Image Source: Getty Images
The global race for artificial intelligence (AI) has become, in practice, a contest between two states. The Stanford AI Index 2026 reports that the leading American model was just 2.7 percent ahead of its Chinese counterpart as of March 2026. In 2025, private AI investment in the United States reached US$285.9 billion, more than 23 times China’s US$12.4 billion, a figure that likely understates state-directed Chinese funding. The material foundations of the field are similarly concentrated. In 2025, the US hosted around three-quarters of the world’s advanced AI computing clusters, while China hosted just 14 percent; one study also found that only these two countries host accelerators sourced from domestic suppliers. Talent follows the same pattern: MacroPolo’s Global AI Talent Tracker found that as of March 2024 the US is home to 60 percent of top AI institutions, and that researchers of American and Chinese origin make up 75 percent of top-tier talent within them.
Despite its economic weight and regulatory influence, the European Union (EU) has not converted these advantages into comparable capability. All major publicly reported European AI data centres combined appear to hold less computing power than the Amazon-Anthropic mega-cluster at New Carlisle, Indiana. Other states, including India, consequently depend on foreign models, chips, and cloud infrastructure with little influence over the rules that govern them.
AI Sovereignty and National Security
AI sovereignty has become central to national security as the technology finds widespread use in the information environment and on the battlefield. During the May 2025 India-Pakistan conflict, AI-generated fabrications and recycled footage circulated widely, threatening public trust between two nuclear-armed states, and in Ukraine, drones with onboard terminal guidance retain their accuracy despite electronic warfare interference.
AI sovereignty has become central to national security as the technology finds widespread use in the information environment and on the battlefield.
Dependence on foreign providers is therefore a security vulnerability rather than a commercial convenience, and ‘sovereign AI’ capability is becoming as fundamental to national power as military and economic policy. As frontier capability is concentrated to an extreme degree between the US and China, so access to state-of-the-art AI is dependent on their technology and foreign policy choices. The June 2026 US AI export-control directive on frontier models reinforced allied fears that Washington could withdraw access to cutting-edge AI at will. Recent analysis warns that reliance on foreign-controlled models is a serious exposure if frontier AI becomes important for defending critical national infrastructure (CNI).
Full-stack self-sufficiency, however, is virtually unattainable today for all states except the US and China in the current environment, since the extreme end-to-end cost of sovereign AI. It is, therefore, imperative that nations such as India focus on realistic milestones that enable the state to make strides in AI without risking excessive reliance on external factors.
Why Small Language Models?
Most states cannot attain full-stack self-sufficiency in AI, but they can still secure a meaningful degree of control over the use cases most critical to national security by isolating those applications and serving them with internally developed models. Small language models (SLMs), defined as language models with a comparatively low parameter count of roughly 100 million to 10 billion, offer a realistic means of doing so.
Minimising Foreign Hardware Dependency
The case for developing SLMs for security and defence applications rests first on hardware. They need markedly less hardware to operate, which matters for countries such as India with limited compute capacity. Reducing dependence on existing compute infrastructure is particularly pertinent because, as of October 2026, American firms supplied 38,000 of the approximately 45,000 graphics processing units (GPUs) in the IndiaAI Mission’s compute facility, with the exact source of the rest unannounced, a further layer of persistent external dependence within the national AI ecosystem. SLMs ease this exposure by not requiring state-of-the-art hardware for training, which reduces training costs and foreign reliance; thus, they lower the barrier to domestic development. This approach does not remove external dependence entirely, but it narrows the capabilities exposed to it and concentrates national effort where the security stakes are highest.
SLMs as a Low-cost alternative to LLMs
SLMs are typically built to perform specialised tasks efficiently in resource-constrained settings, including on-device deployment. They avoid the dedicated data centres needed to operate and sustain general-purpose large language models (LLMs), which further reduces infrastructure costs. In principle, running them on national or edge infrastructure with a cloud compute fallback for failures is viable; this would be an ideal arrangement for military operations at national boundaries and instrumental in keeping sensitive operational data under domestic control. For comparison, TinyLlama, a widely cited general-purpose SLM, was trained using a fraction of the computing capability and time consumed by Meta’s Llama 3.1, a 2024 LLM whose training metrics were publicly disclosed.
Table 1: Training comparisons between TinyLlama and Llama 3.1[1]
| Model | TinyLlama | Llama 3.1 |
| Total Parameters | 1.1 Billion | 405 Billion |
| GPU Model Used for Training | NVIDIA A100-40G (16 GPUs) | NVIDIA H100 80GB (up to about 16,000 GPUs) |
| Training Duration | About 90 days on 16 GPUs, on 3 trillion tokens (approximately 34,560 A100 GPU-hours) | 30.84 million H100 GPU-hours, on 15 trillion tokens |
| Cost per GPU hour (October 2026) | US$1.49 per GPU-hour (March 2026) | US$3.25 per GPU-hour (September 2026 market median) |
Source: Author generated
Existing research also indicates that small, fine-tuned models can match or outperform much larger ones on narrow tasks. This strengthens the case for identifying functions that warrant additional sovereign control on security grounds and for assigning them to purpose-built models designed, developed, and trained within the nation rather than to general-purpose systems developed and controlled abroad, which minimises the cost of achieving a certain level of operational sovereignty.
Strengthening Data Security
Developing SLMs domestically would also strengthen the protection of national security-sensitive data, such as battlefield footage, by allowing workloads to run on state-controlled infrastructure. Foreign state-of-the-art models are generally accessed through cloud application programming interfaces, which raises privacy concerns, and their use exposes data to the legal jurisdiction of the provider’s home state. Models of this size can instead run on national or air-gapped systems, keeping operational data, intelligence inputs, and prompts under domestic control. This matters most in military settings where connectivity may be contested, and in information warfare where analysts are required to classify propaganda or triage suspect claims over sensitive datasets without disclosing them to a foreign provider. A nationally hosted model also stays available if a foreign supplier withdraws access.
Developing SLMs domestically would also strengthen the protection of national security-sensitive data, such as battlefield footage, by allowing workloads to run on state-controlled infrastructure.
India’s SLM Environment
India has already identified focused, domain-specific SLMs with multimodal capabilities as a policy priority, as outlined in the white paper series of the Office of the Principal Scientific Adviser to the Government of India. Firms such as Tech Mahindra, Genloop Intelligence, and Shodh AI are reportedly developing SLMs for various domains and use-cases and the IndiaAI Innovation Initiative, which invited proposals from startups, researchers, and entrepreneurs to build foundational models trained on Indian datasets, has contributed substantially to the present ecosystem. The IndiaAI Mission has additionally enabled smaller developers to access compute and data through the IndiaAI Compute Portal and the AIKosh dataset platform.
Sustained investment in domestic hardware access, the development of original base models, and rigorous evaluation across Indic languages, alongside secure deployment for defence applications, would reduce external dependence and strengthen national defence.
India’s SLM ecosystem is thus reasonably well positioned but needs considerable improvement to deliver a meaningful degree of AI sovereignty. External dependence remains significant, most notably in hardware, where American firms supply all of the IndiaAI Mission’s compute capabilities, and several prominent Indian models are adapted from foreign architectures or base models rather than developed independently. This matters for national security as well as industrial policy, as advanced cybercrime and information warfare increasingly target the data, languages, and public trust on which defence decision-making depends, and foreign reliance exposes sensitive workloads to supply disruption and external control. Sustained investment in domestic hardware access, the development of original base models, and rigorous evaluation across Indic languages, alongside secure deployment for defence applications, would reduce external dependence and strengthen national defence.
Pranoy Jainendran is a Research Assistant, Centre for Security, Strategy and Technology, ORF.
Disclosure: The author acknowledges using Sonnet 5.5 to refine language for clarity.
The views expressed above belong to the author(s). ORF research and analyses now available on Telegram! Click here to access our curated content — blogs, longforms and interviews.