Stripe's Annual Letter: On the 5-Level Model of "AI + Payments"

Hello everyone, I'm Chen Tianyuzhou. Today we're going to analyze a letter—Stripe's 2025 Annual Letter. Why analyze it? Because this letter reveals many aspects of the future we haven't seen before, based on Stripe's $1.9 trillion transaction volume! If you're not in the industry, you might not be familiar with Stripe. Simply put, Stripe is a global payment aggregator, similar to Adyen, Pingpong, and Airwallex. This letter, less than 3,000 words long, is packed with information. It's not just the company's year-end summary, but also a deep analysis and forecast of the impact of payments on the entire globalized economic activity. In 2025, Stripe's total transaction volume reached $1.9 trillion. What does this mean? A year-on-year increase of 34%, accounting for approximately 1.6% of global GDP. In other words, for every 100 dollars of wealth produced on Earth, 1.6 dollars goes through Stripe's channels. This figure reached 1.4 trillion dollars in 2024, a staggering increase of 500 billion US dollars in just one year. After reading this letter, I have only one feeling: The old world we know is collapsing and being reconstructed at an unbelievable speed. In this article, I've tried to break down the key points of Stripe's annual letter in my own words, though some parts have been directly translated to maintain the original flavor. The amount of information is indeed vast, but every point is worth pondering. If you are starting a business or in the payments industry, I suggest you read this several times. Every trend here is impacting your future survival. The letter begins by introducing a very core concept—"sorting machine." They say that the market economy is essentially a giant algorithmic machine. Its role is not to make everyone rich, but rather to ruthlessly filter out profits, capital, and talent, and redistribute them to the most productive companies. Previously, this machine operated slowly, and everyone had enough to eat. But now, AI has added an overclocking engine to this machine. Looking at the data: The top third of the most profitable listed companies in the United States account for two-thirds of the entire stock market's market capitalization. This is the highest proportion since data became available in 1963. The top 10% of companies in the S&P 500 contributed 59% of the profits. And this isn't a valuation bubble; it's a genuine concentration of profits. This differentiation isn't just between large and small companies, but also a life-or-death struggle within the industry. The letter gave several examples: Retail Industry: Offline brick-and-mortar sales have only grown by 5% in the past three years, while e-commerce has grown by 30% during the same period. Airline Industry: The three major US airlines—American Airlines, Delta, and United—have seen their market share and profits increase over the past decade. However, by 2025, Delta and United will have captured almost all of the US airline industry's profits. Other airlines will essentially be lagging behind. Healthcare: Hospitals and insurance companies have seen their profit share shrink significantly since 2019, but the healthcare technology sector is projected to exceed $110 billion in EBITDA by 2029. Money is flowing from traditional players to new players. This is the terrifying aspect of "healthcare": Previously it was big fish eat small fish; now it's fast fish eat slow fish, and it's AI-powered fast fish eating slow, clumsy fish. More macro-level data shows that demand for software, computers, and data centers will drive nearly half of the US GDP growth by 2025. And this proportion will soon become the majority. What does this mean? Previously, we said software was eating the world; now, computing power is driving growth. Industries that haven't capitalized on computing power and software are being relentlessly left behind by this machine. AI-Driven Startup ExplosionWhat has Stripe observed?In 2025, the number of new companies joining Stripe reached a record high, with 57% coming from outside the United States.Moreover, these new companies are growing at an absurdly high rate,50% faster than the batch in 2024. Even more astonishingly, companies that achieved annual revenue of $10 million within three months doubled their initial investment. However, AI companies consider this normal practice. Why? Because code commits on GitHub increased by 41%. Previously, annual increases were only 10%-12%, but by 2025, this figure has skyrocketed to 41%. The number of iOS app releases increased by 60% year-over-year in December 2025. Previously, developing an app required a team of over ten people working on it for six months. Now, a programmer and a few AI agents can write some code over the weekend, and the product can be launched and generating revenue the following Monday. Stripe itself has also jumped on this bandwagon. They created something called "claimable sandboxes," allowing you to launch your Stripe account directly from AI programming tools like Manus, Replit, and Vercel. Finished your beta version and want to launch it? With a single click, a sandbox becomes a live account instantly, without any configuration changes. Currently, over 100,000 sandboxes have been created this way. And then there's Stripe Atlas—the world's simplest company registration tool—which saw a 41% increase in company registrations last year. Moreover, these new companies are monetizing much faster: in 2025, 20% of Atlas startups received their first payment from a customer within 30 days, compared to only 8% in 2020. Faced with these figures, we are forced to ask: Is 2025 an anomaly, or the beginning of a new normal? Stripe's assessment is: The acceleration in 2025 marks the beginning of a major inflection point in entrepreneurship and creativity driven by the large language model. This "machine" is accelerating; if you don't break into the top 10%, what awaits you may not be a slow decline, but rather an instantaneous collapse. Globalization is dead, global localization is new. One particularly interesting point in the letter is that "globalization has become the default state." Previously, Chinese business owners had a relatively fixed approach when venturing overseas: first establish a presence domestically, then send people to Southeast Asia or the US to test the waters, and finally consider expanding into Europe. Coca-Cola took 20 years to bottle its first bottle of soda in Cuba. McDonald's and Starbucks took 27 and 16 years respectively to open their first stores in Canada. After the advent of the internet, free services began to be released globally simultaneously, but there was still a delay in the availability of monetization tools. When Facebook opened to the public in 2006, anyone could register an account, but wanting to run ads? Not so easy. By 2009, the company had already been around for 5 years. Google launched its global search in 1998, but didn't receive its first pound advertising payment in the UK until 2002. This strategy failed. What are AI products like now? OpenAI's ChatGPT, Anthropic's Claude, Replit, Cursor, and Midjourney are all launched simultaneously globally, covering all markets in a second. Every AI product we've heard of is incredibly popular in every country we know of. Stripe's data reveals a very unusual phenomenon: For companies whose main revenue comes from overseas, 30% of their revenue comes neither from their home country nor from the top ten economies like the US, China, Japan, and Germany, but from transparent countries that are rarely seen in the news. What does this mean? It means that demand is dispersed, but the tools are unified. Previously, if you wanted to collect money from the Kenyan people, you had to open a branch office in Kenya and negotiate with local banks to connect to M-Pesa. Now? Stripe goes all in, integrating over 120 payment methods for you—all you need to do is click a button in the backend. For example, Gamma, a California-based AI platform with 70 million users for presentations, saw its Indian revenue surge by 22% that month when it integrated with Stripe and enabled UPI payments in India. This is the benefit of next-generation infrastructure: your product can be in San Francisco, but your money can instantly flow into the pocket of a breakfast stall owner in Mumbai. Ironically, the most geographically limited sector isn't traditional industries, but rather fintech companies themselves: Chime, after 12 years, is still stuck in the US. Want to open a Chime account? Forget about it if you're outside the US. Nubank, founded in 2013, served only Brazil for the first six years, and only added two more countries in the next six. Even Stripe itself, with its Issues product launched seven years ago, only covers 22 countries—faster than most, but still not fast enough. But this is also changing. The new generation of fintech players—Sling Money, DolarApp, Felix, and KAST—are building global financial applications from day one. Stripe also launched its first inherently global product, Financial Accounts, last year, which businesses can use to hold, send, and receive funds. On its first day, it covered more than 100 countries. Why can they do this? Because of stablecoins. Their borderless nature allows fintech to build a complete infrastructure. Born global, this is truly possible for the first time. The crypto winter, the stablecoin summer. 2025 may be a crypto winter, but it will definitely be the summer of stablecoins. For the past decade, stablecoin trading volume has fluctuated with cryptocurrency asset prices. But in 2025, this correlation broke. That year, Bitcoin prices plummeted, dropping 50% from their peak. Logically, shouldn't the crypto world be entering another "winter is coming" phase? But strangely, stablecoin payment volume has doubled, reaching $400 billion. And it's estimated that 60% of that is B2B payments. This is the key quote from the letter: "Crypto Winter, but Stablecoin Summer." Previously, people traded cryptocurrencies to get rich, but now they use stablecoins to save money and make quick money. Stripe's acquisition of Bridge, a stablecoin orchestration platform, saw its trading volume increase more than fourfold. The path for entrepreneurs has changed: A Y Combinator founder can use Stablecoin to raise funds. The money can then be deposited into a Stripe financial account to earn interest. When needed, the money can be directly used to pay the salaries of engineers worldwide—engineers who may be anywhere in the world. SaaS platforms can use smart contracts to collect stablecoin subscription fees, eliminating the need for wallet owners to manually sign every transaction. Companies using stablecoin for international expansion now have better payment tools, allowing them to directly embed digital wallets into their core products. With companies like Privy, Ramp, and Deel, businesses can manage both custodial and non-custodial wallets using a single API. Interoperability between stablecoin and fiat currencies is also rapidly improving. In April 2025, Bridge and Visa partnered to launch a card that allows businesses and consumers to spend their Stablecoins just like any other credit card. The card deducts money from the Stablecoin balance and automatically converts it to the local currency; merchants receive the payment indistinguishably from other payment methods, and may even be completely unaware that the user is paying with Stablecoins. Phantom, one of the most popular crypto wallets with 20 million monthly active users, is using Bridge to promote Stablecoin cards to its users. Even Visa has started issuing cards for Stablecoins. This means that the USDT in your wallet can already be used to buy a Coke at the convenience store on the corner. For payments, Stripe built a blockchain. Since it's preparing for the era of large-scale payments using stablecoins, Stripe has spent considerable time researching blockchain technology. Current blockchains are designed for transactions and DeFi; throughput, reliability, cost predictability, and privacy required for payments are not yet key considerations. Bitcoin can only process fewer than 10 transactions per second. In 2025, a surge in Meme coin trading on a major blockchain caused a Bridge user's payment to be delayed by more than 12 hours, and transaction fees for a single transaction skyrocketed by 35 times. These problems are critical, but the future will be even more critical. Because it is anticipated that most internet transactions in the future will be completed by intelligent agents, we need a blockchain capable of supporting more than one million, or even one billion transactions per second. In September 2025, Stripe and Paradigm jointly incubated Tempo, a blockchain specifically designed for payments. Tempo's logic is simple: I won't play those fancy DeFi games; I'll focus solely on payments. Dedicated payment channels: Not competing with Meme Coin. Sub-second final confirmation: As fast as swiping a credit card. Optional Privacy: Protects trade secrets while ensuring compliance. Interoperability with Compliance and Accounting Systems: Don't talk to me about money laundering; I know more about anti-money laundering than you do. These features may not seem significant, but they are crucial for the payment infrastructure that supports real economic activities. Visa, Nubank, and Shopify are already testing Tempo in various scenarios, including global payments, embedded finance, and remittances. Klarna, whose CEO once called himself a cryptocurrency skeptic, has now become the first bank to issue a stablecoin: KlarnaUSD on Tempo's testnet, using Bridge's Open Issuance to accelerate and reduce settlement costs for cross-border payments. Tempo's architecture is particularly well-suited for agent payments and microtransactions. The mainnet is launching soon; we look forward to seeing what new applications ambitious companies will develop with it. Small Business Financing DilemmaSince 2008, the per capita GDP growth rate of OECD countries has been a paltry 1.0%, while the average growth rate over the previous 46 years was 2.8%.Since 2008, the per capita GDP growth rate of OECD countries has been a meager 1.0%, while the average growth rate over the previous 46 years was 2.8%. Japan's aging population, Brexit, Europe's energy structure......However, there is ample evidence that a significant culprit is the sharp decline in global capital availability. In most OECD countries, bank capital requirements increased after the financial crisis, and capital availability for small businesses decreased. Basel III reforms raised the capital standards for large global banks. In Ireland, bank lending to small businesses declined by more than 66% between 2011 and 2019. In the UK, small business lending has remained weak since contraction in 2012. The US fared slightly better, with non-bank lending institutions providing some access, and per capita GDP growth of 1.7% over the past 15 years. However, the overall trend in the post-Dodd-Frank era remains the same: tightening banking regulations and declining capital availability for small businesses. Since 2010, loans over $1 million have increased by 68%, while loans under $1 million have decreased by 5%. Last year, the approval rate for small business loans in the United States was only 41%, compared to 50% in 2015. Stripe Capital aims to address the global shortage of capital, particularly for small businesses. For businesses using Stripe for payments, their real-time revenue data simplifies loan decisions. Businesses can repay loans with a small percentage of subsequent sales, essentially drawing working capital from their own growth. From 2024 to 2025, funding increased by 45%, and Stripe Capital served over 80,000 companies. Many of these companies received funding through the vertical SaaS platforms they used—GlossGenius for beauty salons, Tekmetric for auto repair shops, Pixieset for photographers, and so on. Over the past two years, Stripe conducted a randomized study to understand the impact of capital: how fast can companies go when we "lubricate their wheels"? The result was, much faster. Companies that secured capital funding grew 27 percentage points faster than their peers who didn't receive funding in the following year. The top 10% of companies grew three times faster than their peers. The next tier also grew nearly 100 percentage points faster. For example, Xirsys, a California-based server hosting company, used funding from Stripe Capital to expand its server presence in China, India, and Japan, subsequently doubling its revenue. It's worth noting that even companies with low credit scores experienced 11 to 18 percentage points faster growth after securing funding. Stripe, in its letter, proposed a hypothesis: as AI improves ROI, capital availability may become a more significant variable in future economic outcomes. "Low-Income Model" OR "High-Income Model" For established companies, the number of clearly identifiable growth levers is pitifully small, and the most enticing ones don't necessarily work. Take advertising, for example. In 2019, Gillette's ad "The Best Men Can Be" won a Silver Lion at Cannes. Later that year, Procter & Gamble recorded an $8 billion impairment charge on this business because the award-winning ad failed to drive explosive growth. Recently, an Indian Railways advertisement won the Grand Prix and six Golden Lions at Cannes 2025—a daily lottery with a chance to win 10,000 rupees with a train ticket every day—but it was discontinued after eight weeks because no growth in ticket sales was seen. This is why Stripe, after 15 years, maintains a missionary-like passion for payments. Compared to more uncertain growth paths, optimizing payment settings is almost a guaranteed way to increase revenue, and one of the growth activities with the highest ROI. Microsoft evaluates the performance of each payment service provider monthly. From June to October 2025, Stripe significantly improved Microsoft's authorization rate using a series of methods, including adaptive acceptance, card account updaters, and network tokens. As a result, Microsoft shifted more payment traffic to Stripe. Gatwick Airport's previous payment provider had issues with payment failures, disputes, and customer complaints. After switching to Stripe, the payment success rate increased by 2.5 percentage points. FICO conducted an A/B test, comparing the previous payment provider with Stripe, and then completely switched, resulting in a 1 percentage point increase in the authorization rate. Telehealth's Royce saw a 2% increase in authorization rates and a 3% decrease in dispute rates over the past 12 months using Stripe, resulting in tens of millions of dollars in additional revenue annually. Most businesses are currently living in what Stripe calls the "low-revenue model": their payment infrastructure isn't optimized, and they're wasting money on conversion rates, authorization rates, and fraud prevention. What does the "high-revenue model" look like? The checkout page should be dynamically adjusted based on each customer. It should be priced in local currency and offer only the most relevant of hundreds of local payment options. Introducing BLIK payment to Polish customers increased checkout conversion rates by an average of 46%; offering Pix to Brazilian customers increased conversion rates by an average of 31%. Behind these achievements lies Stripe's decade-long investment in AI, including their "payment foundation model" and a suite of AI-driven services—Stripe Radar, Optimized Checkout Suite, and Authorization Boost—quietly optimizing billions of dollars in transactions daily. But there's still plenty of room for improvement! Last year, Stripe began testing a new authentication method: allowing customers to tap their cards to their phones. This touch verification validates the card's NFC chip, effectively proving that the cardholder possesses the card. DoorDash participated in the testing and found that compared to previous anti-fraud methods, conversion rates were significantly improved, and chargeback rates were reduced. Stripe also built a new model for Radar to address new fraud methods arising from the AI wave—especially those that steal AI inference computing power through abuse of free trials, which are now very common. Here's an excerpt from the original text: CEOs: Flashy ads are fun, but don't overlook the growth opportunities right under your noses. Your heads of payments absolutely deserve more recognition. How about we hold an awards ceremony for them? In the south of France, to recognize payment optimization. Or if you think the weather there is too nice, San Francisco is also an option; April is perfect. The recent hot topic in the AI world is that models can not only think using training data, but also go online, use browsers, write code, and do things on the internet. For Stripe's world, the most relevant concept is: your AI will soon be buying things for you. Like other AI concepts, agent-based commerce has also been overhyped prematurely. Some people make grand promises, claiming that autonomous agents will know your every preference and plan and execute all your shopping. Stripe breaks this promise down into five levels: Level 1: Eliminating Web Forms You research and decide what to buy yourself. But nobody likes filling out web forms. If you could directly send the agent a link, have them fill in payment and shipping information, and then confirm it with you, that would be much simpler. The system doesn't make any decisions; it just types and clicks "buy" for you. Level 2: Descriptive Search You no longer need to search for products or specific attributes, but rather describe the scenario. "I need to prepare back-to-school supplies for a third-grade child in Chicago, including clothes (not too tight or itchy!), pencils, notebooks, and a lunchbox. My son likes K-pop Demon Hunters and tennis. School starts at the end of August." The system considers weather, material, size, durability, taste, reviews, and delivery time. Niche and long-tail products become easier to find. Annoying keyword searches will become a thing of the past. Level 3: Long-lasting Memory You no longer need to introduce yourself. "Find me Bobby's back-to-school outfit options." The system already knows your preferences and remembers the requirements deduced from previous conversations and purchases. You still decide what to buy, but the options have been filtered based on your taste and budget. Level 4: Delegation You no longer need to make choices. "Get back-to-school shopping under $400." The system handles the searching, evaluating, and purchasing for you. You trust it to weigh the trade-offs and choose what your son likes, just like you do. You only set the budget. This is essentially what most people think about when they talk about agency-based business. Level 5: Proactive Prediction There aren't even any prompts. The system already knows the school calendar, your son's preferences, and your usual budget. You only receive one notification: Here's the list of school supplies already purchased. This is the most futuristic version: what you need appears just before you need it, without you even having to say a word. This industry is currently on the edge of Level 1 and Level 2. Stripe revisited the mid-1990s, the period when the internet architecture we use today was finalized. Netscape created the graphical browser. HTTP and HTML became the shared application layer. URLs and DNS took off. Back then, nobody knew which protocol or player would win. Every Google had an AltaVista in front of it. Now, proxy commerce is at a similar critical juncture—it has the potential to be a generational game-changer. Like the early internet, the future of proxy commerce depends on universal interoperability. To this end, Stripe did a lot last year: It collaborated with OpenAI on the Agentic Commerce Protocol (ACP), establishing a shared technical language between AI platforms and enterprises. Its open design allows it to be used across payment providers and AI platforms. It also launched Shared Payment Tokens, a new payment primitive that allows agents to initiate payments without exposing credentials. Even businesses that don't use Stripe for payment processing can forward these tokens to their own vaults or other processors as security credentials. The Agentic Commerce Suite has been released, providing businesses with a single tool that allows them to sell across multiple AI interfaces and protocols (including ACP and Google's Universal Commerce Protocol released last month) with a single integration. Checkout, payment, and fraud prevention all function normally at the underlying level. Brands already integrated include Anthropologie, Urban Outfitters, Etsy, Coach, and Kate Spade. Machine payments have also been launched, allowing developers to directly receive payments from agents using stablecoin micropayments—via API calls, MCP usage, and HTTP requests. Autonomous agents themselves are becoming a new customer type for internet companies. In collaboration with OpenAI, we've created the first shopping experiences on ChatGPT. We're also working with Microsoft to bring similar capabilities to Copilot. No one can accurately predict where agent-based business will be by the end of 2026, but it's clearly past the hype phase and entered the construction and real-world testing stage. The pace of change will only accelerate. If all goes well, those small agents won't be trapped in walled gardens, but will be racing along the wide highway of open protocols. The Republic of Licensing and the Cost of Innovation Today's entrepreneurs and innovators possess tools and reach unimaginable to previous generations of industrialists. How wonderful it would be if the combined effects of human ingenuity ultimately translated into increased productivity and improved living standards. Last year, Joel Mokyr won the Nobel Prize in Economics. Mokyr is known for emphasizing the importance of culture relative to traditional economic factors such as capital, labor, and technology. Eighteenth-century industrialists relied on more than just coal or geographical location. They possessed a new culture—a "reform mentality"—that viewed the status quo as imperfect and subject to correction. In *The Political Economy of Technological Change*, Mokyr also observes that new technologies have often failed in the past, not because they were not advanced enough, but because technology decisions involve not only suppliers and customers, but also a host of non-market aggregators—regulators, committees, courts—who influence what is adopted. As AI and the internet expand the boundaries of possibility, the synthetic barriers that hinder adoption and adaptation will become increasingly expensive. The economic divergence we are experiencing suggests that growth depends on the application of useful knowledge, not the existence of abstract knowledge itself. AI promises to greatly improve drug discovery… but this potential will only be realized if it makes regulatory processes (including clinical trials) faster and cheaper. European entrepreneurs can use new tools to revive a sluggish economy… but only if they cut back on well-intentioned but ultimately harmful burdens, such as the EU AI Act. Next-generation nuclear energy technology could bring energy abundance… but only if outdated regulatory systems are thoroughly reformed. Autonomous logistics—from long-haul trucks to drones—could drastically reduce the cost of physical goods… provided local regulations don't harden into barriers. Mokyr has written about the importance of “civilian republics” in catalyzing the Industrial Revolution. Today, we live in a “licensing republic”: a filter composed of non-market aggregators. While much regulation is well-intentioned, it is now more important than ever to ensure that it carefully weighs the benefits gained against the potential for being blocked. Fortunately, we can support resilient companies that demonstrate the boundaries of possibility: Mistral AI and Bending Spoons are proving that top European talent can pierce the regulatory freeze. Zipline and Varda are gradually securing licenses for complex new hardware. Spring Health and Maven Clinic are weaving new software layers for modern healthcare. We remain convinced that ideas are crucial to driving economic progress, and that many of the best ideas are undervalued. Stripe Press's book list contains many people, stories, and models that we believe can contribute to the next wave of improvement. The team recently celebrated selling one million books. Works in Progress, our magazine about undervalued, world-changing ideas, has also recently gone into print. The Event Horizon of a Black Hole This letter devotes a large portion to the progress of AI, sometimes making it feel like we're falling behind. The gap between newly released products and last year's most advanced technology is stark. We're reminded of the phenomenon of falling into a supermassive black hole. If you unfortunately experience this, the moment you cross the event horizon doesn't actually feel anything special: the path is locally smooth, although the space of future possibilities has irreversibly changed after crossing that boundary. As we write this letter, we may be standing on the eve of a different, hopefully better, singularity. While many things around us feel similar to previous years in 2026, it's also clear that the next decade will be completely different from the past one. We remain convinced that a vibrant entrepreneurial spirit and a wise culture will help build a more successful future society. We hope Stripe can play a small role. If you are also driving economic growth—whether as an entrepreneur, business leader, or builder of financial infrastructure—we hope to see you at Stripe Sessions in April. As always, there's much to discuss. That machine won't stop; it will only spin faster and faster. Whether you become the selected winner or just redundant data thrown out depends on how you deal with it now. That's all for now, see you next time.