Breaking down the Zepto IPO | AI’s memory race

ELI5/TLDR

Zepto filed to go public, and its filing finally shows the math: it loses about 79 rupees on every order it delivers, and it’s betting it can fix that by cramming more orders through each store rather than charging you more. The second story is about memory chips — the kind that go inside AI servers. A specialised, hard-to-make version called HBM has stopped being a cheap interchangeable commodity, the makers are minting record profits, and Nvidia’s CEO is literally flying to Korea to beg them to make more. The catch: the memory industry has a long history of building too many factories at the top of every boom and then watching prices crash. Everyone insists this time is different. It usually isn’t.

The Full Story

Zepto’s IPO: the economics of a 10-minute delivery

India’s quick commerce race has shrunk to three players: Blinkit, Swiggy’s Instamart, and Zepto. Zepto was the last private one, and its IPO filing is the first time outsiders get to see the books.

The pecking order, by orders in the last quarter of FY26: Blinkit at roughly 46% of all quick commerce orders (274 million orders, ~2,200 dark stores, 200+ cities), Zepto at ~35% (210 million orders, ~1,100 stores, 66 cities), Instamart at ~19% (112 million orders). The whole market they’re fighting over is worth about $11 billion — still under 2% of India’s grocery spend. The bet shared by everyone in the ecosystem is that the runway is long.

The headline numbers look explosive. Revenue went from ~4,500 crore (FY24) to ~11,000 crore (FY25) to ~22,600 crore (FY26). Losses went from 1,215 crore to 4,700 crore to ~6,000 crore.

But that revenue jump has an accounting asterisk. In January 2025, Zepto switched from a marketplace model to a principal model.

Before that when you ordered 500 rupees of groceries the company only counted its service fee in its revenue which is 50 rupees… Now after the shift it started counting the full 500 rupees because as per accounting rules when a company physically controls a good it sells it has to book the whole transaction.

So a chunk of the “growth” is the same business measured differently. This is also why Zepto reports a metric called NRV (net receivable value — order value plus ad and subscription income, taxes included) while Blinkit uses NOV and Swiggy uses GOV. They’re calculated differently. Don’t compare them naively.

Where the money actually goes per order

Follow a single delivery. You pay Zepto the full amount; Zepto pays the supplier and keeps about 72 rupees gross. Then:

~46 rupees to the delivery partner
~34 rupees to running the dark store (rent, pickers, packers)

That’s 80 rupees of cost against 72 rupees of margin. The act of delivering groceries loses money on its own. What plugs the gap is advertising — sponsored listings and banners. Ad revenue went 49 crore → 651 crore → 1,636 crore across FY24–26, and in FY26 Zepto earned more from advertising than it spent on its own marketing. Even so, it lost about 79 rupees per order in FY26 (down from 136 rupees in FY25).

The structural trap: small baskets, fixed delivery cost

But getting a 330 rupees basket to your door cost the delivery partner roughly the same as the 525 rupees one.

Delivery cost doesn’t shrink when the order shrinks. Zepto’s average order value is ~330 rupees, against Blinkit’s 525 and Instamart’s 700 — meaning Blinkit earns 59% more per order to cover roughly the same last-mile cost. Hence everyone’s minimum-order-value rule.

Zepto’s answer is not to push bigger baskets. It’s pure volume: jam more orders through each dark store so the fixed cost per delivery thins out. The plan is to go deeper in the 40 cities it already operates in, opening more stores in the same neighbourhoods. In Q4 FY26 it averaged ~2,140 orders per store per day, up from 1,325 in FY24. Instamart has roughly the same store count but does about half the orders, so Zepto is running its stores at nearly double the productivity.

The FDI handcuff

Here’s a wrinkle most people miss. Blinkit doesn’t just connect you to merchants — it owns its inventory, buying in bulk directly from brands. That buys cheaper prices, better shelf control, and (counter-intuitively) doesn’t tie up much cash, because in quick commerce stock turns over absurdly fast — a packet of biscuits arrives in the morning and is delivered by evening.

Zepto can’t do this. Indian FDI law bars a majority-foreign-owned company from holding inventory for direct retail, and Zepto is classified as foreign-owned and controlled. That’s why domestic investors like Motilal Oswal have been steadily buying larger stakes — the more domestic ownership Zepto builds, the closer it gets to Blinkit’s inventory-led model and better economics.

The cash picture is healthier than the loss

Zepto has ~5,600 crore in cash and liquid investments (March 2026), no debt, and has soaked up over 20,000 crore from investors since inception. The reported ~5,900 crore loss overstates the cash actually burned: dark-store fit-outs are paid once but expensed over years, and stock options are booked as expenses though no cash leaves the bank. Strip those out and the real cash burn was closer to 4,300 crore.

There’s also a working-capital tailwind. You pay Zepto immediately; Zepto pays suppliers over 30–60 days. At March 2026 it owed suppliers ~3,700 crore it had already collected, against ~2,400 crore owed to it. Because it owes more than it’s owed, it’s always sitting on more cash than it strictly needs — customers fund the float.

At its core, Quick Commerce is a logistic business dressed up in a retail interface.

The margins are thin, the operations intense, and every metric is load-bearing. A discount war, a soft city, or a tick up in delivery cost can flip the per-order math fast.

AI’s memory race: why Nvidia’s CEO is begging Korea for chips

Second story. Nvidia CEO Jensen Huang has been spending suspicious amounts of time in South Korea — Korean barbecue with LG, SK Hynix and Naver executives, and at SK Hynix’s Computex booth he signed a wafer of next-gen memory with three words: “Please make more.”

The big three memory makers — Samsung, SK Hynix, Micron — are printing money. SK Hynix posted a 72% operating margin in Q1 2026 (higher than Nvidia’s 65%). Samsung became the second Asian company after TSMC to cross a trillion dollars. Micron’s revenue more than tripled year-on-year. Korea even leapfrogged India to become the world’s sixth-largest stock market on the back of it. And yet all three can’t keep up with demand. There’s an acute shortage of AI memory chips.

What HBM is, and why it broke the commodity model

For most of history, memory was a commodity — a DRAM chip from Samsung was interchangeable with one from Micron, and buyers competed only on price. The only lever to compete was density: cramming more bits onto a flat 2D surface. Until the 2000s, RAM density doubled every 18 months. Then it hit a wall — over the last decade DRAM density doubled just once.

The industry’s escape was to go vertical. Instead of a bigger flat chip, stack chips on top of each other. That’s HBM — high bandwidth memory — a 3D skyscraper of memory layers, typically 8 to 16 tall.

Imagine you’ve run out of floor space in a city, so instead of building wider you build up. The catch is that each floor (each silicon die) has a slightly uneven surface, and you have to align dozens of them perfectly and run thousands of microscopic wires straight up through the whole stack.

Those vertical wires are called through-silicon vias. Doing this across eight dies is brutally hard and error-prone — one bug ruins the entire stack. Making 1 GB of HBM eats three to four times the wafer area of plain RAM, plus extra advanced packaging. SK Hynix built the first HBM device back in 2014, but it was overkill for the computing of that era. AI changed everything — models with billions of parameters need to shuttle enormous amounts of data within milliseconds, which ordinary RAM can’t do but HBM can. Supply still lags far behind demand because it’s one of the hardest manufacturing processes on earth.

And it’s getting harder, because designing HBM is now a co-design effort that starts with Nvidia. Nvidia no longer just makes GPUs — it designs whole AI supercomputers. A single GB200 NVL72 rack packs 72 B200 GPUs, 36 Grace CPUs, networking, and over 13 TB of HBM. Racks are built by contract manufacturers like Foxconn, chips by TSMC, memory by SK Hynix and Samsung — and everyone has to march in lockstep with Nvidia’s architecture, with relentless qualification testing. Samsung repeatedly failed Nvidia’s HBM3e qualification, only passing in late 2025. That tight co-design makes the moat deeper, not shallower.

The contracts changed too

The second structural shift is how memory is sold. For decades DRAM was haggled quarterly or even monthly, with a volatile spot market alongside. HBM, when it existed, ran on annual negotiations — nobody committed supply beyond a year because the industry was too volatile.

Since late 2025, that flipped. For HBM, all three makers have moved to 3-to-5-year long-term agreements with their biggest customers. Two reasons, interlinked:

Scar tissue. Past busts came from open-market volatility. Locked-in volumes at fixed prices give revenue visibility the industry never had, so capex can be planned rationally instead of panic-built.
The buyers changed. Memory’s biggest customers are no longer twitchy consumer brands like Lenovo or Xiaomi. They’re hyperscalers — Microsoft, Google, Meta, Amazon — making multi-year, multi-billion-dollar AI infrastructure bets. Long planning horizons, deep pockets, willing to sign long contracts because their data-centre buildouts are long too.

This is why the makers argue the super cycle will be bigger and longer.

But “this time is different” is exactly what they always say

A bigger cycle is not a cycle that disappears. The brief is blunt about the industry’s track record:

Anytime people show me these curves that just go to the sky with no end that never continues forever. This too will pass.

(That’s Willy Shih, a Harvard Business School professor who studies memory cycles.) The history is a graveyard of “this time is different.” In 2017–18 Micron’s operating margin hit 49%, the three-player oligopoly was supposed to enforce discipline — then revenue fell 30% and the stock dropped 56%. In 2021 chip makers again thought they had the cycle tamed; by 2023 Samsung had halved production. The mechanism is always the same: demand surges, everyone pours tens of billions into new fabs, the fabs take years to build, and by the time they’re online the market has moved.

And the long-term contracts don’t erase forecasting errors — they just shift the risk onto the hyperscalers. If AI monetisation disappoints and the hyperscalers pull back on data centres, the makers are left with idle capacity. Samsung plans 50% more memory capacity this year; Micron is spending $25 billion in capex this year alone; new fabs from all three hit volume by 2027–28, all built on the same demand signal.

China’s quiet flank attack

The wildcard is China’s push for memory self-sufficiency, via two firms:

CXMT (DRAM): Q1 2026 revenue up over 700% year-on-year, Q1 profit ten times all of 2025’s, now ~8% of the global RAM market and prepping the largest Chinese IPO since 2022.
YMTC (NAND flash): share up from 8% to 13% in a year, also heading for a listing.

CXMT is still two to three generations behind — it only recently reached HBM3 parity while the leaders are on HBM4. But China’s play isn’t the leading edge. As Samsung and SK Hynix redirect old RAM fabs toward HBM, they’re vacating commodity memory for PCs and phones — and commodity RAM is now more profitable than HBM, by Samsung’s own admission. CXMT is pouring into that gap.

The twist: YMTC has spent years perfecting hybrid bonding, expected to be the next frontier for taller HBM stacks, and holds a strong patent portfolio. Samsung has begun licensing YMTC’s hybrid-bonding patents for its own NAND. A sanctioned Chinese firm holding patents the Korean leaders need is a remarkable inversion.

The real risk isn’t near-term — it’s that commodity DRAM still accounts for ~80% of global memory volume, and a well-funded, government-backed Chinese player scaling into that segment could break pricing discipline at the low end. That’s the exact playbook by which Japan disrupted the US in the 1980s, and Korea then wiped out Japan. China may be running it again.

Tidbits

Bharat Forge is now supplying components for lithography machines and working with three of the world’s top five chip makers — riding the India Semiconductor Mission, which has drawn over 1.65 lakh crore in commitments.
VinFast (Vietnamese EV maker) is putting 7,000 crore into phase two of its Tamil Nadu plant, adding electric two-wheeler and bus lines, part of a broader $2 billion state commitment; eBus production targeted for August 2026.
Ahead of the third round of airport privatisation, the Civil Aviation Ministry wants to cap bids per operator to prevent monopolisation — one option caps a single entity at two blocks (~four airports), with the second-highest bidder getting a matching right if the same player tops a third.

Key Takeaways

Indian quick commerce is now a three-horse race: Blinkit ~46% of orders, Zepto ~35%, Instamart ~19%. The whole category is ~$11 billion, under 2% of India’s grocery spend.
Zepto’s revenue jump from FY24 to FY25 is partly an accounting illusion — the Jan 2025 switch from marketplace to principal model means it now books the full order value, not just its service fee.
Each delivery has ~72 rupees of gross margin against ~80 rupees of cost (46 delivery + 34 dark store), so the delivery itself loses money. Advertising is what plugs the gap.
Net loss per order: ~79 rupees in FY26, improving from 136 rupees in FY25.
Last-mile cost is roughly fixed regardless of basket size — the core structural problem of quick commerce. Zepto’s average basket (~330 rupees) is the smallest of the three, so it earns less per delivery to cover the same cost.
Zepto’s strategy is volume, not bigger baskets: more orders per store thins the fixed cost. It runs ~2,140 orders per store per day vs Instamart’s ~half on a similar store count.
Indian FDI rules bar majority-foreign-owned firms from holding inventory for direct retail, which blocks Zepto from Blinkit’s cheaper inventory-led model — hence the push for more domestic ownership.
Reported losses overstate cash burn: dark-store fit-outs are expensed over years and ESOPs are non-cash. Real FY26 cash burn was closer to 4,300 crore vs the 5,900 crore reported loss.
Negative working capital is a feature: customers pay upfront, suppliers get paid in 30–60 days, so Zepto floats on cash it has collected but not yet disbursed (~3,700 crore owed vs ~2,400 crore owed to it).
HBM (high bandwidth memory) is DRAM stacked vertically (8–16 layers) connected by through-silicon vias. It broke memory’s commodity model because the stacking is brutally hard — 1 GB of HBM uses 3–4x the wafer area of standard RAM.
HBM existed since 2014 (SK Hynix) but was useless until AI made it essential — billion-parameter models need to move huge data in milliseconds, which only HBM can do.
HBM is co-designed with Nvidia’s full-rack architecture, so memory makers must pass Nvidia qualification — Samsung repeatedly failed HBM3e, passing only in late 2025. The co-design deepens the moat.
Memory pricing shifted from quarterly/annual haggling to 3–5 year fixed contracts with hyperscalers since late 2025 — giving revenue visibility, but mainly transferring forecasting risk onto the buyers.
The boom-bust pattern persists: demand surges → everyone builds fabs on the same signal → fabs take years → market moves before they’re online. Micron 2017–18: 49% margin, then revenue -30% and stock -56%.
China’s CXMT (DRAM) and YMTC (NAND) are scaling into the commodity-memory gap the big three are vacating for HBM. Commodity DRAM is ~80% of global volume and currently more profitable than HBM.
Notable inversion: Samsung is licensing hybrid-bonding patents from sanctioned Chinese firm YMTC for its NAND products.
The historical disruption mechanism: government-backed capacity expansion breaks low-end pricing discipline — Japan did it to the US in the 1980s, Korea to Japan after, China may be next.

Claude’s Take

This is a genuinely good episode — two stories that are both more interesting than the surface, and Zerodha resists the temptation to editorialise into hype or doom. The Zepto segment is the cleanest public explainer I’ve seen of why quick commerce loses money structurally rather than incidentally: the fixed last-mile cost against a shrinking basket is the whole ballgame, and the FDI handcuff on inventory is a real, under-discussed disadvantage versus Blinkit. The point that reported loss overstates cash burn, and that negative working capital is a feature, is the kind of nuance that separates a useful explainer from a headline. I’d push back gently on one thing they glossed: the ad-revenue ceiling. They flag it as “an open question” but it’s arguably the entire bull case, and “more from advertising than it spent on its own marketing” is a slightly cute framing that compares two unrelated numbers.

The memory segment is the stronger half. The framing — HBM stopped being a commodity, contracts went long, but the cycle never really dies — is correct and well-argued, and the China flank (vacated commodity DRAM, the YMTC hybrid-bonding patent inversion) is the freshest insight here, the part you won’t get from generic “AI is booming” coverage. The Willy Shih quote is the right note to end on. The honest tension they hold — everyone has structural reasons this cycle is different, and everyone always does — is exactly the skepticism the topic deserves.

Docking points only for the transcript-level fuzziness (it’s a spoken brief, so “TRAM,” “drram,” “13 tab,” and “Neighbor” for Naver are transcription artefacts, not the show’s fault) and for the ad-ceiling soft pedal. Call it 8/10 — high signal, well-structured, intellectually honest, and it teaches you a mechanism rather than just reporting a number.

Breaking down the Zepto IPO | AI's memory race | The Daily Brief #485