (NASA), (n) Unified Payments Interface (UPI), and (o) Billion (Bn).