Everyone knows Microsoft. You probably touched three or four of its products before lunch an Outlook inbox, a PowerPoint deck, or a Teams call,. That kind of company doesn’t usually go on sale as it is now.

Chart 1. MSFT's monthly closing price since October 2025. After peaking at $539 in late October, fell 21.6% in June alone — the steepest single month drop since December 2000. Chart Source: intrinsiqq.com

Microsoft just had its worst month as a public stock since 2000. The shares closed at $352.83 on June 25, down about 34% from their October high near $539. The main reason is Microsoft is spending a staggering amount of money building AI data centers, and that spending has crushed its reported free cash flow. The market looks at the cash that’s left, and calls the stock expensive.

That’s the consensus, and I think it’s wrong. By FCF measure the stock looks expensive. But free cash flow is the wrong lens for a company in the middle of the biggest build in its history. The right one is owner’s earnings and on that lens, the very same stock is cheap. The whole contribution of this post is that switch: change the denominator from “cash left over after the build” to “cash the business actually earns,” and “expensive” turns into “on sale.”

Here’s the path the rest of this piece takes:

1. Microsoft’s moat — switching costs, scale, and the two real networks

2. Is the AI capex actually turning into revenue? — following the $190B build for 2026.

3. Valuation — 39× free cash flow, or 21× owner’s earnings?

4. The DCF — what Microsoft is worth, and what you’d make buying today

Microsoft’s Moat: Switching Costs, Scale, and Two Real Networks

Microsoft holds three product revenue lines:

• Productivity & Business Processes. The work software: Microsoft 365 (Word, Excel, Teams, Outlook), LinkedIn, and Dynamics. About 450 million paid commercial seats, sold mostly as multi-year subscriptions. This is the most profitable part of the company.

• Intelligent Cloud. Azure, the server software businesses, and GitHub. This is the AI build everyone is watching, and it now does as much revenue in a quarter as the work software does.

• More Personal Computing. Windows, Xbox, and Bing. The legacy bucket and it’s actually shrinking.

  

Chart 2. Microsoft's $82.9B in Q3 FY26 revenue by segment. The work software and Intelligent Cloud are now roughly the same. Source: MSFT Q3 FY26 earnings.

The Microsoft Moat is different in three ways, and it’s worth outlining each one because they don’t all work the same.

The first is switching costs. Once a company runs payroll, email, security, and compliance through Microsoft 365, leaving isn’t a purchasing decision — it’s a year-long migration. That’s why about 97–98% of the commercial revenue is recurring, locked-in subscription money.

The second is scale. Microsoft runs more than 400 data centers, and the cost to build cloud infrastructure at that level is so large that, that most companies couldn’t pay. That cost of entry is itself a moat.

The third defense is network effects. Let me be exact about that, because The Equivara Portfolio owns network-effect businesses and I want to compare it like so.

A real network effect means the product gets more useful as more people use it, which pulls in still more people. For example, Klarna: every shopper it brings makes Klarna more worth offering at checkout, and every merchant that adds Klarna gives shoppers another reason to use it. Shoppers pull merchants; merchants pull shoppers.

Microsoft has network effects in two specific places:

1. LinkedIn is a two-sided network: 1.2 billion members make it the one place recruiters and advertisers have to show up, and that spending makes it the place professionals have to keep a profile.

2. GitHub is the developer version, roughly 150 million developers, the code they share, and now a Copilot trained on all of it; more developers make it more useful, which draws more developers.

Chart 3. Microsoft's Azure flywheel. 430M users generate proprietary data, which trains sharper Copilot and AI products, which make each seat more valuable, which wins more seats and the loop repeats.

But that is not where most of Microsoft’s money comes from. Office, Windows, the server business, Azure those aren’t networks. They’re held by switching costs and scale. So, the accurate version I’d argue is: MSFT has two genuine networks integrated onto a much larger business whose defense is switching costs and scale. That's a different kind of moat than a pure marketplace like Klarna, and a more durable one — Microsoft would still be a fortress if both networks vanished tomorrow.

The market is questioning AI eats seat-based software, Anthropic’s Excel and PowerPoint add-ins are already more loved than Microsoft’s own. It’s a valid worry because the early Copilot numbers are soft, a year ago Microsoft had only about 15 million paid Copilot seats against a base of 450 million, just 3% of its commercial user base.

But the picture is changing in the industry, for decades software was sold by the seat: a fixed price per person per month, no matter how much they used it. With AI, the industry is moving to usage-based pricing instead, where you pay for what the software actually does. Salesforce launched Agentforce at $2 per conversation, then switched to "Flex Credits" at about ten cents per action, that business is already at a $540M run rate, up 330% in a year. Even Figma, a classic per-seat design tool, has layered a usage based AI tier on top of its seats. The pattern is identical everywhere add an AI layer ontop of your existing product, charge for consumption, and revenue starts scaling with how hard people lean on it. Microsoft is on the same path. GitHub Copilot moved to usage based pricing on June 1, about 60% of Copilot customers are already on consumption credits.

Share

Is AI capex actually turning into more revenue?

The question isn't whether Microsoft is spending a fortune because it obviously is. The question is whether that money is turning into more revenue and at what pace.

Microsoft quarterly capital spending climbed from about $21B a year ago to a peak of $37.5B last quarter. For all of calendar 2026 the company guided to roughly $190B, up about 61%, and told investors about $25B of that increase is just higher prices for memory and components, not extra capacity. That guide is the single number that triggered the selloff.

Chart 4. Quarterly capital spending including finance leases. It roughly tripled in a year and stays elevated; the dashed bar is the Q4 guide of more than $40B as new capacity comes online. Source: MSFT FY25–FY26 earnings calls.

Now where is that money showing up? Azure, which grew about 40% last quarter, accelerating, not decreasing, and it keeps growing in the high-20s for six straight quarters. The broader AI business went from a $13B run rate at the end of 2024 to $37B this spring. And the contracted backlog (revenue customers have already committed to but Microsoft hasn't booked yet) sits at $627B, nearly double a year ago. To be honest, you don't sign up for $627B of future cloud commitments unless the build is producing something customers want to pay for.

Chart 5. The return on the build: Intelligent Cloud revenue has climbed every quarter while YoY growth held in the high-20s, and management guides it higher again next quarter ($38.0–38.3B). The two faint bars are estimates extending the guided trajectory into FY27. Sources: MSFT FY25–FY26 earnings.

However and this is a real risk worth breaking down about 45% of that $627B backlog is OpenAI. Backlog is a promise, not money in the bank, and OpenAI is loss-making company funding itself through enormous raises. Microsoft and OpenAI have already renegotiated their arrangement more than once. So a big slice of that committed backlog rests on OpenAI funding and their partnership loyalty.

If you want to see what single customer dependence looks like when it's the whole story, look at CoreWeave. Microsoft was about 62% of its 2024 revenue and roughly 67% in 2025, the top risk factor in its IPO filing was "if one customer leaves, the business breaks." A renegotiation, a cancellation, or that customer simply deciding to build its own capacity would take the forecast down with it. That’s bear sharpest point on Microsoft and I wont disagree with them.

The difference is OpenAI is 45% of the backlog line not 45% of Microsoft revenue. The actual company earns more than $280B a year across Office, Windows, security, and hundreds of thousands of Azure accounts that have nothing to do with OpenAI; strip OpenAI out and the rest of the backlog still grew 28%. So this isn't CoreWeave's existential bet, it's one concentrated line item on an otherwise diversified fortress. But it is real, it is why the market is right to critique the backlog number.

Here's where the spending physically goes. When Microsoft buys a data center, the cost does not reflect in the income statement all at once but gradually, it first shown on the balance sheet as property and equipment, then into profit slowly over years as depreciation. On the graph below: net property and equipment jumped from $205B last June to $283B by March.

Chart 6. Where the capex lands: net property & equipment grew $78B (+38%) in nine months. The rising depreciation on that base (accumulated depreciation went from $94B to $112B) is what flows through reported earnings. Source: MSFT 8-K balance sheets.

Subscribe now

Valuation: 39x FCF or 21x Owner Earnings?

Reported profit and reported cash flow both mislead you when a company is building hard. The fix comes from one Warren Buffet concept, he calls it owner's earnings: how much cash could the owner take out of this business in a year without weakening it? You start with the cash the business actually generates. Then you subtract only the spending it truly needs to keep its current earning power intact. Whatever's left is owner's earnings.

The part that matters for Microsoft is what you don’t subtract. Money spent to grow, like new data centers that aren't serving customers yet, isn't a cost of staying in business. And you should judge a choice by the return it earns, not bury it in the running costs and pretend the company got poorer. When you separate the spending that keeps Microsoft running from the spending that's building its future, the stock stops looking expensive and starts looking like a compounder on sale.

Chart 7. The owner’s earnings calculation. Start with ~$165B of operating cash flow, subtract only the ~$40B of maintenance capex it takes to keep today’s revenue running, and you’re left with ~$125B. Reported free cash flow $68B subtracts the $57B of growth capex . Source: Maintenance figure is an estimate; the rest is from filings (TTM through Mar 31, 2026).

That’s why owner’s earnings is the right lens here. Microsoft is in the middle of the biggest building record in its history. Reported free cash flow is roughly $68B over the last twelve months, which against a $2.62T market cap works out to about 39 times cash flow, genuinely expensive. But strip out the part that’s pure growth, and the same company is doing about $125B of owner’s earnings. That’s roughly 21 times. Microsoft trades near 21× the cash it really earns, while still growing revenue 18%.

One multiple is not clear for a company spending like this, so look at three earnings metrics:

• Trailing P/E — 20.9×. $125.2B of trailing net income is $16.85 a share; at $352.83 that’s 20.9×.

• Forward P/E — 18.2×. FY27 consensus earnings sit near $19.40 a share; the multiple falls to 18.2× on next year’s numbers.

• Price / owner’s earnings — 21.0×. ~$125B of owner’s earnings is ~$16.82 a share, essentially on top of net income — 21.0×, the cash-based confirmation of the earnings number.

So on every earnings-based measure, you’re buying one of the best businesses in the at nearly half off its normal price, while it still grows revenue 18%.

Chart 8. The same business under three earnings lenses against its own decade median of ~33×.

Share

What’s Microsoft Actually Worth?

Assumptions:

• FY26 revenue of about $329B.

• Revenue growth: a base case where the AI layer keeps the top line compounding around 20% next year, fading to 18% by FY31.

• Cash-flow margin: operating cash flow at about 50% of revenue, drifting up a point or two as scale builds.

• Capital spending: Im not leaning to a fast normalization. I hold capex near 30% of revenue through FY27 and FY28, because the build isn’t slowing while demand still outruns capacity, then let it ease to 27%, 24%, and 21% by FY31.

• Discount rate: a dollar arriving five years out is worth less than a dollar today, so I discount future cash back at 8.8% a year.

• Terminal value (4.0% growth): the business doesn’t stop in FY31. The terminal value captures every dollar of cash flow after that, assumed to grow ~4% a year in perpetuity,and discounts back to today.

• Sum the discounted cash flows, add ~$40B of net cash, divide by 7.43B shares.

One thing you should read carefully: the bear case of $317 doesn't mean “the stock falls to $317 five years from now”; it means “if the bear assumptions hold, the company is worth $317 right now, so at $352.83 you're paying about 10% too much.” The return column answers the question that matters to a buyer: pay today's $352.83, hold five years while the gap to fair value closes and the business keeps compounding, and what do you earn per year?

Chart 9. Fair value today under each scenario against the June 25 close ($352.83), with the five-year annual return beside each.

So what does this mean if you’re the one deciding whether to buy at $352.83 today? Two things.

First, look at your worst case. The bear assumes growth slows to the mid-teens and a price war squeezes Microsoft’s cloud margins. Even then, the business is worth $317 today — only about 10% under what you’d pay — and because it keeps growing the whole time you hold it, that bear still pays you roughly 6% a year over five years.

Second, look at what you’re actually betting on. You are not betting that Microsoft set up brakes on spending. The base case, worth $540 (about 53% above today’s price), already assumes capex stays pinned at 30% of revenue through FY28 and the slow normalization. So the only thing you’re really betting on is that the spending stops climbing while revenue keeps compounding.

Put those together and you have the setup every value investor is hunting for: a downside you can live with, and a base case that doesn’t need anything extraordinary to go right.

Subscribe now

The Bottom Line

Microsoft’s worst month since 2000 came from one number FCF being crushed by a data center build that is less a cost than a bet on the next decade. It’s the same misread the market is making with two lenders I’ve already taken a deep-dive here, Klarna and MercadoLibre: accounting forces a lender to book a loan’s full expected loss the day it’s written, so a fast-growing book front-loads the provision and back-loads the credit revenue. Earnings look worse the faster it grows, and the market punishes that upfront charge exactly as it punishes Microsoft’s upfront capex — provision first, revenue later.

Measure Microsoft on owner’s earnings instead of free cash flow and the 39× that looks expensive becomes about 21×, against a 33× decade norm, for a company still growing mid-teens. Even pinning capex at 30% through FY28, the base case is worth about $540, roughly 53% above today. One risk worth flagging is OpenAI is ~45% of the backlog. But this reminds me the setup the market set on Google when it called Chat-GPT the end of Search and Google doubled. Is Microsoft doubling next?

Disclosure: I do not currently own MSFT. Current holdings: NU, SOFI, CHIME, HOOD, KLAR, PLTR, QXO, CRWD, NET, and RBRK. This is my personal investment thesis, not investment advice; I am not a registered investment or financial adviser, and nothing here is a recommendation to buy or sell any security. Do your own research and size positions to your own risk tolerance. Figures come from Microsoft's SEC filings and earnings materials (trailing twelve months through March 31, 2026) and public market data anchored to the June 25, 2026 close of $352.83.

Figma is the one trade I’d take back and not because I was wrong about the company. I was wrong about the day. I bought into the IPO, a tiny position about 0.3% of the portfolio. Six weeks later I sold all and registered a 48% loss. What I got wrong wasn’t the company; it was paying up on the IPO. An IPO pop is the inverse of everything I write here: I look for the dip, the moment the market has looked away, ideally right before profitability. The IPO is the opposite, the entire re-rating happening at once, in public.

The rule I took from it, and now follow without exception: even a business that fits your thesis perfectly is a bad buy on its IPO wait for the hype to pass.

Chart 1 · Figma since the IPO: price and EV/sales. The stock fell from ~$108 to ~$19 (~85% off the first-day high) and the forward EV/sales multiple compressed from ~50x to ~6x.

The two questions I’m working through: does the original thesis still hold after three quarters of public results? And if it does, is the entry point a profitability inflection or a dip?

2. A $33B TAM and Widening as AI Expands Who Designs

Chart 2 · A $33B TAM, ~4% tapped. Figma's addressable market is ~$33B — the global software-design workforce, per IDC in the S-1; FY2026E revenue of ~$1.43B is about 4% of it, and IDC projects that workforce reaches 144M by 2029. Figma publishes a single current TAM, not a year-by-year forecast.

Figma sizes its addressable market at ~$33 billion today and it has cover only about 4% of it. The opportunity widens as Figma extends from design into developer workflows, marketing content, and AI-assisted prototyping, and as AI lets more people design: IDC projects that design workforce reaches 144 million by 2029. That widening is believable because the product offering has kept expanding — into FigJam for ideation, Dev Mode for engineering handoff, Figma Make for prompt-to-prototype, Figma Sites for publishing, and Figma Weave for AI media generation.

Figma's product launches, 2025 vs 2026. The platform widened from design into publishing, marketing, and AI prototyping in 2025; 2026 layered AI monetization on top. Source: Figma product releases.

The product offering has expanded faster than I expected and, crucially, AI has become the thing pulling new usage through it rather than the thing threatening it as many investors have believed during the SaaSpocalypse.

The clearest proof is the new monetization layer. In March 2026 Figma turned on AI-credit pricing — consumption charges layered on top of seats — and the CFO called it, in his words, “an exciting inflection point.” The early signal is strong: as of late April, over 75% of Org and Enterprise users who had blown past their included credit limits kept using credits, and over 95% stayed active. Pro teams that bought credit add-ons spend more than 3x what teams without them do, and adoption of the underlying AI is broad — around 60% of $100K+ customers used Figma Make weekly in Q1, and weekly active users of Figma’s MCP server (which lets AI coding agents read and write Figma files) grew 5x quarter over quarter.

This matters because the original bear case — the one that crushed the stock during the SaaSPocalypse — was that AI would compress seat-based software. Figma’s answer is to make AI a second meter running on top of seats, so more AI usage means more revenue, not less — and the first month of data says customers are accepting it.

Share

3. Figma isn't a design tool anymore but a collaborative product operating system

Figma's moat is that it has become a collaborative product operating system. Three things create it.

1. Network effects: real-time collaboration plus a plugin marketplace and community library that gets richer as more teams join.

2. Switching costs: once an organization builds its design system, component libraries, and brand tokens inside Figma, leaving isn't a software swap, it's retraining everyone whose work lives there.

3. Cross-sell flywheel: as Figma stacks products each new one lands inside the same accounts at almost no extra cost, deepening the lock-in.

The question is whether that moat is still widening and the single number that tells me the flywheel still turns is net dollar retention growth for three straight quarters: 129% → 131% → 136% → 139%.

Chart 3 · Net dollar retention (>$10K ARR). 129% → 131% → 136% → 139% — three straight quarter accelerations. Source: Figma Q1 2026 call.

Net dollar retention rising is the sign of a platform too embedded to leave. And customer expansion it's also growing strong: ~690,000 paid customers (+54% YoY), customers over $100K ARR up 48%, and ~70% using three or more products. The quarter's strength came from "seat expansion across entire organizations." That's the flywheel: land a team, expand across functions, and each product raises the cost of leaving.

Here's how that moat shows up in the financials. Figma doesn't break out revenue by product, seat type, or segment — it reports a single subscription line, split only by geography , so you can't track it’s specific products revenue directly. What you can track is what the flywheel produces in aggregate — 139% net retention, ~70% multi-product adoption, the $100K+ cohort up 48% — and the sum is revenue growth that has held around 40%+ for two years even as the base nearly doubled.

Chart 4 · Revenue and YoY growth. Source: Figma filings.

Revenue growth that stays in the mid-40s while the base climbs from $177M a quarter to $333M is the financial signature of a widening moat — each new seat and product lands on a base that isn't churning. That is the moat translating directly into durable, compounding growth.

4. The Competition is AI-native tools that generate interfaces from a prompt

What changed is AI-native tools that generate interfaces from a prompt — Cursor, v0, and most recently Anthropic’s Claude Design, which shipped in April 2026 and spread the “Figma is redundant” narrative. The sharpest bear case it’s that prompt-to-UI makes the canvas optional. That fear took the stock from ~$26 to ~$19 back in April 2026.

One thing pushes back hard: the largest software franchises are layering AI on top of their offerings and watching it compound revenue, not compress it. ServiceNow's AI suite, Now Assist, grew from ~$600M in annual contract value at the end of 2025 to ~$750M in Q1 2026, with management raising the 2026 target to $1.5B and the number of customers spending over $1M on it up more than 130% year over year. Salesforce’s Agentforce ARR passed $1.2 billion in its latest quarter, up more than 200% year over year, and Benioff calls agentic AI "a tailwind for our business."

The honest read is that the moat is holding against AI-native tools so far, but it's unproven over a five-to-ten-year horizon, and that's the thing I'd most want to keep watching.

Subscribe now

5. When is Figma profitable?

Now the question that matters most for a profitability inflection entry point: when is Figma profitable? Revenue is accelerating, ut grew 46% in Q1 2026 and billings are up ~44%. But, they reported a $142M GAAP net loss in Q1. But the entire loss is stock-based comp — $169M of it, mostly pre-IPO grants — and strip it out and Figma earned $52M of operating income (16% margin) and $89M of free cash flow (27% margin).

Figma isn’t trying to reach profitability it was GAAP-profitable before the IPO and is non-GAAP profitable and cash-generative now. The CFO expects stock-comp’s share of revenue to keep falling as those pre-IPO grants amortize. So the inflection is the GAAP crossover as the IPO stock-comp washes out over 2026–2027, exactly the “right before profitability” window I’m looking for.

Chart 5 · Q1 2026 operating profit, GAAP vs non-GAAP. The $137M GAAP operating loss is more than explained by $169M of (mostly pre-IPO) stock comp. Underneath: $52M non-GAAP operating income and $89M free cash flow. Source: Q1 2026 10-Q.

But here's the real risk to that timing: Management guided full-year FY2026 non-GAAP operating margin to ~9% (down from 12% in 2025), which means margins compress through the rest of the year, for two distinct reasons.

1. Gross margin: it has slipped to 82% (non-GAAP) from ~90%, and the CFO has been explicit that AI-inference spend is "the largest driver" of that compression as Figma rolls out Make and its other AI features — every AI query carries a compute cost Figma partly absorbs.

2. An investment ramp in sales, R&D, and event spend.

The risk for a profitability-inflection entry is direct: if inference costs stay elevated and the margin keeps compressing, the trough deepens and the GAAP crossover I’m waiting for can slip from 2027 into 2028.

Share

6. The Dip Arrived. The Inflection Didn’t Yet.

Which brings me to price — the thing I got wrong the first time. At its first-day close Figma was a ~$68B company on $749M of revenue: north of 50x sales. Today, at $18.65 (the June 18 close), it’s ~5.8x forward sales — and because Figma is cash-generative, I can value it two ways, not just one.

Chart 6 · Two valuation lenses. Left: EV/forward sales — Figma ~6x vs Adobe ~3.3x and Atlassian ~3.7x. Right: EV/FCF (trailing) — Figma ~36x vs Adobe ~8x and Atlassian ~19x. A premium on both, against ~2x the peers' growth. Sources: company filings, market data, mid-June 2026.

On a sales basis, the forward EV/sales is ~5.8x — about $8.3B of enterprise value on ~$1.43B of FY2026 revenue — down from north of 50x at the IPO. The peers I measure against: Adobe, the creative incumbent that tried to buy Figma; Atlassian and Monday.com, the closest collaborative work platforms. All derated hard in the 2026 sell off — Adobe ~3.3x (~10% growth), Atlassian ~3.7x (~20%). So Figma at ~5.8x is a premium, but it's also compounding ~40%+, roughly double the fastest peer.

On a cash basis, the shape repeats. Figma threw off ~$270M of free cash flow over the trailing twelve months, so it trades around ~31x trailing FCF — a premium to Adobe (~8x) and to faster-growing Atlassian (~19x), again because Figma grows faster and its FCF margin is still climbing.

Both lenses land in the same place: Figma is cheap versus its own history but a premium versus its peers, a “fair-to-attractive” entry, not a bargain. But, there's no clear path to inflection right now: the GAAP crossover would get pushed out as long as AI-inference costs compress the margin. The dip is there on price; the inflection is not yet there on profitability.

That splits the scenarios into three cases, each turning on a different mix of growth:

• BEAR ~$13−30% (<12 months): Growth slips toward the low-20s, and the multiple finishes compressing to its peers' ~3.5x sales. The de-rating completes what the SaaSpocalypse began.

• BASE ~$26+39% (12-24 months): Growth holds around 30%, the AI inference and margin trough plays out on schedule, and the multiple stays roughly where it is (~6x sales) on a bigger revenue base.

• BULL ~$45 +141% (into 2028): AI-credit revenue becomes a material revenue line, the GAAP-profitability inflection lands, growth keeps >40%, and the multiple re-rates.

CHART 7 · BEAR, BASE, BULL case

Subscribe now

Disclosure: I do not currently own FIG. I bought a small position at the IPO and sold all of it at a ~48% loss the trade this piece is about. Current holdings: NU, SOFI, CHIME, HOOD, KLAR, PLTR, QXO, CRWD, NET, and RBRK. This is my personal investment thesis, not investment advice; I am not a registered investment or financial adviser, and nothing here is a recommendation to buy or sell any security. Do your own research and size positions to your own risk tolerance. Figures are drawn from Figma’s S-1, its FY2025 10-K and Q1 2026 10-Q and earnings calls, and from ServiceNow and Salesforce earnings releases, price and valuation figures are as of the June 18, 2026 close and will move.

In the first quarter of 2026 MercadoLibre grew revenue 49% year over year to $8.85 billion — its strongest growth since Q2 2022, per the CFO on the earnings call — and both engines were healthy: commerce up 47% and fintech up 51% in dollars, gross margin still at 43.7%, and $611 million of operating income. Yet operating margin fell from a 13.5% peak in late 2024 to 6.9%, net income dropped 16% to $417 million, and the market sold the stock 40% off to about $1,600 from its $2,645 high, MELI is currently near a 52-week low. The selloff is a margin story, not a growth story.

Chart 1 · The jaws. USD revenue growth and operating margin across the last five reported quarters (MELI 8-Ks); margin peaked at 13.5% in Q4’24. Management ties roughly two-thirds of the Q1’26 margin drop to credit provisioning, the rest to the Brazil free-shipping subsidy and first-party inventory.

The compression is almost entirely deliberate: the Q1'26 earnings presentation's own margin waterfall traces most of the drop to front-loaded provisioning on a credit book that nearly doubled, the rest to Brazil free shipping and first-party inventory — and the CFO called margins "a consequence of our investment posture." That's reinvestment, not deterioration, and MercadoLibre has run this playbook before: in 2016 it launched free shipping in Brazil and ate the margin hit for years while building the Envios logistics network — spending that became the delivery lead rivals still can't match. Mercado Pago was the same, a drag at first and 43% of revenue today. Both times, the spent margin came back as moat.

Two of my investment philosophy beliefs shape how I read a moment like this, and they’re worth stating before the numbers.

The first: dips are entry points in high-growth stocks. When a business is genuinely excellent — 40%-plus growth, high-margins, a real moat — volatility isn’t a warning, it’s an entry point. CrowdStrike is my cleanest example: a multi-line security platform briefly priced as broken after the July 2024 outage. I bought that dip, and the stock trades near $650 today — roughly 3x the panic lows.

The second: buy great businesses and then hold on. Selling early to “lock in gains” is the most expensive habit I’ve had to unlearn. I trimmed 60% of my Palantir after it tripled in under six months — the P/E was past 200, and it felt prudent — and rotated into starter positions in Robinhood and Pinterest. Robinhood ran from the $35–40 range to over $100 by early August 2025. But the Palantir I sold would have been worth far more by year-end; the trim cost me more than the new positions made. The lesson wasn’t “never rebalance” — it was that PLTR had what justifies holding through a scary valuation: 40%-plus growth at that time. Compounding only works when you let it. Buy right, then hold on — and that second part is brutally hard.

So the question isn’t whether MercadoLibre is good — it plainly is. It’s whether this is one of those dips, in one of those businesses.

Share

2. Four Revenue Lines, One Customer: the Flywheel That Is the Moat

The shape I look for is many revenue lines stacked on one customer, each new product cross-selling into the same base at almost no extra cost, so every product added raises the cost of leaving. MELI’s four engines feed each other. A marketplace with about 84 million active buyers. A logistics arm handling 95.5% of the platform’s own deliveries. A retail-media advertising business growing 67–73%. And Mercado Pago: payments, a $14.6 billion credit book, roughly $19 billion in assets under management, around 78 million fintech monthly actives. The most important thing about the revenue composition is how the balance has shifted.

Chart 2 · The revenue mix, then vs now.

It’s worth seeing where the margin actually lives, because MercadoLibre is not an 80%-gross-margin software business — its blended gross margin was 43.7% in Q1’26 (10-Q), and the engines underneath are very different. MELI doesn’t disclose per-line margins, so for each line I take the closest public comp and reason from there:

• Advertising — the likely highest-margin line. Amazon doesn’t disclose its ad margin either, but analysts estimate it at 50–60%-plus, and Forrester benchmarks the exact format MELI sells at 70–90%. The logic: the shoppers and the purchase data already exist, so each ad sold costs almost nothing to serve. A ~70% margin for MELI’s ad line — which grew 73% in Q1’26 — is reasonable.

• Marketplace, payments and lending — high-margin, and here MELI publishes a number. Marketplace commissions cost little to serve once the platform exists. On lending, NIMAL — what MELI earns on its loans after loan losses and funding costs — ran 17.8% in Q1’26 (8-K).

• First-party retail — a deliberate drag, and that’s normal. Selling inventory yourself is a thin business even for the best operator on earth: analysts estimate Amazon’s retail operation excluding ads runs near break-even. So as MELI scales 1P fast — GMV up 69% FX-neutral in Q1’26 expect this line to make little or nothing for now. The point is assortment and price, not profit.

• Logistics — a cost, not a revenue line. Analysts estimate shipping absorbs about 15% of MELI’s revenue. That’s the price of the model, not a MELI problem: Amazon’s own 10-K shows shipping costs near 14% of revenue.

That split matters for the next five years. The high-margin engines throw off the cash that funds everything people are worried about — the logistics build-out, the credit book, the 1P inventory. A thin-margin retailer can’t self-fund a $10.9 billion logistics year and an 87%-larger loan book at once; MercadoLibre can. And as the mix shifts further toward advertising and fintech, the blended margin should at 43.7% should drift higher. All four engines are wired into one loop, and the loop is the moat.

Chart 3 · The flywheel. Each one makes the next cheaper to sell. That’s why I weight a multi-line business over a single-product one: the cross-sell is the network effect, and it’s legible in the numbers.

You can see the flywheel in a single user. Someone joins for a cheap purchase, subscribes to Meli+ for free shipping and streaming, pays with a Mercado Pago balance that earns yield, picks up the credit card — and the brands chasing that person’s attention pay for ads to reach them at checkout. One customer, four revenue lines, each making the next cheaper to sell and the whole relationship harder to leave.

3. The Margin Didn’t Break. A Loan Book Just Doubled.

The credit provisioning, is the most misunderstood number in the whole story. Roughly two-thirds of the Q1 margin drop is simply the cost of reserving against a loan book that grew 87% in a year, to $14.6 billion.

When MercadoLibre writes a loan, accounting books the expected loss immediately, at origination; the interest income arrives over the following months. Grow the book fast and you pile up provisions today against revenue that hasn’t shown up yet. Each cohort looks unprofitable the day it’s written and turns profitable as it seasons. The blended book looks strained only because the new, heavily-reserved cohorts outweigh the seasoned ones.

Chart 5 · The cost comes first.Provisions lead because they’re booked at origination; the revenue those loans earn arrives later — which is why fintech revenue re-accelerated to +54% in Q1’26 as the book scaled.

The other third of the compression is the Brazil free-shipping cut — the threshold dropped from 79 reais to 19 — and first-party inventory, both aimed at Shopee and Temu, who’ve taken a real slice of Brazilian volume. The attribution isn’t my guess; it’s the company’s own margin waterfall in the Q1’26 earnings presentation: of the ~600bps decline, bad debt added 3.9 points, cost of goods — the shipping subsidy plus 1P inventory — added 3.0, and sales & marketing 1.0, partly offset by leverage elsewhere. MercadoLibre ran this same play in 2016, and it compounded into the lead it holds today. Early signals say it’s working again: Brazil items sold accelerated from 26% to 56% growth, and unit shipping costs fell 17% year over year. Temporary compression only holds if the investment cycle is finite — if Shopee re-accelerates subsidies, or Argentina’s currency cracks, the recovery slips right. But absolute operating income still grew through all of it. This is a company spending its margin on purpose.

4. Three Valuation Lenses — EV/Sales, P/E, PEG — and All Three Say Underpriced

No single multiple is honest here, so I use three. Earnings metrics punish a company at a deliberate margin trough — the P/E looks inflated precisely because management is spending the margin — while sales metrics ignore profitability entirely. So: EV/sales for the through-cycle view, forward P/E for what you pay on the already-cut estimates, and PEG to tie that multiple to growth. At around $1,600 — roughly an $80 billion company — all three point the same way. On the cut earnings, the forward P/E is in the low-30s, against a trailing peak near 62x last October 2025. EV/sales is about 3x against a historical average near 10x. And the PEG is below 1 — for a business compounding revenue near 49%, the growth itself isn't being paid for. In plain terms: the market is pricing MercadoLibre like a slow retailer whose best days are behind it, while the business underneath is still growing revenue at nearly 50%. One of those two views has to be wrong.

Chart 6 · Three Valuation lenses. The PEG is the one I weight most here: with margins at a deliberate trough, EV/sales understates profitability and P/E overstates the price, but a PEG below 1 says the growth itself isn’t being paid for.

My hurdle is simple: if I can’t see a credible path to roughly 20% a year for five years, I pass. From $1,600, that means about $4,000 by 2031. The math that gets there doesn’t require heroics — three things have to hold:

• Revenue compounds 20–25%. It grew 39% in 2025 and 49% in dollars last quarter, and FX-neutral growth is already easing into the mid-40s — so 20–25% over five years assumes continued normal deceleration, not a re-acceleration.

• Margins mean-revert, nothing more. Operating margin climbs back toward its own 12–13% recent history as the credit cohorts season and the shipping cycle matures — a 2027-and-beyond story, and the thing that keeps the multiple honest.

• No re-rating needed. Hold today’s ~3x sales: 20–25% revenue growth puts 2031 revenue near $72–88 billion, which at the same multiple is roughly a $215–265 billion company — about 2.7–3.3x today’s $80 billion, comfortably above the 20% hurdle.

The nearer-term marks to grade it against: bear ~$1,500 over the next 12 months, if the investment cycle runs long and the discounters keep pressing — roughly today’s price, meaning the market is already paying for that outcome. Base ~$2,200–2,400 by late 2027, as margins normalize on schedule; the street’s 12-month consensus target sits near $2,220. Bull ~$2,800 into 2028, if the advertising and fintech mix pulls the blended margin back faster than expected.

Chart 7 · The scenarios.

5. The Verdict: a Dip Worth Buying, and Holding

Back to the two beliefs I opened with. This is a great business and it’s on sale, about 40% off its high, for reasons that are largely self-inflicted: a credit book reserving up front, a shipping subsidy, a 1P push. That’s exactly the setup my framework points at — volatility in an excellent business is an entry point, not an exit.

If you follow the two beliefs, the sizing writes itself: start small — a 1–2% position, enough to make you follow the company closely, small enough that being wrong is cheap — and let MELI earn a larger slot over time. Add only while three conditions keep holding: 6.9% proves to be the margin trough, delinquencies stay contained as the young credit cohorts season, and Brazil’s shipping investment keeps converting into unit-volume growth. Q2, due in early August, is the first read on all three. If they hold, size up — on schedule or on weakness. If any of them breaks, a 1–2% position is a cheap lesson, not a portfolio problem.

I won’t pretend this is as asymmetric as buying a great business right at the profitability inflection point — in my framework, that’s where the biggest upside lives. The second-best entry is a dip in an already-proven compounder, and that’s what MELI is: the downside here is closer to “fair price” than “screaming bargain,” and much of the upside leans on a margin recovery that’s more a 2027 story than a 2026 one. But a great business, still compounding, bought near a 52-week low at a third of its usual sales multiple, is exactly what my second rule was written for. You buy it, and you hold on.

Disclosure: I do not own MELI yet. This is my personal investment thesis, not investment advice — I am not a registered investment or financial adviser, and nothing here is a recommendation to buy or sell any security. Do your own research and size positions to your own risk tolerance. Figures are drawn from MercadoLibre’s Q1 2026 shareholder letter and results, the FY2025 10-K, the company’s earnings commentary, and the analyst sources referenced above; current-price figures are as of early June 2026 and will move.

NIO is crossing into profitability for the first time while the market still prices the burn. That’s the pattern I trust most — my biggest re-ratings came from buying great businesses around their first profitable print. Nubank near $7. SoFi near $9, averaged to $6.50. Each felt uncomfortable in the moment and looks obvious in hindsight.

Chart 1 · The re-rate comes after the first profit. Nubank's first profitable quarters landed late 2022–early 2023 with the stock near $4 (its 2022 low was $3.26); it has since re-rated to an $18.98 high. SoFi turned its first two GAAP-profitable quarters in Q4 2023 and Q1 2024 around $8 — it dipped to ~$6.50 first, then re-rated past $16 by Jan 2025 and to a $32.21 high by Nov 2025. The big move comes after the inflection, not before it.

NIO just printed its second straight quarter of adjusted operating profit, vehicle gross margin 18.8% — a four-year high — in a seasonally weak quarter. Revenue doubled. The stock is ~$6, with about half its market cap in cash. The question I’m working through does NIO fit the pattern well enough to start a position?

2. Volume doubled, margin jumped 860 basis points

Deliveries hit 83,465 in Q1 2026, up 98.3% and at the top of guidance — and the bear case three months ago was a Q1 miss after a weak January–February. Revenue doubled to ~$3.7B (+112%); vehicle sales (~89% of it) grew 129%. The growth is volume, tilted to pricier models (the ES8, ES9).

Then the margin: vehicle gross margin went 10.2% → 18.1% → 18.8% in four quarters — 860bps, the last step in a soft quarter. Overall gross margin was 19.0% versus 7.6% a year earlier; gross profit rose 428%. A big driver was lithium falling from its 2022–23 peak — but that input has reversed, up 125%+ over twelve months, with the hit landing mainly from Q2.

Chart 2 · The margin turn. Vehicle margin climbed 860bps in four quarters. But NIO isn't GAAP-profitable yet: Q1 2026 swung back to a net loss, with the profit on the adjusted operating line. At the inflection, not past it.

Underneath it, real cost discipline: R&D fell 41% year over year to about $270 million, from roughly 25–30% of revenue in early 2024 to 7–8% now. SG&A is heading toward management's 10%-of-sales target, helped by shared "Sky Stores" that put NIO, ONVO and Firefly under one roof instead of three.

3. The moat problem: one revenue line, not a flywheel

The businesses I own — Robinhood, SoFi, Nubank, CrowdStrike, etc.. — share a shape: many revenue lines stacked on one customer, each new product cross-selling into the same base at almost no extra cost, so the lines compound and every product added raises the cost of leaving.

Robinhood is the clean version. In Q1 2026 its revenue split across at least six material lines: net interest $359M (+24%), options $260M (+8%), prediction markets $147M (+320%), Gold subscriptions $50M (+32%), etc… — $1.07B total, up 15%. Crypto fell 47% and the other five still carried total revenue higher. Each new product — HOOD has gone from nine lines to twelve — deepens the wallet. SoFi runs the same motion across lending, banking and its tech platform; CrowdStrike across its security modules. Cross-sell builds the network effect, and the network effect is legible in the numbers.

Chart 3 · Two kinds of revenue moat. Robinhood’s six-plus cross-selling lines vs NIO’s ~89% one line. Its moat is real but a different kind which is what the rest of this section weighs.

NIO doesn’t look like that, and that’s exactly why its moat is the hardest part of this thesis for me. About 89% of its revenue is one line — vehicle sales — a transaction that’s lumpy with delivery volume, while the recurring pieces (BaaS, energy, swap, software) are folded into the ~11% “other” bucket it doesn’t break out. There’s no six-line flywheel to offset a soft quarter; if deliveries dip, the top line dips. The moat, if it’s there, is a different kind, and I have to test it on its own terms: a switching cost, a flywheel, and the scale gap underneath both.

Switching cost: real, but locked to one car

NIO runs 3,800+ swap stations. A swap takes minutes versus 40+ for a standard charge, and Battery-as-a-Service ($105–165/mo) turns the battery into a subscription model (~$6,300–9,900 over five years). An owner on swaps and a BaaS contract faces real friction leaving — but it’s a per-car switching cost, not a multi-product wallet.

I’m watching that recurring layer is high-margin and runs on infrastructure that’s already built. If the subscriber base and the energy and software lines compound, recurring revenue grows as a share of the mix while one-time vehicle sales shrink. It’s the quality improvement that would earn a higher multiple. It isn’t in the numbers yet.

The real flywheel: cost down, volume up, cost down

First, what it’s not: users pulling in users, or one product selling the next. It’s a cost-and-volume loop — and it’s worth being precise about what has actually moved the margin, because management was explicit on the Q1 call: most of the recent jump was product mix. The new ES8 carries a vehicle margin above 20% and was half of NIO-brand sales last quarter, pulling the blended margin up with it. The rest is cost-out, and the cleanest example is silicon: the in-house NX9031 chip cut about $1,450 per vehicle at launch, internalizing the fat margin Nvidia used to charge — and the chip unit, Shenji, raised about $325 million and is set up to sell chips to other automakers too, turning an internal cost-saver into a potential outside revenue line. Cheaper batteries (a cell-cost program, bill-of-materials cuts, denser packs, and battery reuse across swaps) do the rest.

Where does scale fit? That’s the part that makes a cost edge compound instead of running out — and management is blunt about it: a smart-EV maker needs roughly 2 million vehicles a year to hold a 20% gross margin, and NIO sold 326,000 in 2025, about one-sixth of that. So the two engines scale turns on — buying components at BYD’s or Tesla’s volumes, and spreading fixed costs (R&D, the factories, the swap network, the stores) across far more cars — are mostly still ahead of NIO, not behind it. The one place it’s already showing is the swap network: management said operating cost per station has improved significantly as utilization rises, and “other” sales hit a record gross margin above 20%. That’s the loop in miniature — more cars make each station cheaper to run and more useful, which makes the next car easier to sell. The catch: cost-out is largely a one-time lever, and management flagged a Q2 headwind of more than $1,400 per vehicle from pricier batteries and memory. The flywheel only really turns if volume keeps scaling.

Chart 4 · NIO's cost flywheel. Lower cost — today from product mix and cost-out like the in-house chip — lets NIO defend price in the war, which supports volume, and volume is what finally turns on scale economics. The loop is real but early: at one-sixth of the ~2M-unit scale management says a 20% margin needs.

The weak link: one-sixth of survival scale

At 326,028 vehicles in 2025 — a sixth of the ~2M-unit bar its own CEO calls the survival line — NIO lacks BYD’s or Tesla China’s purchasing power, and BYD just showed a battery that fully charges in nine minutes, narrowing the convenience gap swaps exploit. The moat is real but unfinished; the next two years are about whether the three-brand strategy — NIO premium, ONVO family, Firefly compact — gets it close enough to that bar.

Switching costs: real, but per car — a BaaS contract and swap dependence, not a multi-product wallet. NIO's swap network is a supply-side moat: many customers feeding utilization into one shared, sunk asset base. The compounding shows up in cost-per-swap and coverage density., not the demand-side moat I get with Robinhood or CrowdStrike. Durability: it holds only as long as volume keeps compounding and the recurring layer grows into a real second engine the way HOOD’s lines did. That shift is the one thing I’d track — and it isn’t there yet.

4. The Carvana playbook — and why NIO’s is harder

Carvana in 2022 was cash-burning, priced for bankruptcy. In 2023 it flipped to unit-economics discipline — ads −44%, SG&A −35%, ~19% of staff cut, inventory from $3.3B to $1.5B — lifting gross margin from ~10% toward 17%. More cars at a higher profit per car compounded gross profit and cash, swinging operating cash flow from −$2.6B (2021) to +$0.8B (2023). Buyers near the bottom caught a 100x.

Chart 5 · The same playbook. Cut cost, lift margin per car, until volume turns cash positive. Carvana had pricing power; NIO doesn’t.

NIO is pulling the same levers — R&D down 41% (now ~7–8% of revenue), SG&A toward 10%, the shared Sky Stores its ADESA. Profit per car went from ~10% to 19%, deliveries nearly doubled, gross profit rose 428%. Cash flow turned positive at the operating level in Q3 2025, and ~$7B in cash funds the rest of the climb without Carvana's bankruptcy clock.

Chart 6 · The cash-flow turn. Carvana's operating cash flow climbed from −$2.6B (2021) to +$0.8B (2023). NIO's is earlier and smaller: it dipped deeper in 2024 funding three brands and the swap build-out, and only turned positive at the operating level in Q3 2025. (NIO figures per its 20-F; the 2025 mark is the operating-FCF turn, not a full year.)

Where it breaks. Carvana lifted profit per car partly because it could — a fragmented used-car market with pricing latitude. NIO has none: it's fighting BYD, Tesla and Li Auto in the most brutal price war in autos, so it can't raise price. Every point of margin has to come from cost-out, which caps how fast profit per car can climb and makes the cash-flow turn smaller and more fragile. Carvana's 100x-plus was a two-year snapback off a near-death low; NIO's is a slower, decade-long compounding bet.

5. The bear, the base, and the 10x

NIO trades ~0.71x forward sales against a five-year average near 1.8x — ~60% below its own norm — with about half its market cap in cash. A compressed multiple for a business that just doubled revenue and turned the margin corner. The setup, not a guarantee.

Chart 7 · Scenarios.

• Bear (roughly −40%, ~$9B market cap). Volume growth lands closer to 23% than management’s 40–50%, input costs compress margins back toward the low-teens, the price war intensifies, and the multiple stays near 0.5x sales. Trigger: vehicle margin sliding back under the ~15% ONVO floor.

• Base (roughly +47%, ~$22B market cap). Volume grows toward the low end of the guide, vehicle margin holds in the 18–20% zone, NIO posts full-year non-GAAP profitability, and the multiple re-rates partway home to ~0.9–1.0x forward sales.

• Bull — my own long-term view. Over roughly ten years, if NIO compounds toward 2-million-unit scale, with swap and BaaS becoming a genuine recurring line and Shenji adding chip optionality, the re-rate is a multiple of today — something like 10x, or ~$60 a share from $6. The stock fetched ~$46 five years ago, so this isn’t a fresh-air number; it’s getting back past the old high, but on a profitable, scaled business. A long-duration, high-uncertainty bet.

What the 10x requires. A 10x is roughly a $150B market cap against ~$15B today, and every piece has to land. Deliveries scale from 326,000 toward the ~2M-unit bar (about 6x); vehicle margin holds near 20% with a 7–8% net margin through the price war; the recurring swap, energy and software layer grows as a share of revenue; and the multiple re-rates from 0.71x sales back toward and past its ~1.8x norm. Roughly: ~2M units at a ~$35,000 average price is ~$70B of revenue, ~$5B of net income at a 7–8% margin — which ~25–30x earnings, or ~2x sales, values near $150B. A decade-long stack of things going right, not one catalyst. And it’s a different animal from Carvana’s 100x-plus, a two-year snapback off a near-death low: when a stock falls ~99% and survives, the rebound math is enormous. NIO isn’t priced for bankruptcy — half its market cap is cash — so there’s no near-zero base to spring off. Its 10x has to be earned the slow way, by scaling the business.

Disclosure: I do not currently own NIO. This is my personal investment thesis, not investment advice — I am not a registered investment or financial adviser, and nothing here is a recommendation to buy or sell any security. Do your own research and size positions to your own risk tolerance.