[!warning] Auto-translated Hinglish transcript. Company names and numbers came through garbled. Every ticker below is reconstructed and flagged — verify before treating any as fact. The guest’s name itself did not survive translation.

ELI5 / TLDR

A Pune-based chartered accountant turned full-time research investor walks through how he’s compounded at roughly 30% a year since 2015, then spends most of two hours teaching the data-centre boom as a worked example of his method. His core trick: don’t buy the obvious thing (the data centre itself), buy the picks and shovels around it — the power gear, the cooling, the cables. He maps the whole value chain, explains why India’s data demand is reportedly 16x its supply, and argues the real “AI story” is a power-and-cooling story, not a chip story.

The Full Story

The host is Kushal Lodha; the guest is the founder and managing partner of a SEBI-registered research firm (the auto-translation mangled both his name and the firm — it sounds like an “Alpha Wealth”-type outfit, and he repeatedly says he is SEBI-registered so nothing in the episode is investment advice). He’s a chartered accountant who spent his corporate years at Tata Steel, Saint-Gobain and then seven-plus years at Marico in commercial, controlling and FP&A roles. He opened his first demat account in 2003, got serious around 2009-10, and went full-time into research and advisory about four years ago. His claimed personal CAGR since 2015: “much about thirty percent.”

The pendulum: price versus value

His mental model is a pendulum diagram. One line — call it value — climbs steadily, because the universe of listed Indian companies grows earnings at 15-20% as nominal GDP compounds. The other line — price — swings wildly around it. The two almost never meet.

“Trains are not meant to be standing at stations, ships at ports.”

The point is that if price and value always overlapped, you’d only ever earn the GDP-and-earnings rate, 8-12%. Alpha lives in the gap. At the top of the swing — maximum greed, maximum confidence, “price to perfection” — the market has baked the next ten years of 25% growth into today’s multiple. Miss one quarter, give one soft guidance line, and the stock gets punished. That’s where your margin of safety is lowest even though it feels safest. He pins June-September 2024 as exactly such a euphoria phase, one that then corrected.

The villain in this story is what he calls storification: a company grows 30% for two years, and analysts quietly extend that into a 10-year DCF when a DCF outlook should really be three years. A useful tell for the top: promoters selling. He cites Whirlpool India, where the parent stopped supporting the Indian subsidiary’s valuation because it had floated to a multiple of the parent’s own.

“Bhav bhagwan” — price is god. He half-agrees. Price tells you the crowd is already in. Deep research tells you what it’s worth. The edge is when the two disagree.

Margin of safety and asymmetry

Borrowing the title of Seth Klarman’s book, he frames the opposite end of the pendulum — maximum pessimism — as a short, sharp drop that’s where multibaggers are made. The trick is the asymmetric opportunity: a sector in a genuine tailwind lifts even its 10th-best stock above the 2nd-best stock of a sector in the doldrums. The inflection to hunt for: a sector beaten down 60-70%, but where one company’s balance sheet shows it kept firing right through the downturn.

CCL Products is his teaching example for this and his longest hold. It’s a coffee company that white-labels and private-labels for global brands. He bought it around 2009 at a ~30-40 crore market cap; the conviction event was a 2011 Vietnam plant expansion that signalled disciplined, focused capital allocation rather than diversification “for the sake of it.” It went up roughly 70x; he still holds about 30% of his peak position. The lesson he keeps returning to: research is non-negotiable, and corporate-governance skeletons “don’t like to stay in the cupboard” — he flags a company spending ~3x the normal capex per tonne of output as exactly the kind of red flag that later surfaced as governance issues.

His other named winners: Apar Industries, bought after reading the FY22 annual report and concall when it was a ~3,500 crore company, a global leader in three of its four categories with margins inflecting — “deeply undervalued,” then up several-fold in two years. And Eicher Motors / Royal Enfield, which he calls his best-ever investment: improving capital allocation (quietly shutting tractors and industrial lines), a blue-ocean 350-500cc segment versus the red-ocean 100-150cc fight, and demand so far ahead of supply there were waiting periods. He claims a roughly 60-100x outcome and 21% topline CAGR sustained over 16 years. He’s candid that Eicher “looks like a dream — nothing about it is logical even today,” a nice admission that hindsight makes everything look inevitable.

Study the value chain, not the company

The shift that made him, he says, was moving from company-first to sector-first research. Stumble onto one company via a brokerage report and you’ll rarely land on the best company in the chain. Study the whole value chain and two things happen: you quickly see whether there’s any value there at all, and you learn where the margins actually sit. His textile example: of six-or-seven stages from cotton to cloth, the first four have no margin — so don’t put position-sized money there just because a company says it’ll move up the chain. Each sector needs its own rulebook (in financials, look at return on assets and watch for write-offs, not just growth; metals are pure cyclicals, which he learned the hard way owning Tata Steel in 2002-05).

The data centre, explained from scratch

The bulk of the episode is a data-centre primer used to demonstrate the method. A data centre, he says, is an “unsleeping digital library.” Nothing on your Google Maps query, your Netflix stream, your Zomato order, your online game is stored on your device — your phone and laptop are just interface screens. The “cloud” is aggressively physical: acres of buildings, 24/7, where even a voltage dip can’t be tolerated.

The build order: power must arrive first, then building and construction, then the insides — semiconductor chips, memory chips, CPUs and GPUs, optical fibre (which has replaced copper inside the racks), power cables, system integrators (the “architects”), and cooling.

He sorts the world into four data-centre types: hyperscalers (AWS, Google, Meta — 100,000+ chips, not investable for retail), colocation (shared shells you rent like hotel rooms, no capex of your own), captive (defence, government, banks with sensitive data), and edge (close to the user, where latency can’t be tolerated — self-driving cars, gaming, even military). His latency illustrations run from a BMW 7-series braking for a pedestrian, to Waymo in San Francisco, to a claim that Iran targeted US hyperscaler data centres in the UAE because “data is the new oil” and an Amazon/Meta/Microsoft data centre in your country is effectively a US military-adjacent asset.

“The AI story is actually a power story and a cooling story. Chips will get made — Nvidia does more today, someone else will tomorrow. It’s the story of power and cooling.”

The CPU-versus-GPU bit, fermented: a CPU is the one brilliant kid who solves problems fast but one at a time; a GPU is a hundred thousand kids calculating in parallel, where the total is more than the sum of the parts. AI data centres run on GPUs.

The numbers (treat as directional, not exact)

The translation scrambles units between gigawatts and zettabytes, so hold these loosely. Global data was ~2 ZB in 2010, ~13 ZB by 2013 (the first generative-AI models), ~33 ZB a few years later, crossing ~100 ZB around 2022, and projected to ~400-450 ZB by the mid-2020s — roughly another 4x. The headline he keeps hammering: India’s data demand is around 16x its available supply, versus the US where demand is only ~0.4x of supply (the US built ahead of time). India triggers: 5G rollout, enterprise cloud adoption, a mass shift to the colocation model. Reliance and Adani are reportedly putting $20-30 billion into data centres, helped by India’s capex running roughly half of land-locked Singapore’s.

Unit economics and why he won’t buy the data centre

A full AI-ready data centre runs ~150 crore per MW (~65% IT/servers ≈ 100 cr/MW, ~35% construction ≈ 50 cr/MW), up from ~110 cr/MW pre-AI. Colocation needs only ~50 cr/MW (partial work), earns 8-10 cr/MW, pays back over ~10 years at a 6-8% return — “like commercial real estate,” with leasing risk the only real risk and little capital appreciation. Cloud infra is higher-risk, higher-return: faster payback (4-5 years, some contracts 3) but exposed to quick obsolescence as chips and even cabling change; build-to-order with a 2-3 year client lock-in is how you ring-fence that.

His conclusion is the whole point of the episode: he won’t invest in the data centre itself — colocation is a 6-8% REIT-like return, cloud infra is a still-developing model with obsolescence risk, and hyperscalers aren’t investable for retail. Instead, he buys the ancillaries — power, cooling, cables, niches.

“In a gold rush, the people selling picks and shovels make far more than the ones doing the actual mining.”

The shopping list (names are reconstructed — verify everything)

Dynacons Systems — Indian system integrator/architect, building data centres for government and PSU banks (Bank of Maharashtra, Punjab & Sind Bank) since pre-Covid; DC-cloud segment revenue reportedly 14% → 37%, with an order pipeline of ~3,000 crore against a few-hundred-crore topline.
Black Box — the mirror image: ~93% international revenue, ~70% US, having worked for three of the top five hyperscalers; targeting 25% of revenue from India within three years.
E2E Networks (software/cloud) and Netweb Technologies (hardware/OEM) — E2E reportedly secured a direct Nvidia chip-supply contract in Sept-Oct 2025, holds ~3,700 cloud GPUs including H200s (~2,300) and H100s (~700), and leases servers into tier-3 data centres. L&T reportedly took a ~21% stake in the hardware name (~6,500 crore valuation on ~1,300 crore revenue).
Cooling — Aeroflex Industries (direct-to-chip), KRN Heat Exchanger (rear-door / chillers / dry air cooling), and Blue Star, Voltas and Castrol all developing immersion cooling. He introduces PUE (power usage effectiveness): old data centres ran ~1.3, today ~1.1-1.15, with immersion cooling pushing toward ~1.0 — lower is better.
Cables — Apar (data-centre cabling plus an Adani Navi Mumbai ~1 GW project) and Sterlite Technologies (optical fibre, 5G beneficiary, DC revenue single-digit and guided toward ~30%).
Past multibagger cited as a pattern (not a DC play): Titagarh Wagons, ~20x in 18-19 months when its order book crossed 10,000 crore.

Policy and risk

Two Feb-2026 budget moves he flags as the real “wow”: a 21-year tax holiday for a hyperscaler building a large data centre in India (even if services flow to users outside India), and a transfer-pricing safe harbour — charge a 15% markup and face no assessment or litigation (versus IT/ITeS slabs at 17.5% and above). Risks he names: grid and power constraints (especially tier-2/3, where electrification is still incomplete), technological obsolescence across wiring, chips, racks and cooling, hard thermal limits (a data centre can simply shut down from heat), and physical density/space constraints.

Key Takeaways

Personal CAGR claimed ~30% since 2015; CA background, ex-Tata Steel / Saint-Gobain / Marico; full-time research ~4 years.
Price oscillates around value; the two overlap only ~1-2% of the time. Alpha lives in the gap. “Price to perfection” at the top = lowest margin of safety.
“Storification” = analysts stretching a 3-year DCF horizon into 10 years. Promoter selling is a top-tell (Whirlpool example).
Asymmetric opportunity: the 10th stock in a tailwind sector beats the 2nd stock in a doldrums sector. Buy the company still firing through a 60-70% sector drawdown.
Margin of safety also buys you holding capacity — once 100 becomes 1,500, a 20% drawdown stops hurting.
Study the whole value chain, not one company. Money sits where the margins are; skip the no-margin stages.
Data centre = physical, not virtual. Build order: power → building → chips/memory/GPU → fibre/cables → cooling.
Four types: hyperscaler, colocation, captive, edge. Edge is latency-critical.
“The AI story is a power-and-cooling story.” GPUs (massively parallel) are what AI data centres run on.
India’s data demand reportedly ~16x supply; US demand only ~0.4x. India capex ~half of Singapore’s.
Full AI data centre ~150 cr/MW (65% IT, 35% construction); colocation ~50 cr/MW, 6-8% REIT-like returns.
His verdict: don’t buy the data centre — buy the ancillaries (power, cooling, cables). “Sell picks and shovels.”
PUE (power usage effectiveness): lower is better; immersion cooling targets ~1.0.
Feb-2026 budget: 21-year hyperscaler tax holiday + 15% transfer-pricing safe harbour.
Closing: define whether you want to make money or create wealth (different research); filter noise; don’t DIY serious wealth — apprentice under experts ~5 years first.

Claude’s Take

This is a genuinely good sector tutorial wearing the costume of a stock-tip podcast, and it’s worth separating the two. The value-chain mapping of the data-centre buildout — power, cooling, cables, the colocation-vs-cloud-vs-hyperscaler split, the unit economics per MW — is the kind of structured, first-principles framing most “100x stocks!” content never bothers with. The pendulum and margin-of-safety material is standard value-investing canon (Klarman, Buffett, the price-vs-value distinction), competently retold. If you ignore the ticker name-drops entirely, you come away with a real map of how this sector hangs together.

Now the BS filter. The thumbnail promise — “30% CAGR, hidden stocks, 100X winners” — is survivorship theatre. He names four or five multibaggers (CCL, Apar, Eicher) and admits, almost in passing, that he can’t recall the 40-50 places his thesis was wrong. That asymmetry is the entire problem with learning from a track record narrated by the person who lived it. To his credit, he flags his own hindsight bias (“Eicher looks like a dream, nothing about it is logical even today”), which is more honesty than most. But the format — a Groww-sponsored show, a SEBI-registered advisor reciting a watchlist with a “this is not advice” disclaimer — is structurally promotional, and several of those micro-cap ancillary names are exactly the small, illiquid, story-driven stocks that run hardest in a hype cycle and unwind hardest after.

Treat the specific numbers as directional only: this transcript was auto-translated from Hinglish and the units genuinely flip between gigawatts and zettabytes mid-sentence. “16x demand over supply,” “150 crore per MW,” the GPU counts, the segment-revenue percentages — all are plausible but none are quotable. The frameworks survive translation; the figures do not.

Score: 7. Dense, well-organised, and the value-chain method is transferable to any sector — that’s the keeper. Docked for survivorship framing, sponsor-shaped stock-listing, and specifics too garbled to trust. Not an 8: the teaching is better than the tips.

30% CAGR. The Hidden AI Power & Cooling Stocks To Find 100X Winners! | Kushal Lodha #24