硅谷Fintech观察③｜卖铲子的人：Stripe、Adyen、AirwallexSilicon Valley Fintech Watch ③ | Selling Shovels: Stripe, Adyen, Airwallex

"Why China Can't Produce a Stripe" series · Part Three

In brief — For every hundred dollars that flows across the internet, two or three are quietly skimmed off the top — and it isn't a bank doing the skimming. It's Stripe, Adyen, Airwallex: the people selling shovels. Last time I argued that most of the stars attacking the consumer side of banking have had a miserable run. This piece is about the winners: one that wins on scale and ecosystem lock-in, one that reaches godhood on margins, and one that spent a decade and eighty-five licenses laying global infrastructure across Asia-Pacific. And buried at the end is a fresh landmine: stablecoins, 99% of them pegged to the dollar.

Three years ago, the market handed Stripe a death sentence. The phrase it used was "death spiral."

In 2023, the company's internal valuation was beaten down from its 2021 peak of US$95 billion all the way to US$50 billion. The reporting called it a cash bonfire — growth stalling, "even US$50 billion looks expensive." Even its longtime bulls began to waver.

By February 2026, Stripe's latest employee share sale set the valuation at US$159 billion — up 74% from US$91.5 billion a year earlier. In 2025 it processed US$1.9 trillion in transaction volume, up 34% year on year, a figure equal to 1.6% of global GDP, and it was, explicitly, "solidly profitable." Even Microsoft routed "a substantial share" of its volume over to it, and the client roster added Nvidia and Amazon. It even wrote the largest check in its history — US$1.1 billion for the stablecoin company Bridge.

The bet that everyone dismissed as a "death spiral" paid off, three years on.

Which is the perfect footnote to the conclusion of the last piece: while most of the stars attacking the consumer side of banking are cut in half, halved again, or frozen out of the exit market, the people selling shovels are riding straight through the cycle.

This piece is about those shovel-sellers — why enterprise payment infrastructure is the crown of this whole race, what each of the three protagonists (Stripe, Adyen, Airwallex) stands for, and a track that barely existed in the source material three years ago but has since become a new battlefield: stablecoins. At the end I'll pose a question that carries the whole series to its finish: what makes the moat around this layer of shovel-selling infrastructure so deep? And why does almost none of it grow in China?

I'll use the freshest 2026 figures wherever I can. Let's begin.

I. Why Infrastructure Is the Crown: Shovel-Sellers Keep Their Margins

Let me plant the overall thesis first, then nail it down with the deals.

In every gold rush, the surest business was never digging for gold — it was selling shovels. For one reason only: infrastructure keeps the margin on its own layer, and shares it with no one.

This holds well beyond finance — AI right now is the same play. Through 2024 and 2025, the one quietly minting money wasn't any large model; it was Nvidia, selling shovels. Today the shovel has been handed one rung further upstream, to the memory-makers — Micron, SanDisk, SK Hynix, Samsung. In the bill of materials for a single AI chip, high-bandwidth memory (HBM) alone eats more than half, and Micron's 2026 capacity sold out long ago. The gold-diggers change with every wave; the shovel-sellers never stop counting cash.

The sharpest evidence is a valuation contrast — two fates on the same track. Both started in "corporate spend management" — issuing companies smart credit cards and automating expenses and cost controls along the way. The one that chose to become infrastructure, Ramp, is worth US$44 billion in 2026; the one that invented the category but chose to sell itself, Brex, was acquired by Capital One for US$5.15 billion. The former is 8.5 times the latter's exit price. The company that opened the category didn't earn the category's money — telling, isn't it.

Here's an analogy that makes it click: Stripe is to internet commerce what the water utility is to a city. Nobody gives the water company an innovation award, but every time the city turns on a tap, it's collecting. In its early days Stripe had an even wilder line for this — it wanted to become "a 3% tax on the future growth of the internet."

(In plain terms: embedded finance means slotting the functions that once belonged to banks — accounts, payments, lending — into any non-financial app, like Lego bricks. BaaS, banking-as-a-service, is the wholesale version: a licensed bank rents out its license and account capabilities for fintech companies to use.)

And AI is making this shovel-selling business more valuable still: it upgrades infrastructure from "helping people do things" to "doing things for them." One example is Lassie, which builds AI financial back-offices for doctors' practices and claims to automate 98% of bookkeeping — when infrastructure doesn't just save labor but does the whole job, what it can charge moves into an entirely different bracket.

The class of 2019 fintech stars has been completely reshuffled: Brex absorbed by Capital One, Ramp grown into US$44 billion of spend infrastructure, Mercury going out and getting a bank charter of its own — three companies that started on the same track, ending up in wholly different places. The ones living most comfortably are the few that turned themselves into infrastructure, or kept the spread for themselves. This shovel-selling logic is even being copied onto a brand-new commodity — compute: there are already companies building GPU price indices, aggregated procurement, and metered billing (dubbed "the Stripe of GPUs"), the Chicago Mercantile Exchange (CME) is co-building compute futures with them, and even BlackRock CEO Larry Fink has declared "there will be a new asset class: buying compute futures." Every time a new commodity is born, it needs an entire stack of shovel-selling financial infrastructure — and the people laying that infrastructure tend to outlive the people making the commodity.

I'll boil it all down to one line: shovel-sellers earn software money, not finance money — and that is the root of why they ride through cycles. The three protagonists below are three forms of that same logic.

II. Stripe: The Crown, and One Table for Reading All of Fintech

Stripe discloses almost no financial detail in public, but its playbook is itself the best textbook on selling shovels. It never leans on a single blockbuster product; it leans on a relentless series of moves to make itself the substrate underneath everyone:

• Using Tempo (the payments chain it co-built with Paradigm) to team up with SWIFT, Standard Chartered, and ANZ and rework the underlying standard for cross-border messaging;
• Acquiring Bridge (stablecoins), Privy (wallets), and Metronome (usage-based billing), fencing off the new fee pool of "how you get paid in the AI era" ahead of time;
• Its company-incorporation tool, Stripe Atlas, has seen new company registrations climb 41% year on year since the rise of AI — meaning it holds a leading indicator of the startup wave.

The best way to explain Stripe, and the whole race, is to borrow a table. Fintech companies can actually be sorted along "three business types × three revenue types":

The table's use is this: drop any company into a cell and you can immediately judge its growth potential and reasonable gross margin. The consumer-finance row earns cyclical finance money (covered in Part Two, and it's thin); Stripe's home turf — enterprise finance plus infrastructure — draws more and more of its revenue from the rightmost column: money that's repeatable, that scales, that has software economics. That's why the market is willing to hand it a software valuation.

Follow that bottom "infrastructure" row down and you'll find Stripe isn't alone — there's a whole cohort of dedicated shovel-sellers, each with a death grip on one link nobody can route around:

• Wise grips cross-border — FY2026 underlying income of £1.61 billion, a 26% pre-tax margin, serving roughly 18.9 million cross-border customers, holding that profit while squeezing its take down to 52 basis points; it dual-listed on Nasdaq in May 2026 (primary listing still in London) at a market cap of about US$12 billion. What it faces isn't one market but "a collection of niches."
• Plaid grips data — it's the "data interface" connecting your bank account to all manner of financial apps, valued at around US$8 billion in its latest 2026 round. (In plain terms: think of it as the plumbing socket for financial data; without it, many apps couldn't connect to your account at all.)
• Marqeta grips card issuance — roughly US$625 million in annual revenue, packaging "the ability to issue cards" into an API so any app can spin out its own card in a few lines of code; Knot, which does "one-click card switching," has pushed its monthly call volume to 100 million.

None of them face the consumer directly, yet every financial app a consumer touches has to pass through them. This is the ideal position for a shovel-seller: you don't have to win the users, you just have to become everyone's only way through.

(In plain terms: take rate is the cut — for every 100 dollars of transactions a merchant runs, how many cents the payments company keeps. Stripe's headline rate is around 2.9%, but most of that goes to the card-issuing bank and the card networks; the "net take" that's truly Stripe's own is only a thin sliver — which is why payments is a thin-margin, high-volume business that lives on scale and a software premium.)

The China comparison: many people reach for Alipay or WeChat Pay as the Stripe analog. That's a misreading. Alipay and WeChat are consumer-facing wallet gateways (the C-side); Stripe is developer- and enterprise-facing acquiring infrastructure (the B-side) — a few lines of code, and any website can collect money worldwide. One is the tap; the other is the pipe. That distinction is the core of Part Four.

III. Adyen: The Profitability Benchmark, Stripe's Mirror

If Stripe stands for "scale first," then the Netherlands' Adyen stands for "profit first" — it's the best profitability yardstick on this track.

Adyen's full-year 2025 net revenue was €2.36 billion, up 21% year on year (at constant currency), with an EBITDA margin as high as 53% (up from 50% in 2024), touching 55% in the second half alone, and the company targets a stable 55%-plus by 2028. These are software-company margins, not those of a financial company.

How does it make this much money? Because it runs a "single platform, all built in-house" model: one system spanning acquiring, risk, and settlement — no acquisitions, no bolt-ons, no reliance on outside processors. The cost is that growth isn't as sexy as Stripe's; the benefit is that every dollar of revenue carries an extremely high margin. Its take rate is only about 17 basis points — 17 dollars kept on every 10,000 dollars of transactions — riding purely on astronomical volume and ruthless cost control.

A metric often used to pit the two against each other is revenue-per-employee productivity (how much transaction volume each employee generates on average): a few years back, when Stripe was expanding at speed and its headcount blew past 8,000, its productivity dipped to roughly 40% of Adyen's — one of the market's arguments for writing it off at the time. But Stripe went on to prove, with scale and profit, that the ecosystem lock-in bought by burning through people is itself a kind of moat. So this isn't about who's right and who's wrong; these are two roads that both win — as long as what you're selling is software money.

Put Stripe and Adyen side by side and you have the two ways to live off selling shovels: one wins on scale and ecosystem lock-in (Stripe, US$159 billion valuation, solidly profitable), the other reaches godhood on margin (Adyen, 53% margin, roughly US$30 billion market cap). But both point to the same conclusion — payment infrastructure is a business that can hit software-grade margins, and it's winner-take-all.

Worth a footnote: Adyen too was once wrongly punished by the market, in 2023 (its shares crashed in a single day), for exactly the reason used to write off Stripe: "growth has slowed." And the result? Two years later its margin didn't fall — it rose. The shovel-seller stands up best to the cross-examination of the cycle.

IV. Airwallex: Asia-Pacific Can Grow Global Infrastructure Too — at the Cost of a Decade and Eighty-Five Licenses

The third protagonist is the closest to Chinese readers, and the best bridge into Part Four — Airwallex, a cross-border payment infrastructure company founded by ethnic Chinese entrepreneurs, now headquartered in Singapore and San Francisco (it began in Melbourne).

Its report card is solid: a Series H valuation of US$11 billion in June 2026 (up about 38% in six months from US$8 billion in its late-2025 round), annualized revenue of roughly US$1.3 billion, up 74% year on year, and — by its own disclosure — around 676,000 business customers served, while it throws itself fully at "agentic commerce": an end-to-end financial platform for AI agents to use. This is in fact a brand-new category, not yet locked down as tightly by license moats: when AI agents begin ordering, paying, and managing money on people's behalf, they need a layer of programmable financial rails; whoever first builds "the payment-and-custody interface that agents call" may become the Stripe of the agentic era.

(In plain terms / comparison: what Airwallex does resembles China's LianLian or PingPong — helping businesses collect, pay, and convert currency across borders — but it has spread the battlefield worldwide, building "cross-border money pipes for enterprises to use.")

But Airwallex's real moat hides in an unsexy number: as of mid-2026, it holds more than 85 financial licenses/permits, covering nearly 200 countries and regions. This is the two-sided nature of the business —

The license stack is both the barrier and the bill. It has raised US$1.8 billion in total and spent about ten years grinding these licenses out one at a time; and precisely because it's hard, latecomers must burn ten years of money too, walk the same regulatory drudgery, and only that constitutes a moat no one can copy. In a word: this business is stacked out of capital and time, not code.

Airwallex also carries two details that absolutely have to make it into this piece, and that lead straight into Part Four:

First, it is moving employees and data out of China — the company states plainly that U.S. customers' data lives only in the U.S. or Singapore, inaccessible to the China team. Second, it faces "China backdoor" scrutiny: Khosla Ventures partner Keith Rabois has publicly alleged that its China engineering team and Chinese investors (Tencent, Sequoia China) could leave it "potentially obligated to hand over customer data" under China's national security law; U.S. Senator Tom Cotton has already called for an investigation.

Put those two things together and you get the sharpest bit of foreshadowing in this whole series: a global infrastructure company with Chinese DNA can grow to US$11 billion, but its most successful parts are the ones that placed headquarters, data, and growth outside China — it even actively keeps its distance from the two characters "China." Is that a coincidence? Part Four answers.

V. Stablecoins: A Flawless Reuse of the Shovel-Selling Logic — Issuing the Coin Is Mining, Laying the Rails Is Selling Shovels

In the source material three years ago, stablecoins barely existed. But in 2026 they've become the fastest-growing new track in the shovel-selling story — and they reuse this piece's master framework to perfection.

Start with the seismic shift in the landscape: in 2025 the U.S. passed the GENIUS Act, establishing a federal regulatory framework for stablecoin issuers for the first time; the stablecoin leader Circle is already public.

But Circle's story is exactly the best cautionary tale for "issuing the coin beats selling shovels — until it doesn't." After its IPO the stock slid from a high of US$298 all the way back to US$60–70, its market cap falling to roughly US$16 billion. Why so fragile? Because about 96%–99% of its revenue comes from interest on reserves — you convert US$1 into one USDC, Circle takes that dollar to buy U.S. Treasuries and pockets the interest, sharing not a cent with you.

(In plain terms: a stablecoin is roughly a "USB stick for dollars" — copy dollars out of your bank account and carry them anywhere in the world, 24 hours a day; and the issuer is, at heart, "a money-market fund in crypto clothing" — its lifeblood is Treasury yield and the interest-rate cycle.)

So on this new stablecoin track, the real "shovel-seller vs. gold-digger" split is playing out again: the issuer Circle is mining — capital-heavy, exposed to the rate cycle, and forced to share half its yield with the ones laying the rails; while the true shovel-sellers are the ones who turn stablecoins into payment rails and hold the distribution channels.

The one who understands this logic best is, of all companies, Coinbase. It issues no coin of its own (USDC is Circle's), yet it holds a death grip on distribution: USDC held on Coinbase's platform earns it 100% of the reserve yield, and 50% held elsewhere — that one line alone brought it roughly US$1.35 billion in revenue in 2025. The issuer works itself to death on the rate cycle; the channel-holder skims a cut lying down — that's the "mining vs. selling shovels" split, replayed in the stablecoin world.

And the ones truly turning stablecoins into rails are Stripe, Visa, and Mastercard: Stripe paid US$1.1 billion for Bridge; Mastercard launched round-the-clock stablecoin settlement across mainstream chains and acquired the stablecoin infrastructure firm BVNK for up to US$1.8 billion; and, more explosive still, there are reports that Stripe, Visa, and Mastercard are joining forces to launch a stablecoin platform, with Coinbase possibly joining too. When three payment titans come together to lay this rail, it means the shovel-selling layer of stablecoins has become strategic ground that must be seized.

Go one layer deeper and even banks and sovereign states are now issuing coins themselves, but the battlefield has shifted from "issuance" to "distribution and interoperability": SoFi became the first nationally licensed U.S. bank to issue a stablecoin to all its users; Zelle, jointly owned by seven big banks, launched ZelleUSD — and the most motive-exposing detail is that it chose to launch in India first, and pointedly not as a stablecoin (because India's central bank simply doesn't recognize stablecoins); Tether went ahead and issued a national stablecoin for Georgia, running a sovereign-client model of "write your rules for you first, then issue on your mandate." Behind all this is a rule that keeps being validated: the industry has abandoned the fantasy of "every bank issuing its own coin" — the real battlefield was never issuance; it's distribution, interoperability, and the dirty work of plugging fiat in. Someone captured the relationship precisely: stablecoins are "the rail above the rails" — a Citi token, a JPMorgan token, a Zelle transfer, a SWIFT message, all must ultimately land on the same public layer.

And that reserve business, the "invisible spread," is itself a watershed: for the same dollar stablecoin, an issuer can earn up to 100% of the reserve yield under the U.S. GENIUS framework, only 40%–70% under Europe's MiCA, and about 70% in the UK — the difference in the rules directly decides whose business is fatter. That also explains why stablecoins are almost entirely a "dollar-only game."

Of course, the new track has its wrecks too. Tokenized equities failed their first large-scale stress test: when SpaceX went public, several exchanges abruptly canceled "tokenized pre-IPO" offerings and issued refunds, more than US$1 billion in orders went unfilled, and the entire market's underlying routing turned out to rest 100% on a single broker. The deeper obstacle is that U.S. state law (UCC Article 8) hasn't yet adapted to tokenization — "getting to real convergence will take at least another ten years." The new track is sexy, but the dirty work of laying it still starts at a decade.

This track also buries a big landmine for Part Four: about 99% of all stablecoins are dollar-denominated. In other words, the more successful stablecoins become, the more they are distributing dollars globally. A country that defends its monetary sovereignty through capital controls — should it, and can it, board this track? That's an unavoidable question in "why China can't produce a Stripe."

VI. Devil's Advocate: Could Infrastructure Get Its Table Flipped Too?

As usual, let me max out the bear case before returning to the judgment.

The bear case, said in earnest: shovel-sellers can have their table flipped too.

First, rate risk: in stablecoins, the newest infrastructure business, Circle has roughly 96%–99% of its revenue hanging on Treasury interest — enter a rate-cutting cycle and the whole model warps from the ground up; that's not a moat, it's a tailwind. Second, regulatory and technical risk: tokenized equities crashed their first big exam (SpaceX, over US$1 billion in orders unfilled, UCC still a decade out); crypto's cross-chain bridges are, by historical losses, the worst category in the entire industry, with hundreds of millions of dollars stolen at a time. Third, and sharpest — commoditization risk: even the payment titans' take rates are being pressed down (Adyen at 17 basis points, Wise voluntarily cut to 52), so might infrastructure ultimately be reduced to unprofitable plumbing?

To make the other side's case in plain terms: sell shovels long enough and the shovels themselves become dirt cheap — the real money may be in the gold mine, not the shovel factory.

But back to the judgment: the winners win precisely on the layer that can't be commoditized.

First, infrastructure keeps its margin, and that's been voted on with real feet: Adyen holding a 53% margin, Stripe turning three years of being written off into US$159 billion. Second, the moat isn't in "technology" at all — it's in distribution, licenses, and the "unsexy dirty work": Airwallex's 85 licenses and decade of grind are the thing nobody can copy. Third, the market has already stamped this with valuation — that set of figures from the last piece is worth copying out again: enterprise-grade infrastructure trades at an average multiple of 17x, while payments/transfers get only 7.7x, and in the largest B2B-payment fundraises, enterprise-grade infrastructure took 60%. When even the card networks themselves are putting up billions to lay stablecoin rails, that only proves this infrastructure layer is strategic high ground that must be seized.

VII. Closing: The Crown, the New Track, and an Unavoidable China Question

Three sentences to close Part Three.

First, enterprise payment infrastructure is the crown of fintech. Because it sells software money, not finance money — Stripe at US$159 billion, Adyen at a 53% margin, with the market using a 17x-versus-7.7x valuation gap to price "selling shovels" and "digging for gold" as two plainly different things.

Second, stablecoins are a new track growing out of that crown. But they still obey the same iron law: the issuer Circle is mining, exposed to the rate cycle; Stripe, Visa, and Mastercard, laying the rails, are the ones selling shovels.

Third, and most important — the shovel-selling business is stacked out of capital, time, and licenses. Airwallex proved with a decade and eighty-five licenses that "Asia-Pacific can grow global infrastructure too," but it also spoiled the price: you have to pick the right battlefield (cross-border), burn ten years of money, and even actively keep your distance from "China" while fielding "China backdoor" scrutiny.

So here's the question: why does this layer — neutral, enterprise-facing infrastructure that grows on corporate willingness to pay and on an M&A exit ecosystem — grow into a crown in America, yet almost refuse to grow in China?

That's the finish line of the whole series — next up: "Why China Can't Produce a Stripe."

(The data in this piece uses the freshest figures available in 2026: Stripe's US$159 billion valuation and US$1.9 trillion in transaction volume come from its 2025 annual letter and its February 2026 employee share sale; Adyen's net revenue and 53% EBITDA margin come from its full-year 2025 shareholder letter; Airwallex's valuation, revenue, and license count come from public media reports and company disclosures, with some operating figures on a company basis; Circle's revenue structure and share price come from its post-IPO public disclosures; public-company market caps are on a June 2026 basis and will fluctuate with the trading day; stablecoin and tokenization developments are synthesized from industry media and public sharing by fintech researchers.)