Roadmap: The Swaraj Stack - Indian resilience in a connected world

India is one of the fastest-growing economies in the world, with over a billion people needing access to credit, healthcare and socio-economic systems, all of which are  evolving over new digital rails. Much of India’s ambitions such as universal financial inclusion, energy sufficiency, and the mammoth push against poverty, will be increasingly delivered on AI-mediated infrastructure. 

Which raises a new imperative: how should India build domestic AI rails in a world where foreign tech is embedded into everyday life and systems?

The past year was a stress test on the impact of global interdependence. Cross border policies on advanced chips and AI model access have been rewritten repeatedly. The Hormuz crisis sent Brent past $120, reminding India that it imports over 85% of its crude oil requirements. Hostilities on India’s borders made clear that critical systems are also targets. Each shock impacted a layer of a value chain India depends on but does not yet fully own: compute, energy, security.

The answer is not isolationism. Innovation, after all, has always been a borderless endeavor, which is a longtime belief held at Bessemer. No one nation builds AI alone, and India's growth is inseparable from global trade, capital, and talent. Not too long ago, India successfully tackled a systemic technical infrastructure issue at national scale.

In 2009, India needed to give its citizens an identity, but nothing on the market worked at the necessary scale or price. India’s citizens were invisible to banks, and welfare leaked by the billions. The answer was neither to buy a turnkey system nor to exclude global technology: Aadhaar, India’s biometric ID system, was assembled from vendor neutral, open source parts and tech; commodity sensors, open-source software, affordable servers, and designed from start to finish in India. Today, Aadhaar runs at 1.4 billion enrollments.

Next came payments, built atop Aadhaar. The Unified Payments Interface (UPI) was designed as a public protocol open to every bank and every app, so the ten-thousandth integration became as easy as the first. Today, the same QR code settles a ₹50 roadside purchase and a ₹2,000 dinner bill.

The results: ~50 years of financial inclusion compressed into 6, per the World Bank; over $30B in welfare leakage saved; and UPI now carries nearly half the world's real-time payments by volume.

And now the build for digital rails is carrying over to AI. These digital rails span the many layers of the enterprise and consumer value chain, forming a new stack for India (see why we invested in Sarvam, Mitigata, Seekho, Inrisk).

We call this the Swaraj Stack: a domestic resilience buildout of critical infrastructure for AI and India’s interconnected future. Swaraj (literally self-governance in Sanskrit), to us means mastery of one’s capabilities while learning from the world's. 

What follows is a defining position for the investment team investing in AI in India: where the necessities run deepest, and where the next decade's defining companies will be built. Each layer of the stack has its own technology, economics, and capital needs; each is nascent and will likely take a decade to mature; and each, we believe, is a generational opportunity for Indian entrepreneurs.

The Swaraj Stack 

 

Layer 1 - Energy

India's installed generation power capacity crossed 520 GW in early 2026 (in addition to ~80 GW of captive capacity), and the Chief Economic Advisor's National Generation Adequacy Plan projects it to roughly double to 1,121 GW by 2035-36 simply to meet base demand. This is a 600 GW addition over the next decade, dominated by 509 GW of solar and 155 GW of wind, with peak demand rising from 250 GW today to 459 GW. AI is the new line item on top of that, and it demands unprecedented levels of continuous, stable supply. A hyperscale AI cluster pulls 100 MW or more continuously - the largest such as xAI’s Colossus 2 draw more than a gigawatt. This is on the scale of a major city with near-zero tolerance for instability. 

 

 

The implication is a once-in-a-generation build of energy infrastructure. Datacenters today consume roughly 0.5% of India's electricity; by 2030 that figure is on a path to 3% as installed datacenter capacity rises from 1.5 GW today to 4.5-9 GW. The Chief Economic Advisor (CEA) separately estimates the grid needs 174 GW of dedicated storage by 2035-36, 80 GW of batteries and 94 GW of pumped hydro, to absorb the solar-heavy generation mix. AI's continuous, 24-hour load profile exacerbates that requirement. For instance, utility-scale solar will need to be paired with battery and pumped-hydro storage for daytime load, gas peakers for ramp, and eventually small modular nuclear reactors for steady base. With advancements in AI, we expect transformed grid intelligence software - the system that schedules, balances, and prices the load - to handle megascale and heterogenous energy sources.

At the same time, buildout of domestic AI compute capacity will require innovation at datacenter sites, such as transmission and transformer equipment to address global HVDC trends, and cooling to address denser rack equipment in India’s environment of ever increasing temperatures.

 

Layer 2 - Compute

Colliers projects $20-25B of additional datacenter capex flowing into India by 2030, S&P separately tracks more than $32B in datacenter investment announcements over the last two years. These investments will impact  several distinct markets within the AI data center economy: the developers who finance and build the facilities, the power, cooling, compute, network and storage providers, the operators who run the AI clusters, and the cloud and colocation platforms that sell capacity.

India’s compute layer has significant exposure to international supply chains. The core AI accelerator GPU silicon is designed and fabricated outside India. So is the networking fabric that interconnects GPUs into training and inference clusters. India's semiconductor imports crossed $20B in FY24 alone, up 18.5% year-on-year, with monolithic IC imports alone up 2,000% since FY16. To spur India-based innovation in semiconductors, the Indian government recently announced the India Semiconductor Mission 2.0, committing $13B in incentives over 12 years and targeting expansion of the ecosystem from fabs, assembly and packaging to full-stack chip design and local supply chain requirements - raw materials such as specialty chemicals and advanced manufacturing equipment. Deloitte projects that India's overall semiconductor market could reach $120B by 2030 from $45B today.

We believe that India’s local inference demands will trigger innovation in domestic supply chains - neo-clouds, custom compute and interconnect silicon, and orchestration software engines. Government tenders alone in defence, public sector banks, state cloud are a multi-billion-dollar demand floor that simply did not exist three years ago.

 

Layer 3 - Data

A country’s digital data generation scales super-linearly with the population size - this is where India has a massive advantage. UPI processes 22 billion+ transactions a month, Ayushman Bharat Health Account has crossed 850 million health IDs. India hosts one of the largest sets of population-scale, consent-anchored datasets across 22 official languages and hundreds of dialects. The government has started opening this up. AIKosh (an open-source platform for multilingual text and speech annotation, designed specifically for Indian language datasets), launched under the IndiaAI Mission in 2025, crossed 5,500+ datasets and 250+ models by December 2025.

The gap is still enormous. Indic languages are about 1% of the Common Crawl corpora frontier models train on; India is 18% of the world's population. Indic text on the open web sits closer to 0.1%, against 59% for English. Almost none of it is grounded in Indian regulatory, financial, or healthcare context.

Closing that gap are two interesting business archetypes: 

First, mining and refinement: companies that combine raw India-specific data (DPI, BFSI and healthcare corpora, Indic voice) with synthetic data and data marketplaces into training-grade datasets. We look forward to entrepreneurs building "Bharat Data Commons," indigenous, monetisable layers on AIKosh's scaffolding.

Second, labeling and annotation, where India's human scale and cost structure can be export drivers. The global market for data labeling and annotation was roughly $3.8B in 2024, heading to $17B by 2030, with video annotation growing fastest at 23%+ CAGR. Consider physical AI: robotics and multimodal video need orders of magnitude more labeled data than language models do. Wearable, egocentric data require manual intervention, where India’s abundant manufacturing and farm environments, and English-speaking annotators have a structural advantage. Winners will target specialty sectors such as electronics, and will compete on accuracy rather than price alone.

Layer 4 - Models

This layer of intelligence and reasoning, the kernel of all AI use cases and applications, is perhaps the most critical for India to get right over the medium to long-term (see why we invested in Sarvam). 

Deploying global frontier models in Indian datacenters, or even global open source models in air gapped environments solves for data residency. However, model sovereignty runs deeper.

First, as mentioned in the section on data above, Indic language material constitutes 1% of the Common Crawl corpora, and Indic text on the open web is just 0.1%. For government and regulated enterprise use cases, there is a vast corpora of Indian policy, regulatory, financial and healthcare data, both historical and cultural. For India’s intelligence layer to handle this context, country-native models need to be pre-trained on a high mix of local content.

Second, for mission-critical and regulated sectors, establishing data provenance is paramount. AI use cases in national security cannot afford pre-training data contamination and poisoning, data compliance issues or unfair bias. For example, the EU AI Act specifically calls out data governance as a requirement for high-risk AI providers.

Third and most importantly, for long term autonomy, India will need to control the very intelligence layer that will, over the next decade, run national infrastructure.

This doesn’t mean that India needs to develop frontier-class models from scratch. We don’t believe that the model layer in India is a copy of the global foundation model race, but rather a different race. Local architectures that win will be multilingual by default, voice-native, and tuned for reasoning over Indian context. India's banking and investment services sector alone spent $15B on IT in 2025, up from $13.2B in 2024 per Gartner, with AI inference growing fastest.

 

We believe that India will be a hybrid market constituting use cases for global frontier models, global open source models, and sovereign models. Nation-critical inference workloads will need to be served by sovereign models that are owned, pre-trained, and run domestically.

 

Layer 5 - Infrastructure software

Infrastructure software sits between the models and the applications, constituting the APIs, developer platforms, agentic frameworks, evaluation harnesses (see Sarvam Arya), observability stack, vector databases, routing and caching layers, and guardrails.

This is an area where India has not yet seen significant scale for the local market. As AI adoption increases, we believe opportunities will emerge to solve India-specific problems. For example, serving 1.4 billion users with multi-modal content spread across the country, including tier 3-4 cities, will require specific choices in the inference stack, involving the right combination of latency, throughput and cost. A rapid rise in voice AI based applications at population scale will likely require evolved telephony stacks.

Layer 6 - Applications and Physical AI

At the top of the stack are the consumption surfaces: apps and physical AI.

With more than 1 billion internet users and around 740 million smartphones, the next wave of India’s population comes online using vernacular-first, voice-first, and AI-native technology. Of the 78 million MSMEs registered on Udyam portal (up from 41 million two years ago), most remain undigitized; AI will likely emerge as the first credible way to serve all stakeholders economically. In addition, several thousand Indian enterprises now cross $50M in annual revenue, and most still run core operations such as document processing, compliance filings, customer servicing, on manual or legacy systems built for a pre-digital scale. AI-native document intelligence is emerging as the layer that finally makes large-scale digitization economical for this segment.

Physical AI is the other imminent wave: robotics, autonomous systems, and embodied AI across defence, mobility, and manufacturing. Indigenous production hit a record ~$18B in FY25, up 18%, with procurement policy explicitly favouring domestic systems. AI-led defence technology, including drones and autonomous surveillance robots, and battlefield counter-measures, are likely to see significant outcomes over the next few years.

The economic value at this layer is significant, but its durability depends on the layers below. A made-in-India application or robot built on imported compute, data, and models would export margin to whoever owns the foundational layers.

 

Cybersecurity

The AI era is introducing categories of risk that did not exist a few years ago. On the one hand, AI is accelerating threat vectors and compressing timelines to extract vulnerabilities. On the other hand, agentic applications come with new attack surfaces, including prompt injection, model theft, training-data poisoning, agent hijacking, and deepfake-enabled fraud. We believe that the Indian cybersecurity market needs to evolve to span all the layers of the Swaraj Stack, from operational/OT assets in energy, to hardened agents for consumer and enterprise facing applications. India is facing an acute shortage of cybersecurity talent, at a time when the nascent AI stack is one breach away from becoming someone else’s leverage (see our investment in Mitigata). India's mobile-first consumption surface such as banking apps, UPI rails, OTP-driven authentication has no real parallel anywhere else in the world, and there are significant opportunities in developing indigenous solutions required to defend it (see our investment in Protectt).

Policy

The Aatmanirbhar Bharat policy framework is shaping India’s self-reliance in industrial strategy, with Production Linked Incentive (PLI) schemes underwriting domestic chip and electronics manufacturing, and the new $10B Research, Development and Innovation (RDI) fund routing patient, low-interest capital into AI and silicon. Meanwhile, the Digital Personal Data Protection (DPDP) Act, the RBI's FREE-AI framework, the IndiaAI Mission's $1.2B outlay, the India Semiconductor Mission 2.0’s $13B in incentives, and the localisation requirements emerging across regulated sectors are introducing sovereign infrastructure requirements. Every enterprise that has to comply, every government tender that demands data residency, every regulated workflow that cannot legally run on offshore inference is demand creation for the layers in the stack.

Sectoral opportunities

The Swaraj Stack will manifest in several vertical sectors, each of which will vary in terms of the urgency and timeline of deploying indigenous solutions and the size of the opportunity. The heatmap below scores nine sectors across four dimensions: national security exposure, data sensitivity and citizen risk, cultural and linguistic dependency, and economic leverage risk. 

The top four sectors, Defence, BFSI, Healthcare and Space, represent the frontier of where we believe the Swaraj Stack will have highest adoption over the next decade.

For instance, space surveillance is essential for national security, and detailed imagery data is classified. Last week, with Skyroot Aerospace’s successful Vikram-1 launch, India became the third country in the world (after the US and China) to own a domestic orbital launch capability. Over the next few years, we expect Indian companies in space transportation (launch vehicles, propulsion), infrastructure (space stations, base stations), satellites, and services (AI-powered data and analytics, positioning, communications) to emerge.

Backing Swaraj-enabling founders 

2026 has already proved how interconnected global economies are, and the ways in which technology stacks must evolve in the new AI paradigm. This year is just the beginning of India’s own imperative to address critical build-outs in the AI stack over the next decade.  

Each layer in the Swaraj Stack has distinct character, capital requirements and economics: data, models and applications are digital assets that can be built in months, while energy, compute and robotics are physical assets that take years to develop. These layers are where the category-defining companies of India’s next decade will emerge - and where the building must start now.

Subscribe to Atlas to be the first to read our next installment on the Swaraj Stack. We will go into details of where we believe the biggest opportunities lie across the Swaraj Stack layers and sectoral opportunities. If you are building in these areas, we would like to hear from you. Email us at india_ai@bvp.com.