Stripe data reveals that AI is reshaping the global economy.

Written by: @feigaobox

Translated by: AididiaoJP, Foresight News

In 1987, economist Robert Solow famously said, "You can see the computer age everywhere but in the productivity statistics." This statement perplexed economists for nearly a decade. It wasn't until the mid-1990s that the contributions of computers to productivity finally became evident in the data.

In 2026, the same confusion is repeating itself around AI. The bubble theory comes and goes, scholars argue endlessly, businesses hesitate, and macroeconomic data signals remain vague. But in one area, the impact of AI on the economy is no longer up for debate.

Now let’s look at Stripe.

In recent days, I attended the Stripe Sessions held in San Francisco. The volume of transactions handled by Stripe amounts to nearly 2% of the global GDP, with annual payments reaching $1.9 trillion, and over 5 million businesses on its platform. 86% of the companies on Forbes' AI 50 list are using Stripe. If the AI economy is a newborn baby, then Stripe is the heart rate monitor in the delivery room. It records the baby’s heartbeat earlier and more accurately than almost anyone else.

A study released by the St. Louis Fed in early 2026 shows that AI-related investments have contributed nearly 40% to the marginal GDP growth in the U.S., surpassing the peak contributions from the tech sector during the dot-com bubble. And when these investments turn into revenue, most of the settlements occur on Stripe. More importantly, Stripe is not just recording the heartbeat of the AI economy. At this year's conference, it announced plans to promote a new economic form: Agentic Commerce, where agents become the subjects of transactions. In a collective media interview, co-founder and president John Collison stated that he expects the role of agents as buyers in commercial transactions to become mainstream within 12 to 18 months.

Over two days, 288 products and features were announced, with more than 10,000 attendees, and a defining phrase echoed throughout: Agentic Commerce. Here are my observations from Stripe Sessions 2026, along with my personal thoughts.

How fast is the AI economy really running?

Before discussing agentic commerce, let's take a look at the overall picture of the AI economy. Solow said in 1987 that computers leave no trace in statistical data; nearly forty years later, AI is clearly visible in Stripe's data.

On the first morning of the conference, CEO Patrick Collison presented a set of data. Since the pandemic, the number of new businesses formed monthly on Stripe has remained high, but the curve has been relatively flat. Beginning in early 2026, this curve has shot up almost vertically. The direct reason is that AI coding tools have significantly lowered the barriers to entrepreneurship, allowing many developers to create chargeable products in just a few days using "vibe coding." Patrick described it as a more macro phenomenon—the entire economy is being re-platformed around AI. Maia Josebachvili, Stripe's Chief Revenue Officer for AI Business, added an external comparison: until 2024, the number of apps published on the iOS App Store was declining. After the introduction of AI coding tools, the release volume increased by 24% month-over-month.

Changes are not just about quantity; they are also about quality. Stripe Atlas is one of the easiest ways for founders to register a company in the U.S. It recently celebrated the establishment of its 100,000th company. At the conference, I heard a remarkable statistic: Companies registered through Atlas in 2025 generated double the revenue at the same life-cycle point as companies in 2024. Companies formed in 2026 have been generating five times the revenue of those formed in the same period last year.

In the afternoon's AI economy report, Maia Josebachvili listed several names driving the rise of the AI economy. Lovable achieved $100 million in revenue in eight months, and then surged to $400 million in the following eight months. Cursor reached a $1 billion annual revenue run rate in less than two years, and three months later doubled to $2 billion. Leading AI-native companies on Stripe saw a 120% growth in 2025 and have already grown by 575% in 2026.

Consumer spending is also sharply increasing. The top spending users are spending $371 a month on AI products, exceeding the combined monthly costs of internet access, streaming, and phone bills for the average American. I roughly calculated my own monthly token expenses, which have long exceeded my phone bill.

Patrick also made a comparison: the growth rate of businesses on Stripe is 17 times that of the global economy.

The next day, John Collison directly referenced the Solow paradox and explained it with a historical analogy. In 1882, Edison illuminated the first customers' light bulbs in Manhattan. But during the following three decades of electrification, productivity hardly improved. The reason was not that electricity was ineffective, but that factories were designed around steam engines at the time. Only after the entire factory was rebuilt did productivity gains materialize. John’s judgment is that AI is in a similar phase. Changes are already happening, but the old paradigms have not fully absorbed them. "However," he said, "I doubt AI will take thirty years."

Stripe's data seems to support his optimism. The AI economy is already booming on its platform. Almost every traditional company I encountered has top leadership actively pushing for AI deployment with a high sense of urgency.

Globalized from Day One

In addition to speed, another impressive characteristic of these AI companies is that they have been globalized from day one. Stripe has a saying: default is globalization.

Since I became an AI blogger, I often encounter an experience: AI content creation knows no time zones. AI news from across the Pacific carries the same weight as local news. The same goes for AI products. Large language models blur the interface language and interaction habits traditional software rely on. A unified chatbox allows global users to use products through natural language. In this sense, large language models have for the first time made a unified global software market possible.

Conference data corroborated this observation. In previous SaaS waves, the fastest-growing companies covered about 25 countries in their first year and reached 50 countries by the third year. AI companies are on a completely different pace: 42 countries in the first year and 120 countries by the third year. Maia mentioned that Kazakhstan has now appeared on the market lists of many AI companies. In the second day's "Indexing the Economy" sub-forum, Stripe provided a median figure: the top 100 AI startups sold to 55 countries in their first year.

Emergent Labs is a specific example. Founded in the U.S. in 2024, nearly 70% of its revenue now comes from overseas, with at least 16 countries each contributing at least 1% of revenue. Among leading AI companies, 48% of revenue comes from markets outside their home countries. This proportion was only 33% three years ago. Global revenue is no longer a supplement but the basic condition.

Speed and globalization are the two core characteristics of the AI economy, and both are directly related to Stripe. AI companies need to quickly establish payment capabilities, requiring them to be able to receive payments in 40 countries and regions within the first week. This has been what Stripe has been doing since day one.

It’s important to add some background on the founding of Stripe.

The founders of Stripe, Patrick Collison and his brother John Collison, are Irish, and they are cross-border entrepreneurs themselves. At the conference, I met an Irish colleague who told me that these two brothers are viewed as heroes among AI founders in Ireland. After coming to the U.S., they found online payments extremely difficult: connecting to payment systems required signing contracts with banks, undergoing PCI compliance checks, and interfacing with multiple intermediaries, a process that could take weeks or even months.

So in 2010, two twenty-somethings dropped out of school, moved to San Francisco, and wrote a solution that allowed developers to receive payments in just seven lines of code. These seven lines coincided with the rise of mobile internet and SaaS. Shopify needed to help millions of merchants receive payments, Uber needed frictionless payments from passengers, Salesforce needed to handle global subscriptions…they all chose Stripe. As these global customers grew, Stripe developed localized capabilities in 46 countries, covering 195 markets, and supporting 125 local payment methods.

For consumers, Stripe is not a company that stands in the spotlight. It hides behind the checkout pages of Shopify, the subscription confirmation emails from OpenAI, and the billing notifications from Uber. But this invisibility hasn’t stopped it from becoming the underlying financial pipeline of the internet economy. In the AI era, this global financial infrastructure gives Stripe a first-mover advantage when serving AI companies that are expanding internationally.

At this year’s conference, I also met Abhi Tiwari, Stripe's Global Product Head. He just took on this role three months ago and moved to Singapore. Stripe has engineering centers in San Francisco, Dublin, and Singapore, as well as a Latin America office in São Paulo. Abhi told me that when many AI companies approach Stripe, their first words often are, "We default to globalization; it doesn't matter where the users are." The old model of developing products at headquarters and then pushing them globally is being replaced by a model built locally by teams in the market.

Reaching global users is one thing; getting them to pay is another. The latter is much more complicated because each market has its own currency and payment habits. In this regard, Stripe primarily helps AI companies and other clients through two methods: local currency pricing and connecting local payment methods. The former allows Brazilian users to see prices in reais instead of dollars, boosting cross-border revenue by 18%; the latter enables Indian users to pay using UPI and Brazilian users to pay using Pix, increasing conversion rates by over 7%. When the AI demo tool Gamma integrated UPI in India, its Indian revenue skyrocketed by 22% that month. At the booth, I also saw the presence of the Chinese company MiniMax. From what I understand, many Chinese companies going overseas use Stripe’s financial services via offshore entities.

These AI-native companies also share a common characteristic: very few personnel, many being solo founders. One or two people, along with a group of agents, can support a globally revenue-generating company. The next day, Emily's presentation provided a statistic: the density of independent founders on Atlas has approached 5,000 per million Americans, with increasing numbers of them earning over $100,000 annually.

Emily used the term solopreneur. This reminds me of the rapidly developing OPC (One Person Company) trend in China. John used Ronald Coase's theory of the firm to explain this phenomenon. Firms exist because the internal coordination costs are lower than market coordination costs. But AI may be reversing this logic. When agents can help you discover services, integrate software, and handle payments, the external coordination costs plummet. You no longer need a whole room full of employees to do what previously required an entire department.

From Human Economy to Agent Economy

The AI economy described above, no matter how fast it grows or how globalized it becomes, still has humans as the transaction subjects. Humans are buying AI products, humans are starting businesses with AI tools. But the strongest signal I felt at this year's Sessions was that Stripe's next major focus is on a different transformation: the economic form where agents become market participants, which is Agentic Commerce.

This transformation has already quietly emerged in Stripe's own data. Will Gaybrick, President of Product and Business, showcased a set of figures. For years, the Stripe CLI (command-line interface) was only used by a small group of highly technical users, and its usage hardly changed. After 2026 began, usage suddenly surged. The reason is that agents do not require elaborate graphical interfaces; a simple CLI is often more useful. Maia's data shows that in 2025, traffic from agents reading Stripe documentation grew by about tenfold. If this trend continues, by the end of the year, the number of agents reading Stripe documentation will surpass that of humans. Stripe’s API documentation, honed over more than a decade, has found its new round of most faithful readers.

If agents spending money sounds foreign, consider two scenarios that have already occurred.

The first is that shopping interfaces may be shifting toward model chat windows. Consumers are now commonly using ChatGPT, Gemini, or Instagram to research products. The distance between research and transaction has been compressed into a single interface. Similar cases have emerged in China, such as buying milk tea within AI applications.

In a group media interview, John Collison explained why this compression is difficult to reverse using his experience purchasing a travel power adapter. If an agent completes the entire process from research to ordering, and days later the product arrives at home, it will not return to another website to fill in personal information from the beginning, even if the product on that website might be slightly better. Once a shopping agent completes the search process, the next natural step is checkout.

The second example is even more interesting: OpenClaw. Anyone who has followed the "lobster" wave knows it is one of the hottest open-source autonomous agent frameworks currently. Users give commands to agents via messaging applications like Feishu, Telegram, and WhatsApp, and the agents autonomously execute tasks. The key is that OpenClaw can consume hundreds or even thousands of dollars in token costs in a day. It manages token consumption and usage itself. Although many cases still require human authorization, in the end, it is the agents consuming the tokens, and tokens can be directly translated into money.

The leap from agents managing token consumption to agents spending money is just one step away. At this year's conference, Stripe's demonstration crossed that step.

Demonstration: Agent Buying and Selling

On the second day, a demonstration earned multiple rounds of applause.

John Collison gave an agent a simple instruction: research how AI demand affects the energy market. The agent began searching and discovered that Alpha Vantage had an energy market dataset it needed priced at 4 cents. The agent determined the price was within budget and then autonomously completed the purchase and download using a stablecoin wallet within Tempo CLI, as paying 4 cents by credit card was not cost-effective. It then generated a complete analysis report. This was already impressive. But John then instructed the agent: "Publish and sell this report. Set a price you think is reasonable so other agents can find and purchase it." The agent checked the licensing terms of the Alpha Vantage dataset, confirmed commercialization was allowed, then set up a website, published the report, and generated an instruction file enabling other agents to purchase the data with a request.

In just a few minutes, an agent completed the entire link from research, procurement, production, compliance review, publishing, pricing, to sales. It was both a buyer and a seller. After the demonstration, John remarked, "Agentic Commerce has arrived."

The other two demonstrations on the first day were equally impressive. Will Gaybrick built an API review application that allowed agents to obtain review services for users. Throughout the process, he did not provide the agents any payment information. During task execution, the agents automatically discovered that the application used the Machine Payments Protocol (MPP) and autonomously completed a $2 payment. The human only authorized once with a fingerprint. This zero-configuration payment discovery capability is at the core design of MPP. Developers do not need to write separate payment logic for agents; the agents can find it themselves.

Immediately afterward, Gaybrick combined Metronome (a real-time metering engine), Tempo (a blockchain designed for payments), and stablecoins to demonstrate streaming payments. An application charged in real-time based on AI token consumption, $3 per million tokens. Multiple agents ran simultaneously. The left dashboard displayed rising token consumption, while the right side saw stablecoin micropayments flowing in synchronously. When the Tempo blockchain browser was opened, a total payment of $3.30 consisted of thousands of micropayments, each a mere three-thousandths of a cent. Credit cards cannot do this, ACH cannot do this, UPI and Pix cannot do this. Gaybrick announced on stage that this was the world's first streaming payment business.

The Return of Micropayments and New Consumption Logic

Shopping through chat windows and OpenClaw are examples of agents representing human consumption. But in a group interview, Collison made a more ambitious judgment: agents could create entirely new demand.

He believes that agents could make a business model that has been discussed for years but never truly realized viable: micropayments. Humans struggle with making extremely granular consumption decisions. Spotify replaced single-song payments with a $9.99 monthly subscription because no one wants to decide whether a song is worth 15 cents every time they press play. Agents do not carry this cognitive burden. If this judgment is correct, then a whole class of business models that fail due to human cognitive friction may suddenly become feasible before agents. Maia expressed a similar view in one-on-one conversations with me. She mentioned she just spoke with dozens of AI founders, and when discussing agentic commerce, pricing was the most frequently mentioned topic.

Every transaction involves both buyers and sellers. If the buyer becomes an agent, what should sellers do?

In an interview, I asked Stripe product lead Jeff Weinstein: There is an old saying, "The customer is always right," and businesses need to please consumers. So how should they please agents? Jeff's response was to imagine agents as the best programmers you know. They want perfect information, structured formats, quick readability, and all the context needed for decision-making. Human consumers prefer pretty pictures and smooth animations, while agents want raw structured data, precise logistics information, and the ability to complete transactions with minimal steps.

In another conversation, Meta's VP of Product, Ginger Baker, summarized this shift even more radically: payments will move from "instant" to "strategic." Human consumer purchases are discrete. You walk to the checkout, pull out your wallet, swipe your card, and the transaction is complete. Agent consumption is continuous. You set a set of rules, such as "groceries not exceeding $50 this week," "always prioritize this card," "over $500 must be authorized manually." Then agents autonomously continue to consume within the authorization framework you set.

Security: Computational Power is the New Cash

If agents truly become a new kind of consumer, it will also bring new risks. These risks are fundamentally different from traditional SaaS transaction risks and the risks faced by human consumers.

During the Sessions, I paid special attention to this topic and discussed it with several Stripe executives.

Emily Glassberg Sands, Stripe's data and AI lead, described three rapidly growing fraud patterns. The first is multi-account abuse. The same person repeatedly registers different accounts, each claiming free allowances. According to Stripe network data, one in six AI company registrations involves such abuse. The second is malicious consumption during free trials. This is particularly deadly for AI companies, because every trial consumes real inference costs. She gave an example: a partner company incurs token costs of over $500 for each paying customer because it takes 25 free trials to convert one customer, with 19 of those being fraudulent. The third she calls "dining and dashing." Customers consume a large amount of tokens and refuse to pay at the end of the month. Emily also quoted a saying: "Computational power is the new cash." When traditional SaaS is abused, the marginal cost is nearly zero. But every inference call for AI companies incurs real costs. Stolen tokens are stolen money.

However, there exists a particular dilemma that I find troubling. Many AI founders' response to abuse is to shut down free trials.

Emily said she asked everyone who claimed to "solve" this problem how they did it, and found their solution was to directly shut down the free tier. But Jeff believed this creates another problem. Agents are becoming the primary way to discover new services. If agents cannot trial services by themselves, they will jump directly to another URL. Emily added that if the call to action presented to agents is "join the waitlist" or "contact sales," agents will immediately walk away. Closing self-service registration to prevent fraud could mean handing over the most important growth channel to competitors.

Stripe’s response to this dilemma is its fraud prevention system, Radar. The logic of Radar is simple: every time a transaction is completed on Stripe, Radar learns. Transaction data from 5 million businesses flows into the shared risk identification network. If one company encounters a certain fraud pattern, all companies can benefit. Last month, Radar blocked over 3.3 million high-risk free trial registrations across eight high-growth AI companies.

Jeff also proposed an counterintuitive point: agent shopping might ultimately be safer than human shopping on the web. Human web shopping trust verification relies on inference: how long a user stays on the website, whether the click path is normal, etc. Agent transactions can undergo programmatic authentication. Stripe's Shared Payment Tokens tokenize payment credentials, so agents never touch the original credit card numbers. Users authorize through biometrics and can set transaction limits, time windows, and merchant whitelists. When trust mechanisms shift from inference to confirmation, the security baseline could actually improve.

Ecology, Protocols, and a Piece of History

So far, it should be clear: a well-functioning ecosystem is essential for the realization of agentic commerce. At Stripe Sessions 2026, I met someone from the food industry. He said he attended the conference to learn whether agentic commerce could become a new opportunity for his company, which is a seller's perspective.

Therefore, this cannot solely rely on Stripe; it requires an ecosystem.

Wandering through the Sessions exhibition hall for two days, I saw booths from a multitude of companies in the financial industry chain. Stripe has also launched or joined a series of protocols with upstream and downstream partners, connecting different parts of the ecosystem: buyers and sellers, humans and machines, machines and machines. The Machine Payments Protocol (MPP) allows agents to discover and complete payments via HTTP. The Agentic Commerce Suite enables consumers to complete purchases directly within AI applications from Google, Meta, OpenAI, and Microsoft. The Universal Commerce Protocol (UCP) is a cross-platform commerce protocol initiated by Shopify, joined by Meta, Amazon, Salesforce, and Microsoft. Stripe has joined the UCP Governance Council. A group of companies that are both partners and competitors agreed to jointly establish a shared protocol, as fragmentation would make it difficult for agents to consume smoothly across platforms, which benefits no one.

Speaking of protocols, I saw a special Stripe partner in the exhibition hall: Visa. To me, Visa is essentially a protocol platform.

Noticing Visa immediately reminded me of a book I greatly admire: "One from Many," written by Visa's founder Dee Hock. One core theme in the book is how banks, currencies, and credit cards should be redefined in the electronic age. Money no longer has to be coins and bills; it can also be data, backed by institutions and recorded by networks, flowing globally. In the late 1960s, Bank of America rolled out its BankAmericard nationwide, attracting a flood of out-of-state consumers and causing the old system to collapse. Hock realized the problem was at the organizational level. Dozens of competing banks needed to share infrastructure, but the existing organizational forms could not allow them to cooperate and compete simultaneously. He used decentralized design principles to make all banks equal members of the new organization, with Bank of America relinquishing exclusive control over the system. This organization was later renamed Visa.

Thus, two different eras and two different companies are doing similar things; is there some lineage between them?

Any agent could easily provide answers. Patrick Collison has publicly paid tribute to Hock. After Hock’s passing in 2022, Patrick called him “a severely underrated innovator" with a profound influence on him and his brother. A more explicit signal was the hiring decision: David Stearns, a prominent historian of Visa, later joined Stripe.

Another detail that would elicit a knowing smile from someone familiar with payment history is that on stage, Tempo blockchain CTO Georgios Konstantopoulos showcased the lineup of validators. One name was Visa. Hock’s founded Visa has now become a participant node within the blockchain network incubated by Stripe. The students built the new network, and the teacher became one of its nodes.

When Patrick traced back the origins of Stripe's ideas during the conference opening, he mentioned he started as a Lisp programmer. The core idea of Lisp is "code is data." He translated this idea into Stripe's own language: "The basic notion of Stripe is that money is data. When we launched Stripe in 2011, this was not yet the canonical view in the industry." Hock approached the essence of money from an organizational theory perspective to conclude that money is merely "guarantee of value exchange." The medium that carries it can be anything. Collison entered this from the perspective of programming language, equating currency directly to data: data that can be programmed, called via API, and manipulated by agents. The two expressed the same reality in different languages. That day on stage, Ginger Baker stated even more plainly: "Isn’t currency just another form of digital content?"

If money is data, then it follows that data consumers will also become currency consumers.

Subplot: Stripe’s Content Gene

By now, the story of the AI economy is nearing its end. But let’s take a small detour; Stripe can almost be seen as a peer of content workers.

This company is not only skilled in financial services but also excels in content products. Its publishing brand Stripe Press has exquisite taste, and many know it for publishing "Poor Charlie's Almanack." Its podcast "A Cheeky Pint" is also distinctive, garnering a large audience. Google CEO Sundar Pichai, Anthropic CEO Dario Amodei, and a16z co-founder Marc Andreessen have all appeared on this podcast.

During Sessions, I met with Stripe Press Senior Editor Tammy Winter and designer Pablo Delcan. Tammy joked, "Stripe is a publishing house with a multi-billion dollar company attached." Pablo Delcan discussed his understanding of taste. He mentioned that taste is the result of long-term accumulation and requires time to settle. Regarding design trends, he believes the new question is how to add a certain degree of complexity through details and precision without sacrificing simplicity and clarity of communication.

When discussing books, Tammy told me that within Stripe Press, the series published for founders and builders is called the "Turpentine" series. These books focus on actionable knowledge, tools, techniques, maintenance, and practical matters that keep work running. They are not abstract theories but aim to help readers solve specific operational problems.

The name comes from a story about Picasso: when art critics gather, they discuss form, structure, and meaning; when artists gather, they discuss where to buy cheap turpentine. This series aspires to become the cheap turpentine for founders. If you think about it, for AI companies going global, Stripe's financial services are a form of this turpentine. You don't have to worry about payments, compliance, or forex; you can focus on building products.

This subplot may seem unrelated to the main plot, but there is an underlying connection. Stripe also has a magazine called "Works in Progress," focusing on how economies grow. Its podcast interviews leaders of the AI economy. The Sessions themselves, to some extent, resemble an economics lecture. On the second morning, John Collison delivered a talk on economic data, Coase's theories of the firm, and the Solow paradox. I guess a financial services company has such a keen interest in economics precisely because understanding structural economic changes is the way it discovers the next product opportunity.

As a podcast enthusiast, when I saw John Collison on the first day of the conference, the first question I wanted to ask wasn’t about finance but about podcasts. I asked him, after interviewing such a variety of people, was there an underlying question that runs through all the conversations. He pondered for a moment and said what truly interests him is how these companies operate, what competitive equilibriums they are in, and how they understand their businesses.

Coincidentally, there was a small twist at the end of the first day. The planned final fireside chat was Patrick interviewing OpenAI co-founder Greg Brockman, but just before going on stage, the guest was changed to Sam Altman. Patrick explained that "AI is a rapidly changing field" after all.

Thus, the surprise turned into joy. The audience erupted in applause.

The two have known each other for nearly 19 years. Altman was one of Stripe's earliest angel investors when the Collison brothers were still in their teens. As a result, Altman appeared extremely relaxed throughout the conversation.

As they neared the end, Patrick asked a personal question: why invest in two teenagers back then? Altman recalled that he remembered that the product they wanted to build addressed a problem they personally faced, and he saw that the opportunity could be scaled because many others also needed the same thing.

I felt that his answer regarding the podcast and his answer regarding the investment pointed to the same thing: finding real demand and solving real problems. In the conversation, Altman divided OpenAI's transformation into three stages: from research lab to product company, and finally to supplying intelligence to the world as a "token factory." Each stage corresponds to a different mission. Stripe is quite similar. In 2010, the problem that the two young Irish men were solving was "online payments are too difficult." Along the way, they have solved the same problem for 5 million users. In 2026, they have discovered a new problem: the clients of these businesses may soon no longer be human.

With one hand holding a podcast and the other holding a publishing house, discussing Coase's theory and the Solow paradox on stage, unfolding protocols and APIs in the exhibition hall, Stripe is not only creating the AI economy but also recording it. At the conference, I had a thought that sounds a bit crazy: Stripe holds trading data equivalent to nearly 2% of global GDP. It can see where every dollar of AI revenue comes from, where it goes, and how fast it grows. If Solow had such a heart rate monitor at the time, perhaps he wouldn’t have had to wait ten years to find computers in the statistics.

Maybe one day, Stripe can provide a model for the AI economy. Not a large language model, but a Nobel-level economic model. Who says that's impossible? Just a few years ago, who could have imagined that Demis Hassabis, the founder of DeepMind, would win a Nobel Prize?