Stripe 总裁 Will Gaybrick 解读 AI 战略与智能体商业

Brief Description

Stripe president Will Gaybrick discusses how the company has evolved from payments into a broad financial-infrastructure platform, and why AI should expand what companies build rather than merely cut costs. The conversation covers Stripe's product velocity, agentic engineering, agent commerce, stablecoins, token economics, and the company practices used to preserve product quality as it scales.

Table of Contents

• Stripe as financial infrastructure
• AI companies, fraud, and global expansion
• Startups, enterprises, and a growing market
• Building faster with founder-like agency
• Agentic engineering and smaller teams
• Expanding capacity instead of cutting for efficiency
• The beginnings of agent commerce
• Microtransactions and stablecoins
• Tempo and payments infrastructure
• Tokens as a new financial surface
• Scaling product quality and taste

Stripe as financial infrastructure

Stripe now thinks of itself less as a payments company with add-ons and more as a multi-product platform for financial infrastructure. Its purpose is to reduce friction and increase a business's agency: helping companies change their business models, operate in more countries, and move faster on everything connected to revenue and cash.

The product set has expanded from payments into billing, subscriptions, invoicing, Connect for platforms and marketplaces, Radar for fraud prevention, Tax, and many other products. There are roughly 25 to 30 headline products, with hundreds or thousands of underlying features. The organizing principle remains the same: remove friction and let users act more quickly.

AI companies, fraud, and global expansion

AI companies have exposed new forms of free-trial abuse. Before AI, a software company might lose only a negligible amount of compute when someone repeatedly opened trial accounts. AI products have a meaningful cost structure, so abuse can become expensive quickly. Stripe worked with customers to build a pipeline that combines network-level signals, embeddings, and a reasoning layer that can identify likely trial abusers and explain the relevant signals. ElevenLabs, for example, said it was blocking about 2,000 abusive free-trial users per day using Stripe signals.

The other side of the platform's work is helping digital-goods companies go global. Selling internationally requires registration, tax calculation, and remittance in many jurisdictions. Stripe Managed Payments lets a company remain the seller of record in its home markets while Stripe serves as merchant of record in long-tail markets where the company lacks an entity. Stripe handles tax calculation and remittance, allowing companies to cover more than 100 geographies compliantly.

Startups, enterprises, and a growing market

Stripe's strategy can be summarized as: win all the startups, then win them again. Startups become many of the most important companies of the future, reveal the next opportunities early, and have demanding standards. Large enterprises may say reporting is better than what they previously used; startups are more likely to say a product is inadequate and insist that it be fixed. That pressure makes Stripe better.

Working with a startup over time also pulls Stripe upmarket. The company aims to offer enterprises the same unexpectedly strong experience expected by startups, while recognizing that enterprise sales cycles and post-sale activation are different. Stripe works with major enterprises including Amazon and Microsoft, and a substantial, though not yet half, share of the Fortune 500. The core remains staying close to users, understanding what they need, and demonstrating better outcomes on the metrics they care about.

AI is increasing both company formation and the value of newer software cohorts. The discussion cites first-half signups growing 50% year over year and a 2026 cohort generating 50% more revenue than the comparable 2025 cohort; the 2025 cohort was itself generating 70% more than the comparable 2024 cohort. Products such as Suno and Higgsfield illustrate businesses that either could not have been built a few years earlier or would have been much less capable. Agentic coding has lowered the cost of building software, broadened the opportunity landscape, and driven especially strong growth in usage of Stripe Billing among people starting software companies.

Building faster with founder-like agency

Speed has always been near the top of Stripe's priorities, alongside relentless user focus. The company has borrowed useful operating mechanisms from other enduring institutions: goal setting similar to Google's OKRs, sales organization lessons from Microsoft, a directly responsible individual culture and product-review standards associated with Apple, and operating practices learned from Alan Mulally. The goal is not to copy any one company, but to create an institution that can outlast its current leaders and remain a container for entrepreneurship.

The AI era creates a situation with few clear precedents. A single engineer can now accomplish what two engineering teams could have done two years earlier. Stripe's response is to become an even stronger platform for founders internally, not merely externally. Founders from acquired companies such as Metronome, Privy, Bridge, and Lemon Squeezy have taken major leadership roles, and the company sees a similar opportunity for highly capable engineers.

Many companies have treated agentic efficiency as a reason to shrink operating expenses or headcount. Stripe's view is deliberately different: build everything. The best way to improve the cost structure is to grow more by addressing the large backlog of user requests. To do that, people need founder-like agency rather than being slowed by centralized process overhead.

The constraints have shifted to the back office. More code than ever is being merged, stressing systems for getting products into seller systems, onto pricing pages, and into market before sales teams can be trained. Stripe is optimizing the full critical path from a user request and product idea to a product in users' hands. The two objectives are to give everyone at Stripe more agency and to keep improving the tools that enable them.

Agentic engineering and smaller teams

Stripe created an internal system called Stripe Minions. A Minion is intended to be one-shot: a developer states what they want, the system builds it, takes it through CI/CD and testing, and then presents it for review. It is not primarily an iterative planning workflow. Stripe had reported about 1,200 Minion-created pull requests per week early in the year; the prior week, it had about 7,000, representing around 30% of all pull requests.

Engineers have long been central product leaders at Stripe because the company builds for engineers and technical users. The company is leaning further into that model by providing tools and platforms that enable engineers to take on more front-end development and design work themselves. The short version of the organizational change is flatter, smaller teams. Layers that once coordinated work can be valuable, but some of that coordination is no longer necessary when individual contributors are more powerful.

Stripe Projects was cited as an example: it was largely driven by a product manager, a senior engineer, and a small number of additional contributors, with most pull requests coming from one engineer. That engineer now orchestrates 16 agents from a screen and moves much faster. There is no fixed ideal team size. Flatter organizations may have wider teams and fewer management layers, while higher individual productivity may allow fewer people. Those forces can offset each other: a manager may still have a team of eight, but that team can do three times as much.

Expanding capacity instead of cutting for efficiency

The central question is where to direct newly available AI capacity. Stripe is putting it toward customer-facing products and adjacent opportunities, rather than solely toward internal cost optimization. Its internal knowledge tool, Kai, was built by two people and reached 83% weekly active use and about 60% daily active use at Stripe. Seller productivity rose 20%, but the conclusion was not that Stripe needs fewer sellers. If the payback from sellers is better, the company should have more of them.

The same applies to operations and engineering. Better tools make day-to-day work more productive and enjoyable and eliminate manual tasks. The discussion invokes Jevons paradox: when an asset becomes more productive, demand for it can increase rather than decrease. Stripe does not expect a world with fewer engineers just because engineers are more productive; it sees much more software creation. Its new SaaS-platform cohort in 2026 was said to be 103% larger than the prior year.

For early-career engineers, Stripe should increasingly function as an incubator: find a high-agency project that helps users, provide paved paths and good tools, and get out of the way. The aim is to build more useful adjacencies, not random products. A spend-management project adjacent to Treasury, which might once have been scheduled two years out, can now be picked up by one engineer and moved forward quickly.

There is still value in resource pressure: it drives disciplined, effective use of new tools. In the longer term, budgeting may look familiar—if a company is doing far more with current resources, additional resources can unlock even more. But Stripe sees an effectively infinite list of user asks. It took a long time to build automated U.S. filing for Stripe Tax across many state jurisdictions; global filing, a more complex problem, was built in about a third of that time. Stripe Treasury likewise addresses a long-standing request by letting users hold funds in dozens of currencies across countries and use a banking portal designed for global money movement, even though Stripe is not itself trying to be a bank.

Agentic engineering is compared to injection molding. Once a manufacturer has molds and repeatable patterns, it can produce many items efficiently. For code, the molds are templates, repository guidance, and patterns for integrating with financial institutions. With those foundations, agents can carry out more implementation work, while human effort concentrates on review; agents are also increasingly useful in review. The result is compressed delivery timelines and a much larger opportunity than cost cutting alone.

The beginnings of agent commerce

Agentic commerce has not yet had its Cambrian-explosion moment. There are not yet many canonical use cases repeating at scale, in part because key primitives are missing. Stripe and Tempo created a machine-payments protocol that lets a service indicate what must be paid and how. A request for an image or a piece of online content can return a 402 response that explains how to purchase it. The premise is that machines will increasingly buy from other machines.

What checkout should look like for agents remains open. Browser automation may allow agents to navigate existing forms, but that is a skeuomorphic version of the future. A native version is likely to emerge. One relatively non-speculative outcome is that conventional checkout pages should disappear or become unnecessary. With products such as Stripe Link and Shop Pay, much of the process is already simplified; eventually a person may be able to buy directly from a product display page.

Stripe has launched a Link agent wallet and a Link CLI so agents can use Link credentials with humans in the loop setting permissions. The company is particularly excited about B2B agentic commerce. Stripe Projects can scaffold applications, but its more important potential is enabling agents to provision B2B services. An agent could adopt a hosting service such as Vercel without the person visiting the provider's website. This suggests a useful investment and product question: if agents become the shoppers, is this the service they will want to choose?

An internal demonstration used Browserbase through an agent to research and fill out an NCAA tournament bracket. More practical services that agents can adopt on behalf of people are likely to be popular, while long-running consumer tasks still need more work. The focus today is making adoption of services easier for agents and allowing the social and product norms around that behavior to develop.

Microtransactions and stablecoins

The conversation argues that agents could make ephemeral, one-time consumption much more valuable. Someone who wants an AI service to create a song for a niece's birthday may not want to open a recurring monthly account. Instead, an agent should be able to discover a service, receive a budget such as $15, use the service, and pay. This requires micro-consumption APIs as well as the underlying commerce primitives.

Microtransactions have been discussed since the beginning of the internet, but agents change the economics. They increase human agency by doing tasks that would otherwise require time and many accounts. For complex tasks, agents may collect small pieces of data, temporary storage, and compute from many places. Users need those services to be secure, but should not have to create accounts everywhere. Microtransactions are therefore presented as necessary infrastructure for an agentic economy.

Stablecoins can make such payments practical. They are still awkward for many human users, who must pass through multiple steps, but an agent with an assigned budget can convert dollars to stablecoins or use a stored balance and handle the back-and-forth. This mirrors a broader shift in software from subscriptions and seats toward consumption pricing: consumers can access more services, while providers can serve people who would not commit to a subscription.

Stablecoins and global money movement

Stablecoins are described first as better infrastructure for moving money. Some countries have efficient national payment systems, such as UPI in India and Pix in Brazil. The discussion notes that a very high share of sub-$5 payments in India use UPI, while the comparable share for cards in the United States is in the single digits. National systems can be cheap and fast, but the global economy also needs a common coordination point. Crypto rails could solve that political and interoperability problem by offering a platform that works across borders.

If everyone already held stablecoins, global money movement could be faster, cheaper, and less frictional. That outcome is not guaranteed, but Stripe is working toward it. Stablecoins are native in Stripe Treasury, so users can hold them alongside balances in currencies such as U.S. dollars, euros, and pounds. Stripe is available in fiat in roughly 60 countries, while stablecoin access extends to around 150. This can bring more people into the online economy and make it easier to build an AI company from places such as Thailand or Brazil as well as the United States.

Adoption need not be wholesale. Stripe's broad offering lets enterprises and startups adopt stablecoins at the pace they choose. A remittance company operating between the United States and Mexico was cited as having built on stablecoins and reached between 5% and 10% of remittances on that corridor within a few years. Reaching that share so quickly compares favorably with the long path taken by earlier fiat-money-transfer businesses.

Tempo and payments infrastructure

Tempo is motivated by the particular requirements of payments rather than the idea that the world simply needs another general blockchain. Payment infrastructure needs privacy, since public blockchains can expose patterns that others can reverse-engineer. It needs reliable throughput, and transaction fees that do not spike when trading activity elsewhere overloads the system.

The aim is to move money as efficiently, consistently, and cheaply as possible. Tempo is still early, but the discussion describes encouraging traction and work with companies including DoorDash. Stripe is optimistic that a payments-specific blockchain can become a useful default rail without being the only blockchain used in the Stripe ecosystem.

Tokens as a new financial surface

Tokens increasingly resemble money. The sophisticated attacks on general-purpose token consumers such as Cursor and Replit resemble attempts to steal money from Stripe users. That blurs the line between tokens and dollars and gives Stripe a mandate to protect users as they move, store, and send tokens with the same safety and compliance expected for money.

Over time, tasks that once required payments to people may increasingly be performed or augmented by agents and paid for in tokens. Stripe wants movement between tokens and dollars to become as seamless and safe as movement between dollars and euros. This is early, but it is expected to become an important part of Stripe's future.

There are two related token-management problems. One is operating-expense management: using tokens to build things and understanding their spend. The other is product efficacy: choosing among models and token usage patterns so a token-driven product works as effectively as possible. Stripe is considering how it can help customers with both, while also improving spend management, cash management, revenue operations, closing books, globalization, and growth across the full revenue stack.

The discussion acknowledges a hypothesis that AI could commoditize software severely, perhaps even as models recursively generate models. But the observed reality is the opposite: software creation is accelerating, customer monetization is faster, and companies are adopting more Stripe products. ElevenLabs, for example, was said to use 14 Stripe products. Building software is only part of the challenge; companies also need durable expertise about how abstractions work across the business, especially as employees leave and institutional knowledge decays.

Scaling product quality and taste

The final topic is how to preserve Stripe's product quality and taste as AI performs more work. The speakers note that "taste" may sometimes be used as a vague defense of human value, and that it is not obvious models will never have it. But product quality remains central to Stripe's culture and identity. Carefully crafted tools help users move faster and make the hard journey of building a company more enjoyable. They also make the work of creating products more rewarding than shipping something that barely works.

Scaling quality requires two things. First, leadership must state the standard repeatedly. Just as the company repeats its strategy of winning startups and winning them again, it must keep talking about quality. Second, people must use the product. Stripe invests heavily in simulated product usage, drawing an analogy to aviation simulation and to Nvidia's ability to simulate chip performance before waiting for fabricated chips.

Teams can point to an account pattern and create a privacy-safe, randomized simulation of its day-to-day reality: changing growth rates, seasonality, disputes, refunds, and other operational problems. The events are not real customer activity, but they let teams experience what the user experiences. Engineering managers are asked to lead this work because they control resources and can decide when a product does not feel good enough and the next sprint must elevate quality. In this view, scaling taste is both culture and daily practice: set the standard, use the product, and give people better tools to act on what they learn.