KEY TAKEAWAYS
1. AI is now infrastructure. $270B+ in capex, not R&D. The experimentation phase is over.
2. A two-track market is forming. Global hyperscaler models on one track, sovereign models on local infrastructure on the other.
3. The compliance era has a start date. Modi’s MANAV framework turns AI transparency into a policy directive, not an aspiration.
4. IT services are being unbundled in public. TCS-OpenAI, Infosys-Anthropic, Cognizant-Claude. Survival pivots, not experiments.
5. The most important layer was barely mentioned. $170B+ to compute, $15-20B to models. The layer between models and outcomes? Almost nothing.
6. India is a proof point, not just a market. 100M ChatGPT users. Claude’s second-largest market. 22 languages.
7. AI architecture is now geopolitics. India joined Pax Silica. Supply chain provenance is now a compliance requirement.
The through-line: competitive advantage is migrating from capability to architecture. The layer between models and enterprise outcomes is the least visible and most consequential part of the AI value chain.
Full analysis below.
Nearly 300,000 visitors. 100+ countries. $270 billion+ in investment pledges. Three sovereign AI models launched. One global declaration signed. And an entire technology ecosystem reshaped in six days.
The India AI Impact Summit 2026, held at Bharat Mandapam in New Delhi from 16 to 21 February, was the fourth in the global AI summit series following Bletchley Park, Seoul, and Paris, and the first hosted by a Global South nation. PM Modi, President Macron, UN Secretary-General Guterres, and every major AI CEO shared the same stage.
The headline investments are staggering. But headlines decay. Structure endures. I have distilled six days of proceedings, announcements, and behind-the-scenes signals into seven takeaways that matter for anyone navigating enterprise AI transformation, building the next generation of AI systems, or evaluating where the durable value will concentrate. Because the question is no longer whether AI will transform the enterprise.
The question is who will control the intelligence layer when it does.
01 | The Age of ‘AI as Infrastructure’ Has Officially Begun
The single clearest signal from the India AI Summit is that AI has crossed the threshold from experimental technology to core national infrastructure. This is not a metaphor. It is a capital allocation reality. Adani committed $100B for renewable-powered AI data centers by 2035. Microsoft pledged $50B for AI infrastructure across the Global South by 2030. Google announced $15B for its largest data center hub outside the US in Visakhapatnam plus new subsea cables.
Yotta committed $2B+ for Shakti Cloud using Nvidia’s latest Blackwell Ultra GPUs. Reliance unveiled Jio Intelligence with gigawatt-scale data center capacity at Jamnagar, with over 120MW expected online in the second half of 2026. L&T announced a proposed venture with Nvidia for sovereign AI factory infrastructure. And the Indian government earmarked $1.1B for a state-backed AI VC fund. When governments and conglomerates commit this level of capital, they are not betting on a trend. They are building the rails.
Every major investment announced at the summit, without exception, is infrastructure for production deployment, not research. Capital is migrating from R&D budgets to capital expenditure. That is the single most reliable signal of market maturity. If your AI strategy still lives in an innovation lab budget line, the market has already moved past you. (For the deeper analysis of why AI data center economics reshape competitive positioning, see my essay “From Steel to Silicon: Why AI Data Centers Will Define the Next Economy” in the Return on Intelligence newsletter.)
02 | Sovereign AI Is Not Nationalism. It Is Architecture.
India launched three sovereign AI models at the India AI Impact Summit: Sarvam AI’s twin models (30B and 105B parameters, named Vikram, both using Mixture-of-Experts architecture), BharatGen’s Param2 (17B parameter multilingual MoE model supporting all 22 scheduled Indian languages), and Gnani.ai‘s Vachana voice stack. Union Minister Vaishnaw outlined a “frugal, sovereign, and scalable” AI strategy. The government announced plans to add 20,000+ GPUs to the existing 38,000 base. But this is not just about national pride. It is about architectural necessity.
BharatGen’s Param2 is already deployed in production: MahaGPT powers governance workflows for the Maharashtra government, and an IFSCA deployment converts regulatory manuals into conversational interfaces for international investors. These are not lab experiments. They are operational systems serving real citizens and real capital markets. The structural insight: a two-track AI market is forming. Track One is the global hyperscaler stack (OpenAI, Anthropic, Google on their clouds).
Track Two is the sovereign stack (local models on local infrastructure with local data residency). This is not unique to India. It will replicate across every major economy with regulatory teeth. Enterprises in regulated sectors will need to operate across both tracks simultaneously. The layer that orchestrates this dual-track reality is the most undervalued position in the ecosystem. (This is precisely the shift from model races to system mastery I examined in “From Model Race to System Mastery: Winning the Trillion Dollar Vertical AI Revolution.”)
03 | The ‘Glass Box’ Era of AI Has a Start Date:
Now PM Modi presented the MANAV framework for AI governance at the India AI Summit: Moral and Ethical Systems, Accountable Governance, National Sovereignty, Accessible and Inclusive, Valid and Legitimate. The Delhi Declaration, adopted at the summit, codifies these principles. His emphasis on transparency, robust oversight, and verifiable AI systems was not aspirational language.
It was a policy directive at a 100-country summit. The “V” in MANAV means Valid and Legitimate: AI must be lawful and verifiable. This creates a regulatory trajectory that will cascade through India’s sectoral regulators. RBI for banking. SEBI for capital markets. IRDAI for insurance. DPIIT for industry. Every one of these bodies will need to operationalize these requirements into procurement standards. And this pattern will not stay within India’s borders.
The EU AI Act already points in the same direction. The Pax Silica bloc will follow. We are watching the formation of a global compliance norm in real time. The implication is clear: if your AI vendor cannot demonstrate how their system arrives at decisions, the compliance exposure is not hypothetical, it has a timeline. (For why measurement and grounding become the new competitive moat in this environment, see “The New Moat Isn’t Model Size, It’s Knowing What to Measure.”)
04 | The IT Services Industry Is Being Rebuilt in Public Vinod Khosla did not mince words:
“By 2030, there will be no IT services industry and no BPO industry left.” Provocative, but the incumbents’ response was more telling than the prediction itself. TCS signed OpenAI as its first data center customer under the Stargate initiative, beginning with 100MW of capacity and scalable to 1GW, a strategic shift from services to infrastructure ownership. Infosys partnered with Anthropic to deploy Claude models and agents in enterprises, beginning with telecom. Cognizant deployed Claude Code to 350,000 employees globally to modernize legacy systems.
These are not innovation experiments. These are survival pivots at billion-dollar scale. Microsoft India President Puneet Chandok’s “unbundling” framework gives this disruption a precise mechanism: “AI will not kill jobs. AI will unbundle jobs. Your job is a bundle of tasks. What AI will do is it will unbundle it.” Invoice matching. Resume screening. Claims processing. Regulatory reporting. This task-level automation is exactly how agentic AI systems operate. The disruption is not a future scenario. It is the present tense. (This is the agent washing dynamic I warned about in “Agent Washing: Why 40% of Projects Will Fail”: the gap between labeling something “agentic” and building genuine autonomous capability determines which transformations stick and which get cancelled.)
India’s Chief Economic Advisor V. Anantha Nageswaran framed the stakes precisely: “India can become the first large society where human abundance and machine intelligence reinforce, and not undermine, each other. The window is still open, but it is not indefinite.” The $250B Indian IT services ecosystem is bifurcating. On one side, companies pivoting to own AI infrastructure and the connective layer between models and enterprise outcomes. On the other, pure labor-arbitrage models facing an existential compression. The transition window is narrowing faster than most planning cycles assume.
05 | The Most Important Layer Was Barely Mentioned
Here is what I find most revealing about the India AI Impact Summit. Examine what was NOT announced. The summit validated three layers of the AI stack: compute (Nvidia, Yotta, L&T), models (BharatGen, Sarvam, Gnani), and applications (governance, healthcare, BFSI). But between models and enterprise outcomes, there is an entire layer that determines whether AI investments produce operating leverage or expensive friction.
That layer-the connective tissue between raw AI capability and reliable enterprise operations was conspicuously absent from the announcements. The capital distribution tells the story: Compute and Infrastructure: ~$170B+ (heavily funded, commodity dynamics in 3-5 years) Foundation Models: ~$15-20B (government-backed sovereign play plus global hyperscalers) The Layer Between Models and Outcomes: Minimal Domain Applications: ~$5-10B (fragmented, sector-specific, growing fast) This pattern is not new. In every prior technology wave, from cloud computing to mobile to the internet itself, the value eventually migrates from infrastructure to the layer that connects infrastructure to business outcomes.
Compute will become abundant. Models will proliferate. The question that will separate winners from the rest is: who builds the systems that make intelligence useful, reliable, and compliant inside the enterprise? That is the Work 3.0 question, and it remains unanswered at scale. (The full framework is in “Work 3.0: The Missing Layer in Every AI Strategy.”) When everyone is investing in the same layer of a technology stack, the strategic question is always: what is the adjacent layer that nobody is funding?
06 | India’s 1.4 Billion Are Not a Market. They Are a Proof Point.
Sam Altman revealed India has 100 million weekly active ChatGPT users, second globally after the US. Anthropic confirmed India is Claude’s second-largest market, with revenue doubling since October 2025. Close to half of all Claude usage in India involves coding and technical tasks: building applications, modernizing systems, shipping production software. This is not passive consumption.
This is production-grade adoption at a scale that surprised even the platforms themselves. Altman called India a potential “full-stack AI leader.” Ambani framed India as the primary arena for the Intelligence Era, pledging that Jio would “reduce the cost of intelligence as dramatically as we did the cost of data.” But what makes India uniquely interesting is not just scale. It is complexity. Twenty-two official languages. Massive income disparity. A governance system spanning panchayats and G20 tables.
If AI systems can work reliably across this complexity, across languages, literacy levels, regulatory frameworks, and infrastructure constraints, they can work anywhere. India is not just a market. It is the hardest test case for production AI. And the systems being built here will have global applicability. The transition from delivery center to primary revenue center for AI is already underway, Anthropic opening Bengaluru, OpenAI’s 100M users, and the $1.1B government AI fund are leading indicators, not lagging ones.
07 | AI Geopolitics Now Has a Supply Chain Dimension
India formally joined the US-led Pax Silica initiative on AI and supply chain security, becoming the tenth member of a coalition that includes the US, Japan, Australia, UK, South Korea, Israel, UAE, Singapore, Qatar, and Greece, among others. France and India simultaneously deepened their bilateral relationship with shared emphasis on sovereign AI. The geopolitical dimension of AI is no longer theoretical.
Pax Silica explicitly ties AI governance to supply chain security and economic resilience. For enterprises operating across these allied markets, compliance now extends beyond data privacy and model safety to include supply chain provenance, data lineage, and infrastructure auditability. India’s semiconductor ambitions add another dimension. Four plants are approaching commercial production in 2026, and Qualcomm recently completed a 2nm chip design tape-out at its Indian engineering centres in Bengaluru, Chennai, and Hyderabad, demonstrating that Indian teams now operate at the frontier of chip design.
The Semiconductor Mission 2.0 announcement signals that India is building capabilities across the full hardware-to-application stack. AI architecture decisions are now, unavoidably, geopolitical positioning decisions.
THE THROUGH-LINE
These seven takeaways from the India AI Summit 2026 share a single structural thread: the global AI market has crossed from “should we build?” to “who connects intelligence to outcomes?” Compute is becoming abundant. Models are proliferating. The question that will define the next phase is not who has the best model or the most infrastructure. It is who builds the systems that make intelligence useful, reliable, and compliant inside the enterprise.
This is what I call the Work 3.0 transition: competitive advantage migrates from capability to architecture, from technology to system coherence. The summit title tells the story. We have moved from Safety (Bletchley) to Action (Paris) to Impact (Delhi). The conversation has shifted from “should we regulate AI?” to “how do we deploy AI responsibly at scale in the real economy?” That is the transition from experimental AI to production AI. And production AI in regulated industries requires systems that enterprises can bet their compliance, their reputation, and their operations on.
The window is still open. But it is not indefinite.