In February 2026, the Government of India’s AI Impact Summit hosted the launch of five Indian-built foundation models in a single day. Sarvam-105B from Bengaluru. A 17-billion parameter sovereign multilingual model from BharatGen. Three more from Gnani.ai, Fractal, and Tech Mahindra. The IndiaAI Mission, by then, was operating on a corpus that had risen past ₹20,000 crore between its initial allocation and the 2026–27 Union Budget.
It is reasonable to ask when the conversation that led to this moment actually began. The answer is on the Cypher stage in 2024.
Between Cypher 2024 in September and Cypher 2025 in October, India’s sovereign AI conversation played out across two editions of the conference in four distinct phases. A conceptual warning. A first private attempt. A state-backed national mission. A formal debate. By the time the policy answer arrived in February 2026, every piece of the argument had already been made on the AIM stage, in public, on the record, by named protagonists. It is one of the cleanest examples available of a conference functioning as a leading indicator of national technology strategy.
Figure 1: The five-milestone Cypher arc that traced India’s sovereign LLM conversation, from a 2024 conceptual warning to the February 2026 policy verdict.
Act 1: The conceptual founding (Cypher 2024)
Cypher 2024 opened the sovereign-LLM conversation at three distinct layers, in three distinct sessions, by three distinct categories of speaker. Together they functioned as the conceptual founding of what would become a national mission.
The warning came from Nikhil Malhotra. Tech Mahindra’s Chief Innovation Officer delivered a session titled “The era of Sovereign LLMs and its implications in the AI world.” Malhotra did not present a model. He presented a category. Tech Mahindra had already been working on Project Indus, the company’s indigenous LLM effort, and his Cypher 2024 framing put the strategic argument on the table: India would either build its own foundation models or it would be a permanent customer of someone else’s. That framing, at the time, was contested. It is no longer.
The first private attempt came from Vishnu Vardhan. The founder and CEO of SML presented “Building an AI Company from Bharat: The Journey of Hanooman AI.” Hanooman — SML’s multilingual, multimodal foundation model, developed in collaboration with IIT Bombay and unveiled at NASSCOM’s technology leadership forum earlier that year — was, at the time of Cypher 2024, India’s first major indigenous foundation-model attempt. It was open-source, it spanned models from 1.5 to 40 billion parameters, and it was named after a deity whose defining attribute is responsible power. Two years later, Hanooman has become one of many in a now-crowded field. But on the Cypher 2024 stage, it was the first.
The infrastructure case came from Sunil Gupta. The Yotta CEO presented “Powering India’s AI-First Ambitions With Shakti Cloud,” the Hiranandani Group’s GPU infrastructure platform. Gupta’s argument was structural: no sovereign LLM strategy is credible without sovereign compute. The model layer needs a substrate. Yotta was building that substrate — a domestic GPU cloud that Indian foundation-model labs could train on without depending on US hyperscalers for access to NVIDIA H100s. Sixteen months later, Yotta would be the data centre operator through which the IndiaAI Mission disbursed Sarvam AI’s government-subsidised compute.
A fourth Cypher 2024 session deserves mention: Vivek Raghavan presenting Sarvam AI’s voice-based agents work. Sarvam was, at that point, the application layer of the Indian AI stack — voice agents in Indian languages, deployable through telephony and messaging. It would later become India’s most prominent sovereign-model contender. But at Cypher 2024, the conversation was about applied voice work. The transition from applied AI startup to frontier-model lab would happen across the following year.
Read together, the four Cypher 2024 sovereign-LLM sessions covered the four layers of the stack that any serious indigenous AI ecosystem requires: the strategic argument (Malhotra), the foundation model (Vardhan), the infrastructure (Gupta), and the applications (Raghavan). No other Indian event in 2024 hosted those four conversations together. None did so before national policy caught up. The fact that AIM’s conference did is the first piece of evidence for the leading-indicator thesis.
Figure 2: The fourteen named protagonists who built India’s sovereign LLM story on the Cypher stage across two editions, organised across model builders, infrastructure, policy, and the formal debate.
Act 2: The state-backed mission (Cypher 2025)
If Cypher 2024 framed the question, Cypher 2025 staged the answer. The single most consequential session was Professor Ganesh Ramakrishnan’s presentation of BharatGen: “Scalable Multilingual AI for Bharat.”
BharatGen, by Cypher 2025, was no longer just a concept. It was a government-funded sovereign multimodal LLM initiative, spearheaded by IIT Bombay under the National Mission on Interdisciplinary Cyber-Physical Systems of the Department of Science and Technology, executed by the TIH Foundation at IIT Bombay, and implemented by a consortium of premier institutes including IIT Bombay, IIIT Hyderabad, IIT Mandi, IIT Kanpur, IIT Hyderabad, IIM Indore and IIT Madras. It had been formally launched in September 2024 in Delhi by Dr Jitendra Singh. It was the world’s first government-funded multimodal LLM project. Ramakrishnan’s Cypher 2025 session was the first public stage on which the academic principal investigator presented the project to the Indian enterprise audience.
Around BharatGen, the broader Cypher 2025 sovereign-AI agenda deepened across every layer of the stack laid down the previous year. Amitabh Nag, CEO of Bhashini, presented “Voice-First AI: Powering Multilingual Access for Digital Bharat” — the government’s national language translation mission, now operating at scale. NxtGen Cloud Technologies took the presenting partner sponsorship slot and put three sessions on stage: A S Rajgopal as CEO speaking on “The Next Decade with AI: From Infrastructure to Intuition,” Abhishek Kumar Singh on “Inside M — Engineering India’s Agentic AI Platform,” and Abhisyant A showcasing “How NxtGen’s AIM is Building India’s Own ChatGPT.” The Indian sovereign cloud and agentic-platform category, which barely existed on the Cypher 2024 stage, anchored Cypher 2025’s programming.
Karan Kirpalani, Chief Product Officer of Neysa, contributed the framework piece: “GenAI’s Trilemma: Speed, Cost, or Control — Pick Two.” Kirpalani’s argument is the cleanest single articulation available of the strategic constraint that every sovereign AI builder is operating under: you can optimise for any two of speed-to-market, unit economics, or control over your stack, but not all three. Western-model fine-tuning maximises speed and cost. Building from scratch maximises control. India’s sovereign-AI strategy is, at its core, a series of bets about which two corners of that trilemma to anchor on.
And for the first time, policy voices joined the conference programming explicitly. Krishna Grandhi presented “Shaping AI Policy: India’s Tech Leaders Step Up.” Anuraag Saxena, CEO of the E-Gaming Federation, delivered “Owning the AI Agenda: India’s Policy Push.” Priyank M Kharge, Karnataka’s Cabinet Minister for Electronics, IT and BT, returned with “Karnataka’s Tech Vision: AI, Startups and Global Innovation Hubs.” Cypher 2025 was the first edition where the sovereign-AI conversation included the policy framing alongside the technology delivery.
Act 3: The debate (Cypher 2025)
Then, in the most consequential single session of the entire two-edition arc, Cypher 2025 staged a formal debate: “India Should Build Its Own AI Foundation Models — Not Rely on Western Ones.” Jason Joseph, Chief Information Security Officer at mPokket, argued for the motion. Srivatsa Subbanna, SVP and Head of Intelligent Automation and AI at Axis Bank, argued against. Both sides, by every public account, argued well.
Figure 3: A scorecard of the Cypher 2025 sovereign-LLM debate. Arguments paraphrased editorially. Both sides made defensible cases; the policy verdict arrived four months later.
The pro-sovereignty side made three arguments. First, data residency and control — the position that India’s economic and citizen data should not be the training corpus for Western models. Second, strategic autonomy — the view that a dependency on foreign frontier labs, however technically convenient, is a national-security exposure. Third, Indic-language fluency — the claim that no Western-model strategy can structurally serve India’s twenty-two scheduled languages with the depth that a domestic effort can.
The against-sovereignty side made three equally defensible arguments. First, capital efficiency — the position that building foundation models from scratch when fine-tuning works is a waste of national resources. Second, talent and ecosystem realism — the observation that India’s frontier-AI talent pool, while growing, is still smaller than the rhetoric admits, and that distributing it across multiple full-stack model efforts may produce no winners. Third, time to value — the practical reality that enterprise buyers need models that work today, not in 2027, and that an over-rotation toward sovereignty may delay actual adoption by years.
That this debate happened on the Cypher stage, in October 2025, between named senior practitioners from named institutions, with video on the record, is the single most powerful artefact in this article. It is the kind of debate that other publications would summarise in op-eds and panellists would conduct privately at dinners. Cypher did it in public. Both sides argued well enough that the audience, on the night, could plausibly have moved in either direction.
Act 4: The verdict (February–March 2026)
The policy answer arrived four months after the debate. At the India AI Impact Summit in February 2026, five Indian-built foundation models were launched: Sarvam-105B from Sarvam AI, BharatGen’s 17-billion-parameter multilingual model, and three others from Gnani.ai, Fractal, and Tech Mahindra. The IndiaAI Mission disbursed funding accordingly: Sarvam received approximately ₹99 crore in compute subsidy via 4,096 NVIDIA H100 SXM GPUs provisioned through Yotta Data Services. BharatGen received ₹989 crore in its third Mission tranche — roughly ten times Sarvam’s allocation, reflecting the BharatGen consortium’s status as a multi-institute academic mission rather than a single company.
In total, the IndiaAI Mission operated on a base corpus of ₹10,000 crore, supplemented by an additional ₹10,372 crore in the 2026–27 Union Budget. Sarvam AI was subsequently included in the NVIDIA Nemotron Coalition alongside Mistral, signalling that the international frontier-model ecosystem now treats at least one Indian lab as a peer rather than a regional player. BharatGen, separately, announced a strategic collaboration with IBM in September 2025 to expand commercial deployment of its Indic models.
Read against the Cypher 2025 debate, the policy verdict is unambiguous. The Government of India and the IndiaAI Mission have effectively answered the debate question in the affirmative: yes, India should build its own foundation models, and yes, public capital should flow to that effort at scale. The pro-sovereignty case won, in the only forum where the vote ultimately matters — the national budget.
Cypher as a leading indicator
There is a broader pattern in the timing that is worth naming.
Cypher 2024’s sovereign-LLM conversation preceded the BharatGen launch by weeks. Malhotra, Vardhan, and Gupta presented in September 2024. BharatGen was formally launched on September 30, 2024 in Delhi. The framing on the Cypher stage anticipated the national announcement that followed.
Cypher 2025’s sovereign-LLM debate preceded the policy verdict by four months. The formal debate happened in October 2025. The AI Impact Summit happened in February 2026. The Cypher stage hosted the argument before the IndiaAI Mission disclosed the answer.
Cypher 2025’s GenAI Trilemma framing preceded its mainstream adoption. Karan Kirpalani’s speed-cost-control formulation, presented at Cypher in October 2025, has since become a useful shorthand among Indian AI buyers for triage decisions on vendor selection. The Cypher stage is the first public artefact for that framework.
None of these are coincidences. They are the consequence of a curation strategy that treats a conference as a forward indicator rather than a retrospective showcase. The sessions that matter most on the Cypher stage are not the ones describing what has been deployed; they are the ones surfacing what is about to be deployed, contested, and decided. The sovereign-LLM arc is the cleanest case study available of that strategy paying off.
The 2026 question
What does the sovereign-LLM thread look like at Cypher 2026, then? Three predictions, based on the pattern of the previous two editions.
First, the conversation will move from “should we” to “which model wins.” The 2025 debate framed the strategic question. By Cypher 2026, with five sovereign-model launches behind it and the IndiaAI Mission well into its disbursement phase, the agenda will shift to model evaluation, benchmarking, and the practical question of which Indian foundation model an enterprise should actually build on. Expect a formal benchmark-comparison panel or a debate on enterprise-model selection. Expect, perhaps for the first time, an Indian-built foundation model evaluated head-to-head against a Western frontier model on stage.
Second, sovereign-LLM company representation will move from session-level to sponsor-level. ElevenLabs took a Cypher 2025 platinum tier as the first foreign frontier-model sponsor. The Indian sovereign-LLM companies — Sarvam, BharatGen, Gnani.ai — are now mature enough commercially to justify sponsorship slots. If at least two of them appear on the 2026 sponsor wall, the category will have completed its arc from concept to platform to commercial infrastructure.
Third, the policy conversation will intensify. The 2025 edition introduced Grandhi, Saxena, and Kharge as policy voices. The 2026 edition is likely to host a dedicated AI-governance track — not because the editorial team chose it, but because the regulatory framework around sovereign AI (data residency, certification, model evaluation, procurement) is now the binding constraint on deployment. The companies that win commercially over the next five years will be the ones that solve for governance, not the ones that have the largest models.
Cypher’s record on this thread is, frankly, better than its record on any other in the past decade. The conference identified the sovereign-LLM conversation when it was still an obscure concept, hosted its strategic founding before the national mission launched, ran the formal debate before the policy resolved, and now sits twelve months ahead of the question that the rest of the country has just begun to ask. Whether the same is true at Cypher 2026 will depend on which conversations make it to the stage. The arc so far gives every reason to watch closely.
