"There are only two kinds of companies in the world those who are conquering with AI and everybody else." - Srikanth Velamakanni [00:04:46]
"We are in that world of triple exponential scaling where we are scaling with pre-training... supervised finetuning... and inference time compute." - Srikanth Velamakanni [00:39:40]
Disclaimer: Orignal content owned by or sourced from third parties. It does not represent the views of 'Nuggets' platform or it's team. AI is used extensively across this platform including for summaries. Accuracy is not guaranteed, there can be mistakes. Any info or content on this platform is not a financial, legal, or investment advice. Do your own research. Refer for complete disclosures:- Terms of Use · Full Disclaimer
"The actual question we have to ask really is how many GPU are we consuming in India... we are already the second largest user of all AI products in the world." - Srikanth Velamakanni [00:47:31]
"If you build a business which feels like 'oh spud has got launched maybe my business gets automated and my product becomes a feature', then you're toast." - Srikanth Velamakanni [00:52:41]
"The most important indicator it'll seem simple but very profound is cost of capital versus cost of labor." - Srikanth Velamakanni [01:01:03]
"You know whatever we've estimated so far has been less so my gut feel is it [2 million GPUs] won't be enough." - Sharad Sanghi [00:47:14]
Speakers & Credentials
Akrit Vaish: Host, Founder of Activate. Previously built Haptik (a leading conversational AI platform). Operating as an investor and ecosystem enabler for Indian AI.
Srikanth Velamakanni: Co-founder and CEO of Fractal Analytics. He took Fractal public in 2026, marking India's first AI-focused IPO. A pioneer in machine learning applications with a 25-year trajectory in applied mathematics and enterprise AI.
Sharad Sanghi: Founder and CEO of Neysa (a Neo-Cloud AI data center provider). Previously founded NetMagic in 2000 (acquired by NTT). Backed by a $600M equity and $600M debt line from Blackstone, he is scaling India's AI infrastructure layer.
1. Executive Summary
The dialogue traces India's trajectory in the global artificial intelligence boom, contrasting the massive consumption of AI models against the country's limited localized infrastructure and foundational development.
A core thesis emerges regarding the "Cost of Capital vs. Cost of Labor" dynamic, illustrating why Indian enterprises adopt AI automation at a slower pace due to an 18x delta advantage favoring human labor.
The conversation details the physical evolution of AI, moving from simple logistic regressions and gaming laptops to $6M Blackwell B300 servers requiring immense liquid cooling power (up to 150 kW/rack).
The speakers agree that India's play is not in building horizontal models to compete with OpenAI or Anthropic, but in sovereign, vertically-integrated AI applications (e.g., healthcare, fintech, multilingual interfaces) deployed at population scale.
Ultimately, the panel emphasizes an ongoing era of "early success and peak fear," projecting that AI investments will mirror post-WWII infrastructure scale, driving immense growth and establishing AI as an inevitable cornerstone of all future business operations.
[00:59:32] The Capital vs. Labor Arbitrage Problem
[01:05:55] Building Profitable AI Businesses in India
3. Detailed Thematic Summary
The Evolution from Deep Analytics to True AI [00:09:28]
Fractal initially positioned its services purely as "mathematics" when incorporated as a public limited company in 2000 rather than analytics or AI [00:10:16]. At that time, computer science defined AI as rules-based expert systems, which Srikanth calls an "extraordinarily bad idea" that persisted into the mid-2000s [00:10:55].
Early on, deep learning models failed against simpler statistical models like logistic regression because neural networks were too shallow and acted as black boxes, providing zero incrementality on structured datasets under 1 million rows [00:12:02].
A paradigm shift occurred between 2010 and 2012 when unstructured data volumes exploded and deeper neural networks could utilize vast computational power. Machines began beating human performance in narrow tasks like credit decisioning by 2010/2011 [00:13:26], followed by image recognition outperforming humans in 2012 [00:13:07].
The Infrastructure Reality: Density and Capital Intensity [00:15:22]
Data center mechanics have radically intensified. In 2000, network racks required roughly 3 kW per rack and servers needed 6 kW per rack [00:15:28]. When hyperscalers (Amazon, Google, Microsoft, Oracle) arrived in India in 2015, loads pushed to 10 kW, eventually reaching 40 kW with fan wall cooling [00:15:54].
Today, liquid-cooled AI racks run up to 150 kW per rack, with the industry preparing for a megawatt per rack [00:25:53].
Traditional data centers represent appreciating assets due to real estate value—Sharad cites purchasing 54 acres in Airoli at 13 crores per acre, which appreciated to 54 crores per acre [00:23:12]. Conversely, AI Neo-Clouds center around rapidly depreciating graphic processing units. A single Blackwell B300 server currently costs 5.5 to 6 crores [00:24:10].
Because NVIDIA releases new hardware (Blackwell, Vera Rubin) annually, the asset risk is immense unless usage capacity is instantly pre-sold or tokenized [00:23:51].
Currently, two-thirds of all AI workloads worldwide are dedicated strictly to inferencing rather than training models [00:19:30].
Corporate Transitions: Fractal's AI Aggression and Neysa's Genesis [00:27:17]
Srikanth notes that Fractal allocates 10% to 12% of total revenue exclusively to AI R&D, a discipline maintained since 2012 [00:30:33]. Watching OpenAI's GPT-2 generate coherent story continuations in 2020 served as his true "mind-blowing" inflection point [00:28:54].
Fractal pivoted hard to build localized foundation assets, releasing kaleido.ai (text-to-image), a Marshall Goldsmith AI coach, Vya (an India healthcare AI model), and Cogentic (an enterprise agentic workflow platform) [00:32:09].
Sharad attempted to build a GPU rental infrastructure within NTT (the acquirer of NetMagic) in 2023. NTT declined the capital risk due to the rapid depreciation profile of GPUs [00:24:47]. Sharad subsequently founded Neysa, securing a $600M equity and $600M debt line from Blackstone to seize the supply-demand deficit [00:08:38].
The Global AI Capital Expansion vs. India's Hardware Deficit [00:38:10]
Srikanth estimates global investments in AI sit between $500 billion to $1 trillion annually, constituting 0.5% to 1% of the total world GDP [00:38:10]. In 2025, AI investment accounted for roughly half of the United States' total 2.5% to 3% GDP growth [00:38:29].
The Indian ecosystem is experiencing a severe mismatch in infrastructure capacity. Today, India operates with only 50,000 to 60,000 deployed 3rd-party GPUs [00:45:37], while the United States has deployed an estimated 10 million GPUs [00:45:53].
Sharad expects India's capacity will scale to 2 million GPUs by 2030, though he suspects this is a conservative underestimate [00:45:43].
The Indian data center IT load sits at 1.5 GW today but is projected to hit 7.5 to 8 GW by 2030. Within this expansion, AI workloads will transition from taking up 25-30% of total rack capacity to consuming 70-80% of it [00:45:27].
The Capital vs. Labor Arbitrage Dilemma in India [00:59:32]
The core inhibitor to indigenous AI innovation in India is not capability, but the structural cost of capital vs. the cost of labor [01:01:03].
Akrit and Srikanth model the math: In the US, an entry-level worker costs $60k-$70k, while the same tier in India costs $10k (a 6x labor cost difference). Concurrently, capital borrowing rates in the US hover around 3-5%, while risk capital borrowing in India operates at 15-20% (a 3x to 4x cost of capital difference) [01:02:18].
This creates a compounding 18x delta against automation [01:02:32]. An Indian enterprise considering a $1M GPU investment will repeatedly choose to deploy cheaper human labor because the ROI on hardware automation fails against local wage economics.
However, a reversal is approaching. Outbound AI telecalling operations in India are now achieving sub-human cost structures, presenting the first major breakthroughs where Indian AI fundamentally beats Indian labor rates [01:04:49].
The Reference Vault
4. Data & Figures
Data Point
Value
Context
Timestamp
Neysa Funding Line
$600M Equity / $600M Debt
Capital secured by Neysa from Blackstone to finance Indian AI Neo-Cloud infra.
The "Cost of Capital vs. Cost of Labor" Arbitrage Framework: [01:01:03] A macro-economic mental model explaining why tech adoption occurs at different speeds in different geos. When capital is cheap and labor is expensive (US), automation is highly incentivized. When capital is expensive and labor is abundant/cheap (India), enterprises default to throwing human bodies at a problem. This framework indicates that true Indian AI scale will only occur when the marginal cost of computing falls below local wage minimums.
Early Success and Peak Fear: [00:37:41] A framework used by Srikanth to diagnose the current market psychology. We have achieved highly visible, "early success" with LLM productivity tools, triggering "peak fear" among executives regarding disruption. This fear leads to defensive, massive capital expenditure (FOMO investing), effectively self-fulfilling the infrastructure build-out required to achieve actual AGI.
The Triple Exponential Scaling Law: [00:39:40] A thesis updating the previous assumptions of standard Moore's Law. In modern AI, scaling is occurring in three distinct dimensions simultaneously: pre-training compute, supervised fine-tuning compute, and inference-time compute. This framework guarantees aggressive, compounding advancements, nullifying prior arguments that LLM progress was plateauing.
Sovereign Vertical AI Strategy: [00:51:27] A strategic roadmap for emerging markets. Instead of attempting to raise $100B to build horizontal foundation models (like an OpenAI equivalent) where there is no historical precedent of success, India should build hyper-verticalized applications (e.g., healthcare). By combining global open-source models with proprietary sovereign data and population-scale distribution, local companies can achieve unassailable moats for less than $1 billion in capital.
The "Feature vs. Product" Survival Test: [00:52:41] A heuristic for founders building AI apps. If the release of a new foundational model (e.g., GPT-5 or Anthropic's Spud/Mythos) causes a founder to fear their product just became obsolete, they do not have a real business. A defensible business gets better and more profitable when the underlying foundational models improve.
6. Anecdotes
The "Gaming Laptops in Suitcases" Origin Story: [00:18:03] In 2015, Fractal established cure.ai to use convolutional neural networks for TB detection on X-rays. Realizing there were virtually zero enterprise GPUs available in India, an executive named Prashant flew to the US, purchased 3-4 heavy gaming laptops, stuffed them in his luggage, and brought them back. They literally strung gaming laptops together in a Mumbai office to act as the primary GPU computing cluster for their early healthcare imaging models.
The GPT-2 Mind-Blowing Moment: [00:28:54] Srikanth was tracking OpenAI and DeepMind closely for years. The true "singularity" moment for him wasn't ChatGPT, but the 2020 release of GPT-2. He fed the system a small paragraph of a story, and the machine continued to generate highly coherent, alternate creative narrative pathways. This single experiment convinced him to mandate Fractal's complete re-orientation around generative architectures long before the 2022 hype cycle.
The Hyperscaler Cloud Pushback: [00:24:47] In early 2023, as ChatGPT went viral, Sharad identified massive latent demand for dedicated GPUs from existing local cloud clients. He pitched NTT (the corporate owner of his previous company NetMagic) to aggressively pivot and buy GPU clusters. NTT leadership looked at the extreme asset depreciation rate of silicon chips and refused the risk, preferring the safety of traditional real estate data centers. This rejection forced Sharad to build Neysa with Blackstone's backing instead.
7. References & Recommendations
Companies & Startups
Fractal Analytics: [00:03:30] A global analytics and AI enterprise. Srikanth incorporated it as a public limited company in 2000 and achieved India's first AI IPO in 2026. Mentioned as the baseline for India's AI enterprise success.
Neysa: [00:08:38] Sharad Sanghi's Neo-Cloud startup, heavily backed by Blackstone to solve India's extreme GPU infrastructure deficit.
OpenAI / Anthropic: [00:41:51] The horizontal foundation labs dominating the market. Referenced constantly as the bar for multi-billion dollar capital deployment and model advancement (mention of Anthropic's Spud and Mythos launches) [00:44:08].
DeepMind / Isomorphic Labs: [00:42:21] Mentioned regarding their AlphaFold breakthroughs in protein folding, representing the applied future of AI beyond text generation.
Meta: [00:44:08] Mentioned briefly in the context of their Llama model launches happening concurrently with Anthropic's releases.
NetMagic / NTT: [00:06:10] Sharad's original data center company, representing the classical era of Indian IT infra.
Sarvam AI: [00:50:41] Mentioned as a leader in building multilingual, localized models for the Indian ecosystem [01:08:56].
CoreWeave: [00:08:31] Referenced (as Core) as an international comparable for Neysa; they started as Bitcoin miners and pivoted into the largest Neo-Cloud GPU providers globally [00:19:09].
Products & Technologies
Blackwell / Vera Rubin (NVIDIA): [00:24:02] Successive generations of NVIDIA architecture representing the brutal capital expenditure required to stay relevant in the infrastructure space.
Kaleido.ai / Cogentic / Vya: [00:32:09] Internal products built by Fractal representing the shift from analytics to text-to-image, enterprise agentic workflows, and sovereign healthcare models.
SAS, SPSS, Python, R: [00:11:42] Mentioned as the foundational data analytics tooling utilized by enterprises prior to the modern explosion of deep neural networks.
People
Bill Gates: [00:46:34] Sharad compares the potential underestimation of India requiring 2 million GPUs to Gates' infamous quote that "64KB ought to be enough for anybody."
Marshall Goldsmith: [00:32:09] Mentioned as the basis for Fractal's early generative AI coaching product ("Marshall as a coach").
Organizations & Geopolitical Institutions
NASSCOM: [00:31:05] Srikanth proposed the initial framework for a multi-billion dollar state-backed AI infrastructure fund through this tech association in early 2023.
India AI Mission: [00:31:05] The localized, government-backed infrastructure and foundation model initiative designed to supply compute power to Indian researchers and builders.
Historical Concepts
The 1940s/50s Infrastructure Boom: [00:38:50] Srikanth uses the massive capital deployed globally to build railroads and highway systems (2-2.5% of world GDP) as a historical analogue to justify the seemingly extreme costs of current AI datacenter investments.
DPI (Digital Public Infrastructure): [00:36:40] The Indian payment/identity stack (UPI/Aadhaar) cited as the blueprint for how AI might ultimately be deployed across a massive population scale to bypass legacy friction.
8. The Bottomline (by AI)
India is vastly over-indexing on AI consumption and critically under-indexing on hardware infrastructure, operating with a fraction of the compute capacity found in the West. The barrier isn’t a lack of engineering talent; it is an 18x mathematical delta driven by high local borrowing rates and deeply inexpensive human labor, which severely delays enterprise automation ROI. However, as compute costs inevitably collapse and labor costs creep up, a violent adoption curve will occur. Watch for a massive deployment of localized, verticalized AI applications (fintech, outbound calling, healthcare) specifically engineered to bypass horizontal foundation wars and achieve population-scale distribution.
Full Episode: The AI Industrial Revolution | 2 Jun 2026 | Naval and Nivi
Context: Host Naval Ravikant introduces a roundtable discussion on the "AI Industrial Revolution" with three frontier deep tech and software founders who build their own physical factories and tech infrastructure from first principles rath…
Real Estate Land Appreciation
13 Cr to 54 Cr per acre
Appreciation of a 54-acre data center plot in Airoli, showcasing traditional infra ROI.