Apr 11, 2025 · 52m · neon-show
Can Neysa be India's AI Cloud alternative to AWS, Azure & Google Cloud? | Sharad Sanghi
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of the Neon Show, host Siddharth Ahluwalia interviews veteran infrastructure entrepreneur Sharad Sanghi on founding Neysa, India's sovereign AI cloud platform designed to accelerate enterprise AI adoption and rival global hyperscalers. Drawing on his experience scaling Netmagic, Sanghi shares deep insights into GPU economics, AI data security, enterprise production workloads, and the rapid evolution of India's tech ecosystem.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
Siddhartha as informed peer 3.9 Guest teaching 4.1 Guest disagreement 0.7 Siddhartha pushing back 0.3
speaking balance: gold is Siddhartha, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 25:29 Pushback on enterprise CIO AI preparedness
The guest directly contradicts the host's premise that enterprises lack AI-competent leadership, explaining that top banks have long employed Chief AI Officers and massive ML engineering teams.
Hardest push from Siddhartha ▶ 22:03 Challenging core workload reliability amidst hallucinationsThe host pushes back on adopting AI for core production workloads by highlighting unresolved hallucinations, political biases, and factual inconsistencies in leading models.
Biggest teaching moment ▶ 25:29 Clarifying enterprise AI organizational maturityThe guest educates the host on how Indian enterprise banks have utilized machine learning for fraud detection in production for over a decade with dedicated internal AI divisions.
Siddhartha holds their own ▶ 33:08 Articulating proprietary data investment thesisThe host takes command of the discussion to present Neon Fund's investment thesis regarding defensible proprietary data moats in vertical enterprise applications.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Siddhartha as informed peer | Guest teaching | Guest disagreement | Siddhartha pushing back | Why |
|---|---|---|---|---|---|---|
| The Genesis of Neysa and Current Operational Scale | 3 | 3 | 0 | 0 | The host asks foundational questions regarding the inception and current operational scale of Neysa. The guest explains how enterprise demand for AI workloads during his NTT tenure led him to launch Neysa and deploy 1,200 GPUs. | |
| The Netmagic Journey and the Dawn of Indian Internet | 4 | 4 | 0 | 0 | The host asks about early internet infrastructure in India and notes the national security implications of sovereign AI compute. The guest shares historical context on Netmagic, angel backing from Exodus founders, and NTT's eventual buyout. | |
| Sovereign AI Infrastructure and Hyperscaler Differentiation | 4 | 4 | 1 | 1 | The host lists the few scaled Indian cloud providers and asks how Neysa differentiates against hyperscalers. The guest outlines key differentiators including private clusters, transparent billing, and dedicated MLOps support. | |
| Economics, Gross Margins, and Technical GPU Complexity | 4 | 5 | 0 | 0 | The host inquires about gross margins and compares the trajectory to US players like CoreWeave. The guest educates the host on the order-of-magnitude technical complexity involved in GPU clustering, microcode, and low-latency interconnects. | |
| Netmagic's Scale vs. Modern Indian AI Innovation and Talent | 4 | 4 | 2 | 1 | The host questions Indian AI competitiveness and attributes talent scarcity to a lack of domestic R&D. The guest gently reframes, noting exceptional domestic talent exists but depth of specialized experience needed to scale AI infrastructure is currently thin. | |
| Addressing AI Model Vulnerabilities, Hallucinations, and Aegis | 4 | 4 | 1 | 1 | The host asks how enterprises can rely on models given persistent hallucinations and biases. The guest outlines security vectors like data poisoning and explains how RAG, ethical frameworks, and their Aegis product address these enterprise risks. | |
| Enterprise AI Leadership and the Production Workload Shift | 3 | 6 | 3 | 1 | The host assumes traditional enterprise CIOs lack the depth to evaluate modern AI solutions. The guest directly corrects the host, pointing out that leading Indian banks have had dedicated Chief AI Officers and hundreds of data science engineers for nearly a decade. | |
| The Cloud Revolution (2006) vs. The AI Infrastructure Wave | 3 | 5 | 0 | 0 | The host asks for a comparison between the 2006 cloud wave and current AI infrastructure. The guest delivers an educational breakdown on power density spikes in data centers from 6 kW to 130 kW per rack. | |
| Defensibility in the Application Layer and Vertical AI Strategies | 6 | 2 | 0 | 0 | The host articulates Neon Fund's thesis on proprietary, non-public data moats in vertical AI. The guest agrees enthusiastically, corroborating the thesis with real-world examples like Inference and Data Science Wizards. | |
| Product Management and Go-To-Market Execution | 3 | 4 | 0 | 0 | The host asks about product management and go-to-market strategies. The guest explains how treating product managers as P&L owners and pairing direct sales with vertical ISV partnerships proved essential across both Netmagic and Neysa. | |
| Capital Allocation, Risk-Taking, and GPU Obsolescence | 3 | 5 | 0 | 0 | The host asks about key learnings on scaling infrastructure. The guest details the need to front-load capital expenditure before demand arrives while prudently balancing the acute risk of rapid GPU hardware obsolescence. | |
| DeepSeek's Disruption and the Democratization of Inference | 5 | 3 | 1 | 0 | The host discusses DeepSeek driving down inference costs and links the trend to Moore's Law. The guest agrees that crashing token costs democratize adoption and ultimately expand the market for inference-as-a-service. | |
| Dissecting AI Hype vs. Reality and NVIDIA's Long Bet | 4 | 4 | 1 | 0 | The host asks the guest to separate dot-com-style hype from genuine AI value and highlights Nvidia's evolution from gaming chips. The guest explains that while enterprise pilots can be unfocused, real-world mission-critical deployments are undeniably transforming operations. |
Statements from this episode (25)
Sanghi: Nearly all enterprises want or use AI in production but lack know-how
“Almost all enterprises today, whether it's just AI native startup or an enterprise, everybody either is already using AI in production or wants to use AI in production, but a lot of them don't know what, how to go about it, right?”
Sanghi: Selling standalone software in India is difficult
“The software sale in India is not so easy.”
Sanghi: Neysa has 65 employees and 1,200 deployed GPUs
“Scale in terms of number of employees, 65 in terms of GPUs. We've already deployed 1200 GPUs.”
Sanghi: Neysa has about 15 enterprise clients plus government POCs
“We have about 15 enterprise clients. We also have some research institutes. We also have some AI native startups, and we also have we are doing some POCs with some government institutions as well.”
Sanghi: Neysa raised $50M and spent over $42M on infrastructure
“So we've raised fifty million. We've pumped in more than forty two million in our cloud infrastructure.”
Sanghi: Hyperscalers are cookie-cutter and cannot customize solutions
“Hyperscalers have by the, Term hyperscale itself. They have huge scale. They are cookie cutter. They can't, they don't have, they can't customize solutions”
Sanghi: CoreWeave operates with close to 70% gross margins in the US
“So the leader in the space core weave in the U S has close to 70% gross margins.”
Sanghi: AI cloud gross margins in India will likely be 40% to 50%
“I think in India it'll be good because more price conscious and more competitive. I think the gross margin will be lower, probably be in the range of, Somewhere in the range of 40 to 50%.”
Sanghi: AI GPU infrastructure is an order of magnitude more complex than traditional cloud
“I can tell you that the level of complexity with GPUs and with AI infrastructure, given how latency sensitive and how you know, the number of parameters that you have to worry about is much, much an order of magnitude more complex.”
Sanghi: NTT Netmagic operates 350MW across 4.5M sq ft, exceeding $500M
“Now in entity, we have currently close to 20 data centers. Each of the data centers is approximately 300,000 square foot. Between 200 to 300,000 square foot. And you know, on an average about 30 megawatts per data center. So we have around three 50 megawatts i…”
Sanghi: India is behind on foundation models but will catch up soon
“I think when it comes to foundational models, maybe we are a little bit behind, but I think we'll catch up soon.”
Sanghi: Neysa selects only 1 in 10 candidates due to depth gaps
“At least when we interview candidates for every 10 candidates to interview, we're only able to select about one because, you know, a lot of the talent that we see doesn't have the depth that is required to scale.”
Sanghi: At least 50 Silicon Valley startups focus solely on AI security
“In the Valley. There are like at least 50 startups that are just focusing on AI security.”
Sanghi: Top Indian private banks employ 300 to 400 data and ML engineers
“So some of the leading banks of India, private banks have got three, 400 data science and machine learning engineers.”
Sanghi: 2025 will see a major shift toward enterprise AI production workloads
“People are right now still doing experiments and very little is in production, but I think this year will be a shift where we'll see more production AI workloads. I think last year there was you know, more of experiments and less production, but I think this y…”
Sanghi: AI data center rack power density will soon reach 130 kW
“I remember when we started net magic, our average density of rack was like six kilowatts per rack. Then it went to 10 kilowatts per hour, then went to 22 kilowatts per hour. Now it's at 40 kilowatts per hour and very soon it will go to 130 kilowatts per hour.”
Sanghi: Data center total revenue is expected to quadruple in a few years
“Data centers are already growing at very high rates, you know, anywhere from 25 to 30% as per industry standards. They're expecting the data center Total turnover or total revenue of the data centers to, you know, literally quadruple in a few years, right?”
Sanghi: Domain-specific vertical AI agents will outperform general AI applications
“I think you'll see more and more people focusing on a particular domain and building more and more agents for that domain, and that, that is something that I think will do well in the next few years.”
Sanghi: Over $2B invested in Netmagic post-NTT acquisition
“So in net magic, we've invested more than two billion dollars after entities acquisition a few billion dollars have been invested in there. Almost four, five hundred million a year is what we're investing in net magic now.”
Sanghi: Nvidia H100 lead times are 4-6 weeks, Blackwell takes 6 months
“You want to each 100 is four to six weeks. It takes time. It used to take six months. Now it takes four to six weeks. Now, of course the new black belt series takes six months”
Sanghi: AI workloads require immediate capacity, unlike traditional data centers
“This business, the way this works is when somebody needs an AI workload, they need it immediately. It's not that they're going to wait for you for six weeks for you to set up something and then they'll come to you. In data centers, at least customers give you …”
Sanghi: Hyperscalers give 18-month notice for captive data center builds
“So when a hyperscaler, for example, wants a captive data center build, they'll give you 18 months to build it, right? It's not or if they want a flow, they'll give you six months a headstart here. You don't get that time.”
Sanghi: OpenAI retains moats over DeepSeek in benchmarks and bias reduction
“In some benchmarks deep, deep seek is equal or better. And, but in a lot of benchmarks, open AI is still better because open AI also gives you this They've got the team of people that, you know, work towards reducing biases, et cetera. So they've done a lot of…”
Sanghi: AI capital investment is much larger than dot-com boom
“I wouldn't know this hundred X, but I think it is much larger, but you would say investments. If you look at in terms of pure investment is much larger. I mean, you look at it the way Nvidia stock has gone up. Right. So it's from pure investment perspective is…”