Stripe Data Shows AI Solopreneurs Drive Post-SaaS Expansion
Stripe
Artificial intelligence is shifting global commerce toward lean solopreneurs and autonomous agents while dramatically raising the value of proprietary data, energy, and physical infrastructure.
## Background
Published by Stripe on YouTube on May 19, 2026, this keynote video from the Stripe Sessions 2026 conference focuses on the deep macroeconomic trends and developments shaping the next decade of global trade, business models, and payments. The accompanying page description highlights that the session delves into these transformative trends while providing registration and video links for other keynotes and fireside chats from the event.
## TL;DR
In this opening keynote, Stripe's John Collison and Dr. Emily Glassberg dissect deep economic data to challenge common misconceptions regarding SaaS valuations, tariffs, and a K-shaped economy. They demonstrate how artificial intelligence is accelerating business velocity, expanding cross-border trade, and fueling an unprecedented global surge in highly automated solopreneurs. Grounding their analysis in Coasean economic theory, they introduce five levels of agentic commerce alongside a live demonstration of autonomous software-to-software micropayments. Ultimately, they outline five key economic complements that will dramatically gain value as raw intelligence cheapens, framing the current AI era as a profound structural re-platforming of the global economy.
## Atmosphere & style
The presentation carries an intellectually stimulating and highly forward-looking atmosphere, balancing rigorous economic analysis with an engaging, tech-forward energy. The dynamic between the speakers is collaborative, grounded, and authoritative, translating academic concepts like Coase's theory of the firm into actionable, real-world insights. Watching the session feels deeply immersive and validating for tech builders, moving from analytical scrutiny of market data to lighthearted curiosity during live agentic software demonstrations. It leaves the audience with a sense of rapid momentum, painting a compelling picture of an economic transformation that is unfolding at an unprecedented pace.
## Key takeaways
- [00:00](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=0s) Software market fears over AI substitution remain prospective, with Stripe's real-time payment data showing SaaS transaction volume and underlying revenue continuing to grow steadily.
- [01:00](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=60s) Equity markets have returned to prioritizing strict profitability, heavily rewarding highly profitable firms rather than inflating unproven valuations.
- [02:20](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=140s) Tariff policies introduced in recent years inverted the historical trend of durable goods deflation, with businesses increasingly passing higher import costs to consumers.
- [03:18](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=198s) Stripe platform data contradicts the 'K-shaped' economy narrative, showing a narrowing spending gap between high-income and low-income consumer demographics.
- [03:58](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=238s) Post-pandemic recalibrations, tighter immigration, and higher interest rates—rather than AI job destruction—are the primary drivers of recent labor market cooling and unemployment trends.
- [05:37](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=337s) AI's impact on macroeconomics is most visible in business formation metrics, driving a global surge in high-earning, single-person non-employer firms (solopreneurs).
- [07:46](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=466s) Modern AI startups follow a new playbook, launching internationally on day one, maintaining minimal headcount, and acquiring cross-border revenue at unprecedented speeds.
- [10:46](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=646s) Applying Coasean economic theory, AI agents lower external transaction and coordination costs, enabling smaller firms to achieve high output through market mechanisms rather than large internal hierarchies.
- [13:30](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=810s) Agentic commerce is progressing from frictionless form-filling and reasoning discovery toward fully autonomous financial execution by AI agents.
- [17:46](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1066s) Machine-to-machine micropayments rely on low-friction financial rails like stablecoins to make sub-cent agent transactions economically viable.
- [23:47](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1427s) As AI makes raw intelligence cheap and abundant, economic value shifts toward its complementary assets: compute infrastructure/energy, proprietary data, network effects, real-world execution/hardware, and complex system interfaces.
- [28:12](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1692s) Much like historical adoption curves for electricity and early computing (the Solow Paradox), the full productivity impact of AI requires structural re-platforming before reflecting in broad macroeconomic stats.
Opening the keynote at Moscone, John Collison welcomes attendees to Stripe Sessions 2026, noting the palpable tectonic shifts happening across the tech landscape. Reflecting on the prior day’s presentation by Will Gaybrick—who unveiled 288 new Stripe products designed to assist businesses as the economy re-platforms around artificial intelligence—Collison shifts to his own passion: analyzing deep economic data. Before diving into major macroeconomic shifts, he takes stock of the turbulent conditions since the previous year's event.
The past year brought significant volatility to the software industry. In January 2026, the software sector lost $1 trillion in market value over a span of less than 30 days. This sell-off was fueled by fears that AI makes software more abundant, easily substitutable, and less sticky than traditional SaaS models implied. However, Collison highlights that this panic remains purely prospective. Real-time transaction data from Stripe reveals that SaaS payment volumes are actually higher today than they were before the market downturn, demonstrating that underlying SaaS revenue continues to grow steadily.
Instead of a structural collapse in SaaS, what has actually returned to fashion across equity markets is a strict focus on profitability. Historical market trends show that during the dot-com boom, market valuations became detached from reality, briefly valuing the least profitable companies higher than the most profitable ones. Today, despite widespread concerns about market bubbles, equity markets are behaving "tediously rational," heavily rewarding the most profitable firms with outsized valuations.
Addressing global trade, Collison looks back at the trade policies and tariffs introduced over the past year following what he calls "Liberation Day." While early consensus suggested tariffs were an economic "nothing burger" because overall trade flows ended 2025 only slightly below 2024 levels, Stripe’s analysis shows a different picture emerging underneath. Historically, durable goods experience steady intra-year deflation due to manufacturing productivity gains and global competition—making products like flat-screen TVs and washing machines cheaper over time. In 2025 and continuing into 2026, that downward slope completely inverted. Businesses initially absorbed the higher import costs, but quarter by quarter, they have begun passing those costs directly through to consumers.
- **$1 Trillion:** Market value lost by the software sector in under 30 days during January 2026.
- **SaaS Payment Volumes:** Stripe transaction data shows higher overall volume today than prior to the software stock sell-off.
- **Durable Goods Inflation:** Tariff costs have broken the multi-year trend of durable goods deflation, with price increases continuing to pass through to consumers three months into 2026.
---
[03:18](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=198s) Debunking the K-Shaped Economy and AI Job Loss Myths
Collison examines two widely discussed economic narratives: the rise of a "K-shaped" economy and the concern that AI is causing widespread job destruction. Popular media and viral social media posts have highlighted a reliance on high-income consumers, citing examples such as United Airlines' reconfigured 787-9 cabin map dominated by business class seats, or Delta Air Lines reporting that premium seat revenue surpassed economy revenue for the first time. However, Stripe’s transaction data directly contradicts this narrative.
When tracking the ratio of high-income to low-income consumer spending across Stripe’s platform, the line is not slanting upward as a K-shaped model would predict. Instead, it shows a gradual downward trend, indicating that the spending gap between high-income and low-income demographics is shrinking rather than expanding. Stripe’s underlying data does not reflect an over-reliance on high-end luxury consumers to sustain the economy.
Turning to the labor market, Collison acknowledges that US unemployment has ticked up roughly one percentage point over the past three years. While public concern frequently blames AI for job losses, the primary factors driving the labor market cooling are macroeconomic adjustments and post-pandemic recalibrations. Companies are still unwinding the aggressive hiring binges of 2021 and 2022, tighter immigration policies have altered labor supply, and sustained higher interest rates have increased the cost of capital. The Federal Reserve's end-of-2026 interest rate forecast sits half a percentage point higher than projected two years prior, enforcing tighter monetary conditions that naturally slow hiring. While AI will have a profound long-term impact, it remains just one variable among many and is not the dominant driver of current unemployment trends.
- **High vs. Low Income Spending Ratio:** Stripe's platform data indicates a shrinking gap between high-income and low-income consumer spending.
- **US Unemployment:** Up roughly 1 percentage point over a 3-year period.
- **Fed Rate Projections:** The Federal Reserve's forecast for the end of 2026 is 0.5% higher than predicted two years ago, dampening hiring through tighter credit conditions.
- **Stripe Global Presence:** Stripe currently processes almost 2% of global GDP, providing a direct lens into macroeconomic shifts.
---
[05:37](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=337s) The Surge in Solopreneurs and Rapid Global Scaling
To detail structural shifts in global economic dynamism, Collison invites Dr. Emily Glassberg, Stripe’s Head of Data and AI, to the stage. Glassberg addresses the cognitive dissonance felt by AI enthusiasts who wonder why AI isn't showing up clearly in top-line macroeconomic stats like labor numbers. She explains that AI is indeed appearing in macro data—if one knows where to look, specifically in business formation metrics.
US Census Bureau data reveals a dramatic spike in new business formations during the pandemic that is now re-accelerating. Crucially, virtually all recent growth is driven by "non-employer firms," commonly known as solopreneurs. While single-person operations were historically viewed as minor side projects, AI tools have enabled them to achieve meaningful enterprise scale. Today, nearly 5 million Americans earn their living as solopreneurs, with a rapidly growing cohort generating over $100,000 in annual revenue. This phenomenon is global: new business registrations have surged 40% in Australia, 70% in Finland, and 80% in France.
This structural surge is mirrored directly inside Stripe’s product ecosystem:
- **Stripe Atlas Milestone:** Stripe recently passed its 100,000th company incorporated via Stripe Atlas. The 100,000th company, Empirical, founded by Rachana, builds AI software to optimize profitability for battery energy storage systems.
- **Accelerated Revenue Velocity:** Atlas companies incorporated in 2025 earned twice as much revenue as the 2024 cohort did at the same point in their lifecycle. Remarkably, the class of 2026 is tracking to generate **five times** the revenue of the 2025 cohort over the same timeframe.
- **Cross-Border Growth:** Five years ago, only 11.6% of Stripe businesses earned the majority of their revenue outside their home country; today, that figure has doubled.
- **Long-Tail International Expansion:** Among companies making most of their revenue cross-border, about 25% derive the majority of that income from outside the top 10 global markets.
- **AI Startup Playbook:** The median top 100 AI startup on Stripe earns most of its revenue internationally and sells into 55 countries within its first year. For example, Emergent Labs (founded in the US in 2024) generates nearly 70% of its revenue internationally, with 16 distinct countries each accounting for at least 1% of its top line.
```
Stripe Atlas Cohort Revenue Trajectory (Relative to Class of '24 Baseline)
2024 Cohort: [1.0x Baseline]
2025 Cohort: [2.0x Baseline] (2x growth vs '24)
2026 Cohort: [10.0x Baseline] (5x growth vs '25)
```
> "This is a totally new startup playbook. You launch globally on day one, you keep headcount very lean, and you automate aggressively." — **Emily Glassberg**
---
[10:46](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=646s) Coase's Theory of the Firm in the Age of AI
Collison and Glassberg contextualize these structural shifts through the lens of economic history. Collison shares the story of British economist Ronald Coase, who in 1931—as a 21-year-old traveling across Depression-era America—questioned why centralized corporations exist inside a broader market economy. Coase observed that while free markets rely on price mechanisms, internal corporate structures function like mini centrally planned economies. His insight, which eventually won the Nobel Prize in Economic Sciences 60 years later, was that firms exist because the transaction and coordination costs of operating internally are often lower than navigating open markets.
Applying Coasean economic theory to AI, Glassberg breaks down the near-term and medium-term impacts on organizational structures:
1. **Within-Firm Efficiency (Near-Term):** Organizations leverage AI quickly internally because employees share context, unified systems of record, and aligned incentives, lowering internal coordination costs.
2. **Market Efficiency (Medium-Term):** AI agents dramatically lower external market coordination costs. Autonomous agents simplify product discovery, make software integration trivial, streamline contract negotiations, and execute micro-transactions at frequencies and scales impossible for humans.
As external transaction costs drop, the traditional necessity for large corporate hierarchies diminishes. The market is shifting toward fewer employees per firm, significantly higher output per firm, a higher overall count of independent businesses, and market-based coordination mechanisms taking over work previously held inside large organizations.
> "Coase would have liked that idea. He always thought firms were an inefficient setup. The more we can do via markets, the better." — **Emily Glassberg**
---
[13:30](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=810s) The 5 Levels of Agentic Commerce and Live Research Demo
Collison shifts to the second major structural trend: the transition toward fully agentic commerce. Stripe frameworks this transition across five distinct levels of purchasing autonomy, moving from simple user assistance to full agent decision-making:
- **Level 1 (Form Filling & Frictionless Checkout):** Software automatically populates user details and handles form friction on behalf of the user. An example is Stripe’s native in-app checkout experience with Meta, where an agent leverages stored credentials to complete purchases seamlessly.
- **Level 2 (Reasoned Search & Discovery):** Transitioning beyond traditional keyword search to reasoning assistants that evaluate constraints. Examples include asking ChatGPT for non-obvious birthday gifts under $100 for a brother interested in calisthenics, or using Wayfair’s AI assistant to scan catalogs based on mood or room descriptions.
- **Levels 3–5 (Autonomous Decision-Making & Execution):** AI agents independently evaluating choices, making financial commitments, and executing transactions within set parameters.
To illustrate early frontier demand for autonomous transactions, Collison points to OpenClaw (Claw), where payment-related skills recorded approximately 125,000 cumulative downloads within a 12-week span despite high technical setup friction.
Transitioning to a live demonstration of software-to-software commerce, Collison demonstrates voice interaction with Anthropic's Claude. Rather than typing, Collison verbally prompts Claude: *"Hey Claude, how is AI demand affecting commodity prices and supply and demand for different energy sources?"* The agent initiates real-time web research to analyze how massive data center capital expenditures and power grid constraints collide with long-standing US energy market equilibria.
- **125,000 Downloads:** Cumulative downloads of payment skills on OpenClaw within 12 weeks.
- **Levels of Autonomous Commerce:** Defined from Level 1 (automated checkout execution) to Level 5 (fully autonomous financial agents).
Here is a detailed, faithful walkthrough based on the transcript:
[17:46](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1066s) Agentic Data Purchasing and Micropayments Infrastructure
The demo continues with an AI agent analyzing the presenter's request, evaluating necessary data sources, and deciding to purchase commodity and equity data from Alpha Vantage to ground its analysis. Finding the total cost of 4 cents to be well within its allocated budget, the agent proceeds to buy and download the data autonomously. This marks a fundamental shift from existing AI workflows—where models generate ideas but force human users to execute the "grunt work" like deployments, checkouts, or sign-ups—toward agents fully executing tasks on their own.
To facilitate these machine-to-machine transactions, the agent uses the Tempo CLI and a dedicated stablecoin wallet. While traditional machine payments can run on fiat, tiny micro-transactions require a different financial infrastructure. Stablecoins offer near-zero transaction costs, making sub-cent machine transactions viable for the first time. As the process computes, the presenter jokes lightheartedly about needing a guitarist on stage to fill the waiting time, noting that despite using a paid "fast mode," complex agentic reasoning still takes a few moments to resolve.
- **Data source purchase:** 4 cents for Alpha Vantage commodity and equity data, evaluated and approved by the agent against its budget.
- **Payment infrastructure:** Uses Tempo CLI paired with a stablecoin wallet, leveraging near-zero transaction fees essential for AI micropayments.
[19:19](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1159s) Interlude: Works in Progress and Agentic Side Projects
While waiting for the agent to finish processing, the presenter highlights several articles and experiments featured in *Works in Progress*, Stripe’s magazine focused on economic and technological development. These examples illustrate the novel ways AI and data are being applied, alongside historical lessons on infrastructure.
- **Nuclear Energy in Britain:** An article exploring how Britain achieved rapid historical progress in nuclear technology before stalling out due to administrative and political hurdles.
- **The Guinness Index:** A real-world application where an AI agent autonomously phoned every pub in Ireland to map real-time prices for a pint of Guinness across the country.
[20:00](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1200s) Monetizing and Publishing Agent-Generated Reports
Once Claude completes its data processing, it opens a generated report in the browser showing significant macroeconomic trends in energy and technology: US data center electricity demand is projected to nearly triple by 2080, hyperscaler capital expenditure has surged 62%, and natural gas prices have risen 104% to meet energy demands.
The presenter then instructs Claude to publish and commercially monetize the report, allowing the model to choose an appropriate price point for other agents or humans. After briefly verifying that the Alpha Vantage terms of service permit commercial redistribution, the agent builds and deploys a live web application via Vercel.
- **Live Site:** Deployed at `johnsreport.vercel.app`, allowing human visitors to purchase the report for $5 via Link.
- **Machine Readability:** Includes an automatically generated `llms.txt` endpoint containing instructions and Tempo API parameters so other AI agents can programmatically buy the document in a single request.
- **Supported Ecosystem:** Highlighting early pioneers enabling agentic commerce, including **Parallel** (browser-based web browsing for agents) and **Postalform** (allowing AI agents to physically mail physical letters).
[23:47](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1427s) Economic Complements: What Grows More Valuable as Intelligence Cheapens?
Transitioning into economic theory, the presenter poses a central question: In a world where AI-driven intelligence becomes cheap and abundant, what becomes vastly more valuable? According to classic economic rules, when a core input's cost collapses, its **complements**—the supporting goods or assets required to utilize it—see a dramatic spike in value.
Historical parallels illustrate this joint demand curve:
1. **Containerized Shipping:** When containerization collapsed shipping costs, deep-water ports capable of handling massive container ships exploded in value.
2. **Spectrum Auctions:** Early radio spectrum auctions in the 1990s brought in hundreds of millions of dollars; once mobile phones became ubiquitous and cheap, the value of that same spectrum rocketed into tens of billions of dollars.
When applied to artificial intelligence, five key complements stand to gain immense value as raw intelligence is commoditized:
1. **Compute Infrastructure and Energy:** GPUs and dedicated chips shifted from gaming components to core computing assets. Similarly, nuclear power is undergoing a resurgence to power data centers, while gas turbine demand has surged, driving market valuations up for key manufacturers like Siemens Energy.
2. **Proprietary Data:** Proprietary data assets become far more valuable when superintelligent models can reason over them. Companies that once provided data for free have shut off automated web scraping to monetize their datasets. For example, Reddit's non-advertising data licensing revenue reached $35 million per quarter, and Stripe's proprietary data powers increasingly accurate fraud models in Stripe Radar.
3. **Network Effects:** Two-sided marketplaces remain vital as venues for buyers and sellers to meet. Across ten major marketplaces, take rates remained flat until a post-pandemic rise driven by better AI-driven matching techniques.
4. **Real-World Execution and Hardware Integration:** Software systems tied tightly to physical equipment create enduring competitive moats. John Deere exemplifies this by integrating vision sensors, GPS guidance, and machine learning into heavy machinery deployed across 130 countries—a moat built on physical presence rather than code alone.
5. **Complex System Interfaces:** Systems capable of managing complex, cross-platform interactions between legacy operations and modern software retain high strategic value.
[28:12](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1692s) The Solow Paradox and Re-Platforming the Global Economy
Concluding with a historical reflection on technology adoption, the presenter draws a parallel to Thomas Edison lighting up Lower Manhattan with six dynamos in 1882. Despite the debut of electricity, national productivity growth remained flat—and even slowed—for nearly 30 years. The bottleneck was not the core technology, but the structural layout of the economy: factories built around centralized steam engines with overhead shafts and belts had to be entirely redesigned from scratch to take advantage of unit-electric motors. It was not until the 1910s that worker output per hour doubled.
A similar lag occurred with the advent of computing, famously summarized by economist Robert Solow in 1987: "You can see the computer age everywhere but in the productivity statistics." It took until the mid-1990s for computing to visibly impact macro productivity stats.
"Transformative technology looks for a long time like it's not doing much. If you're looking at the economic gauges, you're going to sit there and you're tapping the gauges, is this thing on?"
The AI economy is currently navigating this exact digestion phase. While foundational metrics are changing—solopreneurs scaling to seven figures, agent-to-agent transactions, and day-one global launches—these shifts are early indicators of an economy re-platforming itself. While electrification took three decades to unlock its full productivity dividend, the transition to an AI-driven economy will happen much faster.
## Notable quotes
> [08:57](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=537s) This is a totally new startup playbook. You launch globally on day one, you keep headcount very lean, and you automate aggressively. — Emily Glassberg
> [10:46](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=646s) Coase would have liked that idea. He always thought firms were an inefficient setup. The more we can do via markets, the better. — Emily Glassberg
> [15:30](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=930s) Hey Claude, how is AI demand affecting commodity prices and supply and demand for different energy sources? — John Collison
> [28:12](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1692s) Transformative technology looks for a long time like it's not doing much. If you're looking at the economic gauges, you're going to sit there and you're tapping the gauges, is this thing on? — John Collison
## Mentioned
Stripe, John Collison, Will Gaybrick, Liberation Day, United Airlines, Delta Air Lines, Federal Reserve, Emily Glassberg, US Census Bureau, Empirical, Rachana, Emergent Labs, Ronald Coase, Meta, ChatGPT, Wayfair, OpenClaw, Anthropic, Claude, Alpha Vantage, Tempo, Works in Progress, Guinness, Vercel, Link, Parallel, Postalform, Siemens Energy, Reddit, Stripe Radar, John Deere, Thomas Edison, Robert Solow
Speaking at Stripe Sessions 2026, John Collison and Dr. Emily Glassberg argue that artificial intelligence is already restructuring the global economy through lean solopreneurship, agentic commerce, and new complement assets, even if aggregate productivity indicators exhibit a historical adoption lag.
Macroeconomic Realities vs. Prevailing Narratives (0:00)
Collison challenges several popular economic assumptions using platform and market data. While the software sector shed $1 trillion in market value in January 2026 over fears that AI would commoditize SaaS, Stripe transaction data indicates SaaS payment volumes remain higher than before the downturn. Equity markets have shifted toward penalizing unprofitable firms rather than abandoning software altogether. Additionally, platform figures contradict the idea of a 'K-shaped' economy driven solely by luxury consumption; the spending gap between high- and low-income demographics on Stripe is gradually narrowing.
Labor Market Drivers and Tariff Pass-Through (3:18)
Collison attributes the approximately one percentage point rise in US unemployment over the past three years to macroeconomic adjustments rather than AI automation. Key drivers include the unwinding of pandemic-era hiring surges, tighter immigration policies, and sustained high interest rates, with Federal Reserve end-of-2026 rate projections sitting 0.5% higher than anticipated two years prior. On trade, Collison observes that new tariffs broke the multi-year intra-year deflationary trend for durable goods, with businesses systematically passing higher import costs onto consumers into 2026.
Solopreneurship and Coasean Organizational Shifts (5:37)
Dr. Glassberg points out that AI's macro footprint is visible in business formation data. Nearly 5 million Americans now operate as non-employer solopreneurs, bolstered by tools that enable significant revenue scale. Stripe Atlas recorded its 100,000th incorporation, with the 2025 cohort doubling the revenue velocity of 2024, and the 2026 class tracking at five times the 2025 baseline. Furthermore, the median top 100 AI startup on Stripe sells into 55 countries within year one.
Agentic Commerce and Machine Micropayments (13:30)
Stripe categorizes agentic commerce into five levels, spanning basic form-filling up to autonomous financial decision-making. To demonstrate the frontier, Collison demonstrated an Anthropic Claude agent independently purchasing research data from Alpha Vantage for 4 cents using the Tempo CLI and a stablecoin wallet, which provides the near-zero transaction fees necessary for machine micropayments. The agent subsequently generated, monetized, and deployed a live web report equipped with an `llms.txt` endpoint to allow automated purchases by other software agents.
Economic Complements and the Solow Paradox (23:47)
Collison argues that as raw intelligence cheapens, classic economic theory dictates that complementary assets experience sharp value increases. These complements include physical compute and power infrastructure, proprietary datasets (such as Reddit licensing data or Stripe Radar fraud models), network-effect marketplaces, and tightly integrated physical hardware like John Deere machinery. Collison compares today's delayed productivity statistics to the Solow computer paradox of 1987 and the decades-long lag of factory electrification, concluding that the AI structural transition is quietly underway.
Collison and Glassberg view the current economy as navigating a structural digestion phase where corporate hierarchies shrink and autonomous agents interact. While aggregate macroeconomic productivity statistics have yet to reflect these advancements, Stripe anticipates this re-platforming will materialize far faster than previous industrial transitions.
## Background
该内容由 Stripe 官方于 2026 年 6 月 7 日发布在 YouTube 平台，录制自年度科技盛会 Stripe Sessions 2026 的主题演讲。根据页面官方简介，本场演讲聚焦于分享即将塑造未来十年全球支付演进与商业贸易格局的核心宏观与技术趋势。
## TL;DR
Stripe 联合创始人 John Collison 与数据负责人 Emily Glassberg Sands 依托平台底层海量交易数据，接连破除了 SaaS 崩盘与 K 型消费分化等舆论迷思，揭示了“一人公司”激增与初创企业“首日全球化”的新范式。演讲结合科斯定理阐释了外部协调成本暴跌带来的组织微型化，并现场生动演示了 AI Agent 使用稳定币钱包进行微支付、自主购买数据并一键建站分销研报的智能体商务闭环。最后，演讲总结了廉价智能时代算力、能源、专有数据、网络效应与物理执行力这五大高价值互补品，并借电气化历史的索洛悖论，指出全球商业底层架构正在经历一场远超以往迭代速度的系统性重构。
## Atmosphere & style
整场演讲洋溢着极客技术探索与宏观经济学严谨思辨相交织的生动氛围，台上的经济学博士与风趣的创业者互动频密、妙语连珠，将沉重的数据论证转化为举重若轻的智力漫游。在现场演示 Claude 借助稳定币以 4 美分微支付自主买数据并建站发售时，屏幕上飘落的庆祝彩带与台下的掌声将极客科技的酷炫感与前沿商业落地的确定性推向高潮，带给观众一种亲眼见证历史齿轮正在加速咬合的振奋与笃定。
## Bulletin
- [00:00](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=0s) Stripe 交易数据破除“SaaS 衰退”与“K型经济复苏”迷思：平台 SaaS 交易量远超抛售潮前水平，且高低收入群体消费差距正在逐步收窄而非极端分化。
- [05:36](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=336s) 新型“一人公司”与初创范式爆发：全美年营收超 10 万美元的无雇员企业已近 500 万家，Stripe Atlas 2026 届初创企业同期营收轨迹达到 2025 届的整整五倍，且前百大 AI 初创企业首年向平均 55 个国家销售产品。
- [13:30](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=810s) 商业全面进入 Agent 化阶段，现场演示 Claude 借助 Tempo CLI 稳定币钱包完成 4 美分微支付并自主定价上架研报，基于 llms.txt 开启软件向软件自主采购的新形态。
- [23:46](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1426s) 根据互补品经济学理论，随着推理智能成本归零，算力硬件、基荷能源、专有数据、双边网络效应以及打通物理现实执行能力的资产将迎来价值倍增。
## Key takeaways
- [00:00](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=0s) 软件板块市值在 30 天内蒸发 1 万亿美元属于前瞻性恐慌，Stripe 底层 SaaS 处理量依然高于抛售前，且公开市场对盈利确定性极高的企业赋予了更高溢价。
- [02:04](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=124s) 关税政策并非“雷声大雨点小”，耐用消费品价格在 2025 至 2026 年初逆转了多年的通缩走势并加速上涨，企业正持续将供应链成本转嫁给消费者。
- [03:16](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=196s) 达美航空高端舱收入超经济舱等轶事助长了 K 型经济叙事，但 Stripe 真实数据显示高低收入者支出比率持续平缓下滑，底层消费动能依然广泛。
- [04:21](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=261s) 劳动力市场降温主要受疫情期间过度招聘出清、移民政策趋紧及高利率资金成本等传统宏观因素主导，AI 绝非当前失业率微升的决定性推手。
- [05:36](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=336s) AI 工具赋予了单人创业者（一人公司）极高的规模化杠杆，全球主要发达经济体的新企业注册量同步激增 40% 至 80%。
- [07:43](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=463s) 新一代初创企业颠覆了“先本土后出海”的传统扩张顺序，呈现出“上线首日即全球化、极简人头编制、全流程激进自动化”的三大范式特征。
- [10:52](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=652s) 科斯定理在 AI 时代被重新激活：AI 大幅削减企业间的信息发现、接入与签约成本，将驱动组织变得更小更敏捷，更多协调机制重归外部市场。
- [13:30](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=810s) 商业的 Agent 化经历从表单代填（Level 1）、约束推理导购（Level 2）到全权自主决策（Level 3-5）的跃迁，开源生态对支付技能的需求在 12 周内即达 12.5 万次。
- [16:21](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=981s) 在软件向软件的自主微采购中，稳定币因接近零的手续费特性成为打通机器间分级微支付（如 4 美分调用）的核心底层结算管道。
- [21:16](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1276s) 通过在网站根目录部署机器可读的 llms.txt 标准规范，外部 AI 智能体仅凭单次请求即可自主调取支付与履约逻辑，实现端到端无人化交易。
- [24:46](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1486s) 智能要素的极度廉价将直接引发互补品价值暴涨，包括以英伟达为代表的硬件互连网络，以及以核电和燃气轮机为代表的基荷电力供给。
- [25:45](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1545s) 高防爬壁垒的私域专有数据价值正在被大模型深度激活，如 Reddit 单季度实现 3,500 万美元数据授权收入，Stripe Radar 欺诈防护精度也大幅提升。
- [26:31](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1591s) 软件构建门槛归零抹平了工具切换壁垒，使头部平台型企业的撮合网络效应更具黏性，其抽成率在过去三年内依托 AI 优化持续走高。
- [27:20](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1640s) 单纯的纯软件代码不再构成决胜护城河，类似约翰迪尔将算法与传感器阵列深植全球 130 国田野的物理世界综合执行力才具备持久壁垒。
- [28:12](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1692s) 索洛悖论与电气化历史表明，革命性技术往往需要长达数十年的组织架构重塑才能兑现宏观生产率红利，但 AI 驱动的底层再平台化重塑进程将远快于历史。
## Detailed walkthrough
### [00:00](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=0s) 宏观审视：SaaS 抛售恐慌与重回理性的资本市场
Stripe 联合创始人兼总裁 John Collison 登台，热情欢迎大家回到旧金山莫斯康展览中心（Moscone Center）。他感慨道，如今几乎能切身感受到“脚下的地壳板块正在剧烈震动”。就在前一天，Stripe 联席总裁 Will Gaybrick 刚刚发布了 288 项全新产品功能，旨在协助企业应对由 AI 引领的底层经济架构重塑。不过，John 表示今天他要带大家进入自己的“快乐老家”——深入剖析宏观经济数据。在深入未来趋势之前，他首先带领观众复盘过去一年市场的波折历程。
今年 1 月，整个软件行业在不到 30 天内蒸发了整整 1 万亿美元的市值。市场的担忧显而易见：人们普遍认为 AI 会让软件变得更加过剩、更容易被替代，其客户黏性或许远不如过去的估值模型所假设的那样稳固。然而 John 指出，这种恐慌纯粹是“对未来的前瞻性担忧”，而非当前正在发生的收入疲软。从 Stripe 平台上的实际 SaaS 支付交易量来看，目前的交易规模甚至远高于抛售潮爆发前的水平，SaaS 行业实际上依然保持着稳健的增长。
紧接着，资本市场展现出另一个强烈的风向转变：盈利能力重新成为最核心的考核指标。John 展示了历史上最具盈利能力与最缺乏盈利能力公司的市值分布图。在 2000 年互联网泡沫的高峰期，市场陷入某种非理性狂热，最不赚钱的公司市值一度短暂超越了最赚钱的公司；而反观当下，尽管外界对 AI 泡沫的讨论不绝于耳，二级市场的表现却显得“极其理性甚至有些乏味”——资本正毫不吝啬地将超高估值溢价赋予那些盈利能力最强、基本面最扎实的企业。
- 软件板块于 1 月份在 30 天内市值缩水达 1 万亿美元。
- Stripe 平台上的 SaaS 付款处理量不仅未见下滑，反而明显高于此轮抛售发生前的水平。
- 当前公开资本市场的估值逻辑高度向头部高盈利企业倾斜，与互联网泡沫时期的投机现象截然相反。
---
### [02:04](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=124s) 关税滞后效应与耐用品价格趋势的逆转
在过去一年里，贸易与关税政策始终是全球热议的焦点。John 打趣说，虽然距所谓的“解放日”（Liberation Day）关税政策出台仅仅过去了一年，但整个商业社会感觉像是熬过了漫长的八年。当前市场普遍的主流观点认为，新关税在宏观经济层面上几乎是个“雷声大雨点小的事物”（nothing burger），就像那只“没有叫吠的看门狗”——毕竟全年的整体跨境贸易流动仅比 2024 年略微下滑了一点。
但这种表面上的风平浪静掩盖了价格传导的真实轨迹。John 指出，在政策推行初期，许多企业选择依靠自身利润空间消化上涨的供应链成本；但随着时间推移，企业每个季度都在逐步将这些增加的成本转嫁给终端消费者。以耐用品价格走势为例，在 2024 年及更早的正常年份，由于制造业生产率的持续提高和全球贸易竞争的深化，液晶电视和洗衣机等耐用消费品呈现出平缓且可预测的逐年价格通缩曲线。
然而进入 2025 年后，这一长达多年的良性走势被彻底打破，价格曲线逆势大幅抬升。到了 2026 年的前三个月，这一上涨势头不但没有平息，反而在加速恶化。Stripe 的核心判断是：关税带来的实际成本压力仍在层层向终端消费者传导，“整个宏观故事的结局远未尘埃落定”。
---
### [03:16](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=196s) 破除“K型经济”迷思：Stripe 数据揭示的消费真相
近来广为流传的另一个宏观论调是所谓的“K型经济复苏”——即富裕阶层凭借高收入和资产增值支撑着整体消费，而普通大众的消费能力则在被逐渐侵蚀。社交网络上充斥着印证这一论调的典型案例：例如美联航全新的波音 787-9 客机座位图在 Twitter 上爆红，整个客舱的大部分物理空间全被改造成了商务舱；达美航空的最新财报也显示，其高端舱位（Premium Seat）的客票收入历史上首次超越了普通经济舱。
```mermaid
flowchart TD
accTitle: 社交网络叙事与真实底层消费数据对比
accDescr: 对比舆论场中的K型经济假象与Stripe实际交易数据反映出的高低收入群体消费差距收窄的客观事实。
subgraph Narrative["舆论与轶事（热门叙事）"]
A["高档舱位收入激增<br/>（达美航空高端舱超经济舱）"]
B["豪华空间扩张<br/>（美联航 787-9 大面积商务舱）"]
A --> C["推论：消费严重K型分化<br/>依赖高净值人群支撑"]
B --> C
end
subgraph Reality["Stripe 真实底层交易数据"]
D["追踪高收入与低收入群体消费比例"]
E["高低收入消费比率持续缓慢下降"]
D --> E
E --> F["结论：实际消费差距在收窄<br/>并未过度依赖超高净值人群"]
end
```
然而，John 强调，这些轶事虽然极具传播力、能轻松收获成千上万的转发，但与 Stripe 庞大的底层交易数据所展现的客观事实并不吻合。如果经济真的呈现出极端分化的 K 型结构，那么高收入与低收入消费者的支出比率曲线理应陡峭上升；但数据显示，该比率反而呈现出平缓的下降趋势。换言之，高收入群体与低收入群体之间的消费差距正在逐渐缩小，全社会并不存在舆论所渲染的“经济增长全赖富人一力支撑”的过度依赖现象。
- 达美航空高端客票收入虽首次超过经济舱，但更反映航空公司产品策略，而非全貌。
- Stripe 内部高收入与低收入消费者支出比率指标正在逐步走低，证明底层消费动力依然广泛存在。
---
### [04:21](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=261s) 劳动力市场放缓：AI 并非主因，多重宏观因素交织
针对外界普遍担忧“AI 正在抢走所有工作”的焦虑情绪，John 进行了客观拆解。从数据上看，美国劳动力市场确实在略微降温，过去三年间失业率上升了大约一个百分点。但如果仔细剖析背后的成因，AI 在其中所扮演的角色被过度放大了。
劳动力市场的放缓主要是多种传统宏观经济力量综合作用的滞后结果：
1. **疫情后招聘狂欢的消化周期**：企业在 2021 年和 2022 年进行了狂热的过度招聘，如今整个商业环境仍在持续消化并出清这一时期的过剩人力冗余。
2. **移民政策收紧**：跨境劳动力的流入放缓对劳动力结构和就业统计产生了直接影响。
3. **高利率与紧缩货币政策**：美联储针对 2026 年底的最新利率预测，比两年前所预期的水平整整高出了 50 个基点；资金成本的高企不可避免地导致了企业招聘步伐的迟缓。
John 总结道，AI 毫无疑问已经置身于这场变革之中，并且在未来一定会释放巨大的影响力，但在当下肉眼可见的宏观就业统计数据中，AI 绝对不是那个起主导作用的决定性推手。
---
### [05:36](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=336s) 经济活力的结构性爆发：AI 赋能下的“一人公司”浪潮
抛开上述外部舆论与传统指标，Stripe 目前处理的资金流已接近全球 GDP 的 2%，这使其成为了洞察全球数字经济最前沿的绝佳显微镜。John 指出，Stripe 数据呈现出的第一个最重大的结构性变化，便是经济活力的爆发式增长：新企业成立数量激增，它们在极长时间内保持着极小的人员规模，但其营收扩张速度却达到了前所未有的高度。为了从专业经济学视角解析这一现象，John 邀请了 Stripe 数据与 AI 业务负责人 Emily Glassberg Sands 博士登台。
Emily 解释说，很多笃信 AI 颠覆力量的人常感到认知失调，因为在官方公布的传统劳工和宏观指标里很难找到 AI 的影子。然而，AI 实际上早已深刻融入宏观数据，关键在于“你是否找对了观察的窗口”。回顾美国过去 20 年的企业注册数据，除了疫情期间众所周知的一波激增外，当前全美创业浪潮正在经历更强劲的“二次加速”。
这波新增动力几乎全部来自于美国人口普查局定义的“无雇员企业”（non-employer firms），亦即大众熟知的“单人创业者”（solopreneur，一人公司）。过去，这类单打独斗的小生意往往被视作打闹性质的兼职或微型营生，但 AI 工具的赋能彻底改写了这一逻辑，赋予了一人公司前所未有的商业规模化能力：全美年营收突破 10 万美元的一人公司数量迎来井喷，如今已有近 500 万美国人以此谋生。
这一创业活力的大爆发绝不仅限于美国：
- 澳大利亚的新注册企业数量增长了 40%。
- 芬兰的新注册企业数量上升了 70%。
- 法国的新增企业注册量激增了 80%。
- 各大发达经济体均在同步迎来一场前所未有的草根商业活力浪潮。
---
### [07:43](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=463s) 新一代初创企业范式：首日全球化与指数级营收增长
随后，John 与 Emily 共同分享了来自 Stripe 自身产品生态的一线佐证。作为全美最简便的公司注册服务，Stripe Atlas 在上周正式迎来了第 100,000 家注册企业。这项里程碑的达成时间远超团队早前的预期。戏剧性的是，这家第 10 万个里程碑公司正是 Empirical——一家利用 AI 软件优化电池储能系统（BESS）盈利能力的初创企业，其创始人 Rachana 也受邀来到了现场。
不仅注册数量出现爆炸式增长，新企业的造血能力更令人咋舌。整体而言，2025 届通过 Atlas 成立的公司，在同等生命周期节点所赚取的营收达到了 2024 届公司的两倍；而更惊人的是 2026 届新生代——尽管今年才刚刚过去几个月，2026 届初创企业的同期营收轨迹已达到 2025 届的整整五倍之多。
```mermaid
flowchart LR
accTitle: 新旧两代初创企业国际化路径演进
accDescr: 展示从过去先做本土再出海的传统模式，转变为主打首日全球化、极简人头与自动化扩张的新型AI初创范式。
subgraph Legacy["传统初创路径"]
L1["深耕本土市场"] --> L2["团队人员大幅扩张"]
L2 --> L3["成熟后缓慢出海试水"]
end
subgraph Modern["新一代 AI 初创范式"]
M1["Day 1 设立全球化架构"] --> M2["全流程高强度 AI 自动化"]
M2 --> M3["极精简团队运转<br/>（小团队甚至一人公司）"]
M3 --> M4["迅猛穿透长尾全球市场<br/>（首年覆盖数十国）"]
end
```
支撑这种超常规扩张曲线的核心引擎，在于初创企业全球化模式发生的代际跃迁。五年前，在 Stripe 平台上仅有 11.6% 的公司能够实现“跨国收入占据其总营收的绝大部分”；而如今，这一比例直接翻倍。更重要的是，他们不再局限于美、英、日等主流大国，在跨境收入为主的企业中，有大约四分之一的收入来自全球排名前十以外的长尾小众市场。
传统的商业扩张顺序被彻底颠覆：过去是初创公司先做透本土市场，羽翼丰满后再涉足海外；现在的“Stripe 新生儿”则是从成立第一天起就在全球长尾市场中狂奔。在 Stripe 平台排名前 100 的 AI 初创企业中，中位数企业在成立第一年内就向 55 个国家销售产品，且大部分收入均来自海外。典型的代表如 2024 年在美国创立的 Emergent Labs，其近 70% 的营收直接源自海外市场，并在多达 16 个国家中每个单一市场贡献度均超过总收入的 1%。
这一全新的创业范式可以归结为三大特征：“上线首日即全球化、极度精简的人头编制、以及全方位的激进自动化”。演讲最后，Emily 幽默地调侃道，尽管现场台上站着一位经济学博士（Emily 本人）和一位大学辍学后在酒吧录播客的创业者（John），但接下来却要由 John 来带大家回顾一堂生动的经济史课。John 随即风趣回应：“Dr. Sands，如果你知道历史上有多少经济学经典理论其实是在酒吧里就着啤酒推导出来的，你一定会大吃一惊。”
### [10:52](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=652s) 科斯定理视角下的企业与 AI：协调成本的重构
演讲者以一段经济思想史作为开场：1931 年，一位 21 岁的英国年轻学生在大萧条时期的美国进行公路旅行。彼时正值中央计划经济在知识界享有崇高声望的年代，许多年轻知识分子笃信自上而下的集中规划远优于杂乱无章的自由市场。然而这位年轻学生观察到了一个有趣的现象：在一家企业内部，资源的调配绝非依赖价格机制，企业内部并没有设立小型自由市场，其本质就是高度自上而下的中央计划。他因此提出了一个发人深省的问题：**既然市场机制如此高效，为什么人们还要在市场的大海中不断建立一座座中央计划的小岛？**
这位学生正是后来的诺贝尔经济学奖得主罗纳德·科斯（Ronald Coase）。他的核心答案是：企业的存在，是因为在企业内部协调各项业务的交易与沟通成本，往往显著低于在公开市场上通过合约与交易进行协调的成本——瑞典皇家科学院在整整 60 年后才为此授予他诺贝尔奖，足见其决策之审慎。
那么，如何用科斯的理论来理解 AI 带来的变革？演讲者指出，短期来看，**企业内部效应（within-firm effect）**最为立竿见影：企业内部拥有共享的业务上下文（context）、统一的记录系统（systems of record）以及一致的激励机制，这让 AI 能够极其轻松地赋能内部协作，也是当下绝大部分企业最直观的体会。然而从**中期来看，外部市场协调效率将迎来爆发式提升**：
- AI Agent 具备极强的外部信息发现能力（discovery）；
- 极大地降低了接入和集成新软件的门槛；
- 让跨企业签署合同与协议变得异常清晰直接；
- Agent 能够以极高频次、极小金额完成交易，其敏捷程度远超人类。
这三者合力将大幅削减企业间的外部协调成本。最终的经济图景将是：**每家企业的员工人数更少、单店产出极高、市场上涌现出更多精简敏捷的企业，而绝大部分协调活动将重回市场化机制**。科斯若在世定会对此大为赞赏，演讲者风趣地将其形容为“来自另一个时代的独立黑客（indie hacker）”。
---
### [13:30](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=810s) 商业的 Agent 化：从表单代填到自主决策的五级跃迁
伴随着企业活力的激增，演讲者切入核心主题：**商业交易本身正在全面 Agent 化（Agentic Commerce）**。Stripe 将购买流程中的自动化演进划分为从简单辅助到全面自主的五个不同层级，并展示了 2026 年所处的现实水位：
- **Level 1（表单代填与代结账）**：软件代表用户机械地填写繁琐表单并完成结账。例如 Stripe 与 Meta 合作的 In-App 结账体验——用户在社交媒体中看到心仪广告并表达购买意向，Agent 读取保存好的个人信息并瞬间代办结账。虽然体验极为丝滑，但在技术上尚未展现深层推理。
- **Level 2（基于约束的智能导购）**：从传统的关键词搜索（Keyword search）跨越到能够基于多重约束进行推理的 AI 助手。演讲者幽默地举了一个“完全假设”的例子：“我想给 38 岁的哥哥挑个生日礼物，他在这个‘高龄’莫名其妙迷上了街头健身（calisthenics），但相关装备已经一应俱全，请推荐一款 100 美元以内、非同寻常的礼物。” 传统的关键词搜索对此无能为力，而 ChatGPT 或家具零售商 Wayfair（支持用户仅描述房间风格、氛围即可在目录中翻找匹配）等系统已能自如处理。在此阶段，用户依然保留最终决策权。
- **Level 3 至 Level 5（具备真正自主权的购买决策与执行）**：Agent 将能够完全自主地制定决策并执行购买。
前沿领域的强劲需求已经显现。以开源生态 OpenClaw 为例，尽管对于普通公众而言仍有较高门槛，但其平台上的支付相关技能（payment-related skills）在短短 12 周内就斩获了大约 12.5 万次累计下载量。这证明自主商业的核心不再是“是否存在需求”，而是“如何将最前沿的实验推向主流商业舞台”。这也印证了前一天 Stripe 演讲中强调的观点：AI 经济基础设施必须补齐三大支柱——**Agent 具备付款能力**、**商家具备接收 Agent 付款的能力**，以及贯穿其间的**信任层（trust layer）**。
```mermaid
flowchart TD
accTitle: 商业交易 Agent 化的自主等级演进
accDescr: 展示从 Level 1 辅助到 Level 5 完全自主交易的演进路径，当前正处于软件向软件采购的突破期。
subgraph Level_1_2["当前成熟落地 (2026)"]
L1["Level 1: 辅助代填<br/>(Meta In-App 结账 / 自动填表)"]
L2["Level 2: 约束推理导购<br/>(ChatGPT / Wayfair 自然语言搜品)"]
end
subgraph Frontier["前沿爆发：软件向软件采购"]
L3["Level 3-5: 自主商业决策与执行<br/>(OpenClaw 技能 / 跨 Agent 自动化交易)"]
end
L1 --> L2
L2 --> L3
```
---
### [16:21](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=981s) 软件购买软件：实时演示 AI Agent 的微支付与调研
在所有 Agent 落地场景中，推进速度最快的当属“软件向软件自主采购”（Software buying from software）。演讲者现场通过语音向 Claude 下达了一项复杂宏观调研指令：
> “Hey Claude，AI 的算力需求是如何影响大宗商品价格，以及不同能源来源的供需关系的？”
演讲者向观众剖析了该问题的现实背景：AI 算力大扩张带动了空前的资本开支（capex），新建数据中心四处林立，并受制于电力供应瓶颈；而美国电网长期未见大规模扩建，面对突如其来的巨量电力需求，原有的供需均衡被彻底打破。
Claude 接单后开始自主检索多方信源，随后在终端界面明确反馈：“目前的市场格局非常惊人，但我需要获取宏观商品与股票数据来支撑深入分析。Alpha Vantage 拥有我所需的数据集，我打算购买该服务，总费用为 4 美分（$0.04），这完全在我们的授权预算额度之内。”
```mermaid
flowchart LR
accTitle: Agent 自主数据采购与报告生成流程
accDescr: 展示 Claude Agent 接收指令、自主寻找付费数据、发起微支付、生成报告并准备分销的端到端链路。
subgraph Step1["调研与预算评估"]
A["用户语音提问<br/>(AI 对能源与商品影响)"] --> B["Claude 解析需求<br/>寻找外部数据源"]
B --> C["锁定 Alpha Vantage 数据<br/>报价 4 美分 (在预算内)"]
end
subgraph Step2["执行微支付与分析"]
C --> D["Tempo CLI 发起支付<br/>(稳定币钱包 / 零手续费)"]
D --> E["拉取数据并自主撰写研报"]
end
subgraph Step3["二次变现"]
E --> F["Claude 接收指令<br/>自主定价并发布上架"]
end
```
在底层支付机制上，不同于前一天演示中由人工使用 Link CLI 向 API 审查员付款，本次 Claude 是通过底层调用的 **Tempo CLI** 借助其专属**稳定币钱包**直接划扣。演讲者特别说明：虽然机器间交易亦可使用法定货币，但面对 4 美分这种极其细微的**微支付（micropayments）**，传统的法币支付通道在手续费上不可行；具备接近零手续费特性的稳定币，才真正使机器间的微交易成为现实。
---
### [19:00](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1140s) 算力引发的能源激增，以及 Agent 的研报自主分销
在等待 Claude 运行推理的间隙，演讲者打趣现场如果能有个吉他手即兴弹奏暖场就更好了。随后，他向观众展示了 Stripe 旗下探索人类进化的深度读物《Works in Progress》杂志中的趣味内容：既有英国核电发展史上的起伏反思，也有被戏称为“AI 首个真正实用落地案例”的 **The Guinness Tank**——一个 AI Agent 通过自动拨打爱尔兰所有酒吧的电话，以此实现全爱尔兰健力士黑啤（Guinness）价格的实时物价追踪看板。
随后，Claude 顺利完成运算，浏览器自动弹出了排版精美的深度研报。报告揭示了一组惊人数据：
- 到 2080 年，美国数据中心的用电需求预计将激增至当下的近三倍；
- 超大规模云厂商（Hyperscalers）的资本支出激增了 62%；
- 作为主要发电过渡能源的天然气，其价格在短周期内已大幅飙升 104%。
面对这份详尽报告，演讲者并没有逐页细读，而是紧接着演示了闭环商业的下一阶段——让 Agent 将产出的成果向市场变现。他通过语音向 Claude 下达追加指令：
> “Hey Claude，把这份报告发布上架出售。请你自行评估并制定一个合理价格，供其他 Agent 或人类用户搜索并购买。”
Agent 随即启动了自动建站与分发部署流程。不过演讲者此时突然幽默地自言自语道：在公开叫卖之前，自己最好先仔细核查一下刚刚花 4 美分买来的 Alpha Vantage 数据集的商业授权许可条款（licensing terms），看看自己是否真正拥有将包含该数据的研报用于商业再分销（commercially redistribute）的合法权利。
### [21:16](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1276s) 智能体商务实操部署与先发优势
在演示现场等待报告完成发布的间隙，演讲者强调了智能体商务（Agentic Commerce）所蕴含的巨大机遇。当前智能体商务已经不是遥不可及的概念，而是切实行进在商业现实之中。率先入局的先行者优势（First-mover advantage）极为可观，这也是 Stripe 竭尽全力快速推进该领域基础设施建设的核心原因：如果一家公司的产品或平台具备承载机器对机器（M2M）支付的潜力，现在就应当立即着手构建。在过去的开发范式中，开发者必须繁琐地手动调取 API Key，或者频繁在 Vercel 等平台的管理后台中来回切换配置；而如今依托于 Stripe Projects，整个编排流程已被完全自动化接管。现场展示中还列举了数家已经接入智能体商务生态的创新公司，例如提供浏览器级智能体网页浏览能力的 Parallel，以及支持智能体代寄实体信件的 Postal Form，实现了前沿机器交互与传统实体世界的无缝对接。
```mermaid
flowchart LR
accTitle: 智能体商务交互与履约架构
accDescr: 展示用户智能体如何通过 llms.txt 获取指令并完成全自动机器对机器支付。
subgraph Discovery["协议发现与交互"]
A["智能体 Agent"] -->|"读取规范"| B["llms.txt 指令文件"]
end
subgraph Execution["支付与编排"]
B -->|"单次调用请求"| C["Stripe Projects 编排"]
C -->|"触发交易"| D["机器对机器支付结算"]
end
subgraph Fulfillment["终端交付"]
D -->|"自动化交付"| E["在线报告 / 实体服务"]
end
```
随着后台流程走完，部署在 `johnsreport.vercel.app` 的网页正式上线。页面上不仅为人类用户提供了直接点击购买报告并触发确认链接的传统结账路径，更关键的突破在于网站根目录下的 `llms.txt`。该文件结构化地提供了全套机器可读指令，使得各类外部 AI 智能体仅凭单次请求便能直接调用支付逻辑、自动化采购产品。演讲者在台上风趣地向全场听众呼吁：“我非常欢迎、甚至恳求大家自掏腰包花上 5 美元的‘巨资’买一份我的报告，好帮我补贴一下 Token 预算。”随着演示成功，屏幕上飘落起庆祝彩带，演讲者打趣道：“快给他们放点彩带，现在的年轻一代（Gen Z）就吃这套”，生动宣告了机器与机器之间自主进行商务往来的时代已经揭开序幕。
- **跨世代兼容工具**：除了数字服务，类似 Postal Form 的工具允许自主智能体代用户邮寄纸质信件，体现了新旧交互方式的协同。
- **机器可读入口**：通过配置标准的 `llms.txt`，网站将购买协议直接暴露给 LLM，彻底去除了传统网页 UI 对机器人的阻碍。
- **现场定价与结算**：报告标价 5 美元，直接打通真实生产环境的端到端支付与交付链路。
---
### [23:46](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1426s) 廉价智能时代的经济学：AI 时代的互补品理论
由智能体商务的成熟，演讲者引出了当天的第三个核心论题：当智能无处不在、任何人都唾手可得且成本急剧下降时，世界上什么东西反而会变得更值钱？这触及了古典经济学中的一个基本定律——“当某种要素变得极其廉价时，它的互补品（Complements）价值就会大幅飙升”。这种联合需求曲线（Joint demand curves）的现象在商业史上屡见不鲜：当集装箱海运技术的普及使得全球货运成本雪崩式下跌时，能够停靠接卸大型集装箱巨轮的深水码头港口资源价值瞬间暴涨；在 20 世纪 90 年代初，最初的无线电频谱拍卖仅能募得数亿美元，但随后移动电话迅速普及并彻底平民化，极具互补性质的频段资产身价陡增数个数量级，最终引发了多国政府动辄数百亿美元的频谱竞拍狂潮。
面对当前的大模型与生成式 AI 浪潮，各行各业的决策者都必须从根本上重新审视：什么是智能的互补品？伴随机器推理成本趋近于零，究竟哪些资产能够坐享价值倍增的红利？
---
### [24:46](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1486s) 算力、能源与专有数据：AI 催生的前三大价值高地
经济学规律在物理层面的显现最为直观，首先反映在芯片与能源这两项显而易见的互补要素上。GPU 在人工智能爆发前已然是一门不错的生意，但英伟达（NVIDIA）近年来的出货量和市值飙升清晰地印证了计算硬件作为智能互补品的价值飞跃——其营收结构从早期主要依赖游戏显卡，彻底转变为数据中心计算与网络通信业务的爆发式增长。能源领域亦然，电力唯有接入能够承载智能的系统时才具备最高价值；当前全球核能产业正经历历史性复兴，其主要推动力正是为了给巨型 AI 数据中心提供不间断基荷电力。而在核能扩建落地前的过渡阶段，全球对重型燃气轮机的订购需求呈爆炸式增长，制造龙头西门子能源（Siemens Energy）等公司的市值攀升便是最佳明证。
```mermaid
flowchart TD
accTitle: 廉价智能对五大互补资产的价值催化机制
accDescr: 展示由于推理智能成本雪崩，带动算力、能源、专有数据、网络效应及物理执行五大要素价值上升。
Core["智能成本急剧下降"]
subgraph Physical["硬件与物理基础设施"]
Core --> Com1["算力硬件：NVIDIA 计算与网络架构"]
Core --> Com2["基础能源：核电复兴与燃气轮机"]
Core --> Com5["物理世界执行力：重工业设备与落地网络"]
end
subgraph Informational["数据与网络架构"]
Core --> Com3["专有数据资产：高防爬壁垒与非广告变现"]
Core --> Com4["双边网络效应：撮合交易平台抽成率提升"]
end
```
相比硬件和能源，较少被充分讨论但更为关键的第三大互补品是**专有数据（Proprietary Data）**。当高智能代理拥有深度逻辑推理能力时，企业私域的高质量数据资产便被彻底激活。衡量数据升值的一个直观信号是：曾经免费公开数据以供抓取的互联网平台，如今纷纷构筑高墙禁止 AI 爬虫随意采集，转而寻求商业化变现。以 Reddit 为例，其多年积累的各子板块（Subreddits）用户讨论以往只是资产负债表上的沉睡资产，但如今通过对外授权数据协议，单季度非广告收入已达到 3,500 万美元。Stripe 自身的反欺诈防护产品 Stripe Radar 也是典型写照——虽然该系统历来能够在全量交易网络语料上进行分析，但随着底层 AI 推理模型的迭代演进，所沉淀的历史交易与风控专有数据的商业防护价值得到了成倍放大。
- **英伟达产品重心迁移**：业务核心彻底脱离消费级游戏图形渲染，全面由数据中心级大规模互连与计算网络驱动。
- **能源基础设施承压与转移**：数据中心能耗迫使能源网重构，短期利好燃气轮机制造商，长期推动核电投资复兴。
- **Reddit 商业模式质变**：停止提供免费语料抓取，单季度通过 AI 训练数据许可授权斩获 3,500 万美元纯数据收益。
- **Stripe Radar 效益升级**：底层大模型推理能力的进化，使系统对海量底层专有交易数据的提取与欺诈拦截精度显著提高。
---
### [26:31](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1591s) 网络效应与物理世界壁垒：第四与第五大互补品
智能泛滥时代所催生的第四大互补品是**网络效应（Network Effects）**。无论底层算法如何进化，买家与卖家始终需要达成交易的撮合场域；当软件构建门槛降低、用户在工具间的切换成本被抹平时，平台型网络效应的锁客价值不降反升。对全球排名前十的公开头部平台型企业（Marketplaces）的抽成率（Take rate）进行追踪可以发现：平台抽成率在经历数年横盘后，先是在疫情期间出现跳涨，随后在过去三年间伴随各大平台全面应用更先进的 AI 撮合与风控推荐算法，整体佣金变现能力持续稳步攀升。
第五项、也是最具持久护城河特征的互补品，是**打通复杂软件系统与物理现实世界执行能力的综合实体运营资产**。演讲者以百年农机巨头约翰迪尔（John Deere）为例：如果询问大众谁是 AI 时代的核心受益者，约翰迪尔几乎很少上榜；然而，这家公司多年来持续将高精度 GPS 制导、机器视觉识别以及全套高密度传感器阵列无缝集成至其大型农机装备中。在这个商业闭环里，纯粹的 AI 模型本身并非绝对壁垒——竞争对手或许也能组建团队写出合格算法，真正的护城河在于“把搭载智能系统的拖拉机开进全球 130 多个国家的农田土壤之中”。
| 互补品维度 | 核心价值体现与壁垒来源 | 典型代表案例 |
|:--- |:--- |:--- |
| **算力与硬件** | 极高工艺与互连网络架构门槛，承载一切模型推理与训练 | NVIDIA 算力与通信网络 |
| **基础能源** | 支撑大型数据中心运行的稳定基荷电力保障 | 核电投资再起、西门子能源燃气轮机 |
| **专有数据** | 具有极高法律与业务访问壁垒的独家结构化语料 | Reddit 数据许可、Stripe Radar 支付数据 |
| **网络效应** | 买卖双边高黏性聚合生态，AI 提升撮合效率反哺抽成比例 | 头部互联网 Marketplaces（抽成率近 3 年连涨） |
| **物理世界执行力** | 长期硬件积累、跨国线下分销、与复杂现实场景软硬深度耦合 | 约翰迪尔（遍布 130 国田间地头的智能拖拉机） |
这一格局彻底颠覆了科技行业的传统竞争逻辑：过往各团队引以为傲的是“打造所属赛道中最顶尖的软件”，但当下单纯的软件代码本身已不再是决胜核心。真正能够在高度智能的丰沛世界里立于不败之地的，是那些难以复制的专有数据、强黏性网络效应，以及“花费长达十年方才调优成熟的物理现实运营系统与综合工具链”。
---
### [28:12](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1692s) 索洛悖论与电气化启示：AI 重构经济范式的漫长与加速
在演讲的尾声，演讲者借由一段厚重的技术经济史来审视当下：1882 年，托马斯·爱迪生（Thomas Edison）使用六台发电机点亮了曼哈顿下城 82 位商户的电灯，人类正式迎来电气时代。然而在其后的数十年中，尽管电力技术不断渗透，宏观经济的劳动生产率增速却异乎寻常地平缓，甚至随着铁路投资狂潮的退去而出现放缓趋势。当时的经济学家普遍深感困惑——这项技术如此颠覆，为何在经济数据报表中毫无显现？其症结并非在于电气技术本身存在缺陷，而是在于“整个工业经济体系必须花时间完成对新技术的消化与吸收”。当时的既有工厂完全是围绕蒸汽机动力构建的，垂直排布的传动轴、纵横交错的皮带网络与多层建筑空间，从根本上违背了电力驱动最优的横向流水线逻辑。直到工厂被推倒重来、基于电机动力学从零重新设计车间结构，迟到许久的生产率飞跃才终于喷薄而出。
```mermaid
flowchart TD
accTitle: 技术红利释放的历史延迟与重构模型
accDescr: 解释新技术诞生到生产率飞跃之间必经的漫长组织消化与流程重构期。
subgraph Historical["历史范式：电气化转型（耗时约 30 年）"]
E1["1882 年：爱迪生点亮曼哈顿"] --> E2["1880s-1900s：蒸汽时代旧工厂无法适配电机"]
E2 --> E3["1910 年代：工厂推倒重来、彻底采用横向流水线"]
E3 --> E4["生产率红利爆发：人均产出增长率翻倍"]
end
subgraph Current["当代范式：AI 重构经济架构"]
A1["基础模型能力全面涌现"] --> A2["消化吸收期：单兵企业崛起、机器自动商务"]
A2 --> A3["全社会组织架构再平台化（Re-platforming）"]
A3 --> A4["宏观经济飞跃显现（耗时将远短于 30 年）"]
end
```
这种系统性重构整整经历了 30 年之久——从 1882 年一直跨越到 1910 年代末期；唯有在 1910 年代那短短十年中，工人的平均单位产出增长率才实现翻倍。同样的规律在信息革命初期再度重演，诺贝尔经济学奖得主罗伯特·索洛（Robert Solow）在 1987 年留下了著名的论断：“计算机无处不在，唯独没有出现在生产率统计数据中”（Solow Paradox），这一停滞表象直至 1990 年代中期才随着企业信息化改造完成而终被打破。任何重塑底层的革命性技术，在相当长的时间里看起来都像是毫无动静；经济学家们盯着宏观指针，宛如困惑地拍打着仪表的乘客在喃喃自语：“这玩意儿到底开了没有（Is this thing on?）”。
这恰恰是当前全球在面对生成式 AI 时所处的时间节点。重构的种子已经深埋于当前的微观实践中：严肃商业实体生存所需的最小有效规模（Minimum efficient size）正在极速坍塌；“超级个体”凭借单人力量便能将营收规模拓展至七位数乃至更高量级；软件代理与代理之间实现全自主商业往来；初创企业自创立第一天起便实现全球化展业。由于所有这些新型运作形态均无法嵌套进过去的宏观经济统计模型中，明天的经济报表或许依然无法即刻兑现亮眼的生产率红利，但这正是整个经济体系正在全面进行底层架构重塑（Re-platforming）的前置信号。
演讲者在结束致辞中总结道：各位开发者当下所构建的企业形态，绝非人工智能宏大历史中的边缘注脚，其本身就是 AI 推动经济与生产率转型的核心历史篇章。电气化曾经耗费了人类社会整整 30 年时间来完成经济范式的重组，但面向人工智能的重塑进程，“我们绝不需要等待那么久”。
## Notable quotes
> [10:52](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=652s) 既然市场机制如此高效，为什么人们还要在市场的大海中不断建立一座座中央计划的小岛？ — John Collison
> [07:43](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=463s) Dr. Sands，如果你知道历史上有多少经济学经典理论其实是在酒吧里就着啤酒推导出来的，你一定会大吃一惊。 — John Collison
> [16:21](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=981s) Hey Claude，AI 的算力需求是如何影响大宗商品价格，以及不同能源来源的供需关系的？ — John Collison
> [16:21](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=981s) 目前的市场格局非常惊人，但我需要获取宏观商品与股票数据来支撑深入分析。Alpha Vantage 拥有我所需的数据集，我打算购买该服务，总费用为 4 美分（$0.04），这完全在我们的授权预算额度之内。 — Claude
> [19:00](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1140s) Hey Claude，把这份报告发布上架出售。请你自行评估并制定一个合理价格，供其他 Agent 或人类用户搜索并购买。 — John Collison
> [21:16](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1276s) 我非常欢迎、甚至恳求大家自掏腰包花上 5 美元的‘巨资’买一份我的报告，好帮我补贴一下 Token 预算。 — John Collison
> [21:16](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1276s) 快给他们放点彩带，现在的年轻一代（Gen Z）就吃这套 — John Collison
> [23:46](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1426s) 当某种要素变得极其廉价时，它的互补品（Complements）价值就会大幅飙升 — John Collison
> [26:31](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1591s) 把搭载智能系统的拖拉机开进全球 130 多个国家的农田土壤之中 — John Collison
> [28:12](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1692s) 计算机无处不在，唯独没有出现在生产率统计数据中 — Robert Solow
> [28:12](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1692s) 这玩意儿到底开了没有（Is this thing on?） — John Collison
> [28:12](https://www.youtube.com/watch?v=5wGqWRv1Z1s&t=1692s) 我们绝不需要等待那么久 — John Collison
## Mentioned
John Collison, Will Gaybrick, Emily Glassberg Sands, Ronald Coase, Thomas Edison, Robert Solow, Rachana, Stripe, Stripe Atlas, Stripe Radar, Stripe Projects, Link CLI, Tempo CLI, Claude, ChatGPT, Meta, Wayfair, OpenClaw, Alpha Vantage, Parallel, Postal Form, Vercel, NVIDIA, Siemens Energy, Reddit, John Deere, United Airlines, Delta Air Lines, Emergent Labs, Empirical, Works in Progress, The Guinness Tank, llms.txt
## Video description
Hear about trends that will shape the next decade of payments and global trade.
More information: https://stripesessions.com
Register for Stripe Sessions 2027: https://register.stripesessions.com/2027?promocode=s27youtube
Opening keynote: https://youtu.be/lIsHZfRl2zw
Developer keynote: https://youtube.com/live/-vRY2dtD7iQ
Indexing the economy keynote: https://youtu.be/5wGqWRv1Z1s
Sam Altman fireside: https://youtu.be/5eouRdDYM2c
Nat Friedman and Daniel Gross fireside: https://youtu.be/l9wzs_QIyp0
[Source](https://www.youtube.com/watch?v=5wGqWRv1Z1s)
Source quotes
[10:52] John Collison: 既然市场机制如此高效，为什么人们还要在市场的大海中不断建立一座座中央计划的小岛？
[7:43] John Collison: Dr. Sands，如果你知道历史上有多少经济学经典理论其实是在酒吧里就着啤酒推导出来的，你一定会大吃一惊。
[16:21] John Collison: Hey Claude，AI 的算力需求是如何影响大宗商品价格，以及不同能源来源的供需关系的？
[16:21] Claude: 目前的市场格局非常惊人，但我需要获取宏观商品与股票数据来支撑深入分析。Alpha Vantage 拥有我所需的数据集，我打算购买该服务，总费用为 4 美分（$0.04），这完全在我们的授权预算额度之内。
[19:00] John Collison: Hey Claude，把这份报告发布上架出售。请你自行评估并制定一个合理价格，供其他 Agent 或人类用户搜索并购买。
[21:16] John Collison: 我非常欢迎、甚至恳求大家自掏腰包花上 5 美元的‘巨资’买一份我的报告，好帮我补贴一下 Token 预算。
[21:16] John Collison: 快给他们放点彩带，现在的年轻一代（Gen Z）就吃这套
[23:46] John Collison: 当某种要素变得极其廉价时，它的互补品（Complements）价值就会大幅飙升
[26:31] John Collison: 把搭载智能系统的拖拉机开进全球 130 多个国家的农田土壤之中
[28:12] Robert Solow: 计算机无处不在，唯独没有出现在生产率统计数据中
[28:12] John Collison: 这玩意儿到底开了没有（Is this thing on?）
[28:12] John Collison: 我们绝不需要等待那么久
Source quotes
[8:57] Emily Glassberg: This is a totally new startup playbook. You launch globally on day one, you keep headcount very lean, and you automate aggressively.
[10:46] Emily Glassberg: Coase would have liked that idea. He always thought firms were an inefficient setup. The more we can do via markets, the better.
[15:30] John Collison: Hey Claude, how is AI demand affecting commodity prices and supply and demand for different energy sources?
[28:12] John Collison: Transformative technology looks for a long time like it's not doing much. If you're looking at the economic gauges, you're going to sit there and you're tapping the gauges, is this thing on?
Original transcript
[0:00] Good morning. It is great to be back at
[0:03] Moscone.
[0:04] You can almost feel the tectonic plates
[0:08] shifting underneath us. Yesterday, Will
[0:10] walked you through the 288 new products
[0:13] that we're launching to help you as the
[0:15] economy re-platforms around the eye.
[0:19] But today, I want to take you to my
[0:21] happy place, getting deep into the
[0:23] economic data.
[0:25] I have a few big trends that we're
[0:26] seeing that I want to talk you through.
[0:30] But first,
[0:31] let's take stock of what's happened
[0:33] since we were together here at Sessions
[0:35] last year.
[0:37] So,
[0:39] since we last got together,
[0:41] it's been a little wobbly.
[0:43] In January, the software sector lost a
[0:45] trillion dollars of market value in
[0:48] under 30 days.
[0:49] You're all familiar with the worry. AI
[0:51] makes software more abundant, more
[0:53] substitutable, potentially less sticky
[0:55] than prior models had assumed.
[0:58] But interestingly, this is a prospective
[1:00] worry. It's not revenue softness that's
[1:03] happening right now.
[1:05] SaaS payment volumes on Stripe, they're
[1:07] actually a good chunk higher today than
[1:09] before the sell-off. So, SaaS is still
[1:11] growing just fine.
[1:15] What does seem to have come in back into
[1:16] fashion though is profitability, and not
[1:18] just in software, but across the entire
[1:21] equity market. So, this is a chart
[1:23] showing how much market cap has
[1:25] historically accrued to the most
[1:28] profitable and the least profitable
[1:29] companies in the index. So, you see in
[1:31] the middle here, this is the dot-com
[1:32] boom where the markets went a bit mad
[1:34] for a moment, and briefly the least
[1:36] profitable companies were worth more
[1:38] than the most profitable ones.
[1:40] But if you go forward to today, now,
[1:43] despite all the talk about bubbles, it's
[1:44] the opposite. The markets are being
[1:46] tediously rational. The most profitable
[1:48] companies are the ones being rewarded
[1:50] with outsize valuations.
[1:55] We heard a lot about trade policy over
[1:57] the past year. Believe it or not, it's
[1:58] only been a year since Liberation Day.
[2:00] It feels like eight years, but one year.
[2:04] The common wisdom by now is that the
[2:07] tariffs were kind of, economically
[2:08] speaking, a nothing burger, you know,
[2:10] the dog that didn't bark.
[2:12] And you can make that case. Trade flows
[2:14] ended the year only a little bit below
[2:16] 2024 levels.
[2:19] But there were higher prices to absorb
[2:21] as a result of tariffs.
[2:23] What happened is early on, you had
[2:25] businesses using the higher costs.
[2:28] But now, quarter by quarter, they're
[2:30] starting to pass them through. So, we
[2:32] can throw up durable goods prices here.
[2:35] And this is how they behave intra-year
[2:37] in normal years. You might be wondering,
[2:38] you know, why is there there this
[2:39] downward trend? It's cuz you have
[2:41] deflation in durable goods.
[2:42] Manufacturing productivity generally
[2:44] increases, global trade adds
[2:46] competition, and your flat screen TV
[2:47] gets cheaper, or your washing machine
[2:49] gets better for the same price. So, you
[2:50] usually, in most years, get this gentle,
[2:53] predictable slope downwards.
[2:55] And you see this only goes up to 24. If
[2:57] we add 25,
[2:59] totally different. Bucks the trend.
[3:01] And if we add 26, it's only 3 months in,
[3:04] but same thing is happening and even
[3:06] more so.
[3:08] So, our view is that tariff costs are
[3:10] still working their way through to the
[3:12] consumer. The story is not written yet.
[3:16] Another phenomenon you might have heard
[3:18] about is this K-shaped economy. The idea
[3:20] that wealthy consumers are holding
[3:22] things up. They're an increasingly large
[3:24] share of the spending. So, maybe you saw
[3:26] this go viral on Twitter. This is the
[3:28] new United 787-9 seat map. And you can
[3:32] see that most of the cabin is now
[3:33] dedicated to business class seats.
[3:38] Or you can look at at Delta's ticket
[3:40] revenue. So, they reported that premium
[3:42] seat revenue is now bigger than economy
[3:45] for the first time.
[3:48] This stuff
[3:50] gets a lot of retweets, but we don't
[3:52] actually see it in Stripe's data. So, if
[3:55] we look at the ratio of high-income to
[3:59] low-income consumer spending in our
[4:01] data, you know, in a K-shaped economy,
[4:04] you'd expect this chart to be going up.
[4:05] But, it's the opposite. The line is
[4:06] trending gradually downwards. The gap
[4:08] between high-income and low-income
[4:10] spending has been shrinking.
[4:12] So, we don't actually see, when we look
[4:14] for it, this over-reliance of the
[4:16] economy on high-end consumers that
[4:18] everyone has been talking about.
[4:21] The other major economic topic is the
[4:23] worry that
[4:25] AI is taking all the jobs.
[4:27] And if you look at the numbers, it's
[4:28] true that the labor market is cooling
[4:30] slightly. US unemployment rates are
[4:32] ticking up about a percentage point over
[4:34] the past 3 years.
[4:37] But, you can ask the question, how much
[4:38] of that is actually AI?
[4:40] Part of it is the delayed pandemic
[4:42] hangover. If you look at the hiring
[4:43] rate,
[4:44] companies really binged on hiring in '21
[4:46] and '22, and they're still unwinding
[4:48] that. And tighter immigration played a
[4:50] bit of a role. And then, of course, you
[4:52] have interest rates. So, the Fed's
[4:54] latest forecast for the end of 2026 is
[4:57] half a point higher than they predicted
[4:59] at this time 2 years ago. So, money is
[5:02] tighter, and tight money slows hiring.
[5:05] So, AI is in the mix. It's presumably
[5:08] going to have a big future impact,
[5:10] but it's one force amongst many, and
[5:13] probably not the dominant one in the
[5:15] labor market numbers you've seen
[5:16] today's.
[5:18] Now, you're probably wondering what
[5:19] we're seeing at Stripe, and how it
[5:21] corresponds to what's going on in the
[5:23] wider world.
[5:24] Stripe now processes almost 2% of global
[5:27] GDP. So, we've really useful window into
[5:29] the forefront of the internet economy.
[5:32] There's three areas that I want to dive
[5:34] into deeper with you.
[5:37] The first is that we're seeing a
[5:39] structural increase in economic
[5:41] dynamism.
[5:43] More firms are getting started. They're
[5:45] keeping head count lean for longer,
[5:47] but they're scaling revenue faster than
[5:49] we've ever seen before.
[5:51] To explore this more, I'd like to call
[5:53] in some help from an actual economist,
[5:55] Stripe's head of data and AI, Emily
[5:57] Glassberg.
[6:07] >> Thanks, John.
[6:11] There's plenty of talk about AI not
[6:13] showing up in the macro data yet, like
[6:15] the labor stats we just saw.
[6:18] If you're as AI-pilled as most of us in
[6:20] this room, that probably creates some
[6:22] cognitive dissonance.
[6:24] But here's the thing.
[6:25] AI actually is showing up in the macro
[6:27] data. You just have to know where to
[6:29] look.
[6:30] Here's US business formations over the
[6:32] past 20 years.
[6:34] You can see a huge bump during the
[6:35] pandemic, not surprising, but you can
[6:38] also see it re-accelerating now.
[6:41] Basically, all of the recent growth is
[6:44] coming from what the Census Bureau calls
[6:45] non-employer firms, or what we all call
[6:48] solopreneurs. That's the blue line here.
[6:52] Now, these didn't used to be considered
[6:54] serious businesses, but AI means more of
[6:57] these businesses are getting to real
[6:59] scale.
[7:00] This is the number of solopreneurs in
[7:02] the US doing over 100,000 in revenue.
[7:06] Yes, solopreneurism is now how close to
[7:09] 5 million Americans earn their living.
[7:12] And we're fortunate to have many
[7:14] thriving solopreneurs building with
[7:16] Stripe.
[7:25] It's a great crew. This isn't just an
[7:28] American story. New business
[7:29] registrations are up 40% in Australia,
[7:32] up 70% in Finland, up 80% in France. So,
[7:36] the surge in dynamism is happening
[7:38] across advanced economies.
[7:40] At Stripe, we're seeing this firsthand.
[7:43] >> Yeah, Stripe Atlas is the simplest way
[7:45] to get incorporated in the US. And last
[7:47] week, we celebrated our 100,000th Atlas
[7:50] founded business.
[7:51] Somewhat topically,
[7:53] um yes, a 100,000. So, somewhat
[7:55] topically for this talk, that business
[7:57] is Empirical, which is building AI
[7:59] software that optimizes profitability
[8:01] for battery battery energy storage
[8:03] system. Sounds like something we're
[8:05] going to need. And I actually think
[8:06] Rachana, the founder, is here today.
[8:08] But the point is, we hit 100,000 Atlas
[8:11] incorporations way earlier than we
[8:13] expected. The number has exploded.
[8:15] There's more of them, and they're
[8:17] scaling up like crazy.
[8:18] >> Yeah.
[8:19] >> They really are. So, in aggregate, Atlas
[8:21] companies incorporated in 2025 are
[8:24] raking in twice as much revenue as the
[8:26] class of 24 was by this point. Nothing
[8:29] shabby about a doubling.
[8:31] But it's the class of 26 that's really
[8:34] cooking. So, let's zoom in a little bit.
[8:37] And you can see that just a few months
[8:39] into the year, the class of 26 is
[8:42] tracking to five times the revenue of
[8:44] last year's cohort.
[8:47] These are very unusual growth rates, and
[8:49] they're driven in part by a generational
[8:51] shift in how companies grow
[8:53] internationally. So, one simple way to
[8:55] see this is to ask, across Stripe, how
[8:58] many companies are earning most of their
[9:00] revenue outside their home country?
[9:03] Five years ago, that was 11.6%.
[9:07] Since then,
[9:08] it's doubled.
[9:10] And they aren't just selling in the
[9:11] obvious places. Among companies making
[9:13] most of their money cross-border, around
[9:15] a quarter are making most outside the
[9:17] top 10 global markets.
[9:20] It's a real long tail story.
[9:23] The international pecking order has also
[9:25] flipped. It used to be that young
[9:26] companies sold mostly at home, and only
[9:29] ventured out to the open sea of global
[9:30] commerce once they'd gotten big.
[9:33] Now, it's the newborns on Stripe who are
[9:35] globetrotting their way through the long
[9:37] tail of markets.
[9:39] Take the top 100 AI startups on Stripe.
[9:42] The median earns most of its revenue
[9:44] internationally and sells into 55
[9:47] countries within its first year of
[9:49] existence.
[9:50] Emergent Labs was founded in the US in
[9:52] 2024,
[9:54] but already nearly 70% of their revenue
[9:56] comes from abroad.
[9:58] They're doing material business in many
[9:59] markets. No less than 16 countries drive
[10:02] at least 1% of Emergent's revenue. So,
[10:05] this is a totally new startup playbook.
[10:07] You launch globally on day one, you keep
[10:09] headcount very lean, and you automate
[10:12] aggressively.
[10:13] But, not every business can be a
[10:15] solopreneur or a top AI startup. So,
[10:19] what does all of this mean for the
[10:20] median firm?
[10:22] Despite the fact that one of us has a
[10:24] PhD in economics and the other dropped
[10:27] out of college and now hosts a podcast
[10:31] in a pub.
[10:34] John is going to teach us a little bit
[10:36] about economic history.
[10:46] >> I think you might be surprised to learn
[10:48] just how many economic theories have
[10:50] been devised over a pint, Dr. Sands.
[10:52] Um okay, some economic history. Are we
[10:55] ready?
[10:56] Uh in 1931, a 21-year-old British
[10:59] student took a road trip across
[11:01] depression-era America. And back then,
[11:04] as now, bunch of young people pretty
[11:06] into communism.
[11:08] Central planning had this real
[11:10] intellectual prestige at the time. All
[11:12] the smart people thought that top-down
[11:14] coordination might be the way to do
[11:15] things, much better than messy, chaotic
[11:18] markets.
[11:19] And the student noticed something
[11:20] interesting.
[11:21] Within a firm, resources aren't
[11:24] allocated by prices. There's no little
[11:25] market inside the company. Companies
[11:28] themselves are centrally planned.
[11:31] And so, he wondered, if markets are so
[11:34] efficient, why do we keep building
[11:36] little islands of central planning
[11:38] inside markets?
[11:40] His name was Ronald Coase.
[11:42] And his answer was that firms exist
[11:44] because coordinating inside a company is
[11:46] often actually cheaper and easier than
[11:48] coordinating through markets.
[11:50] And that answer won the Nobel Prize
[11:53] 60 years later. They don't rush into
[11:55] decisions over there in Stockholm.
[11:57] >> I know what you're thinking. John's
[11:59] little economic stories
[12:01] heavy on story and light on economics.
[12:04] But here's where we're going.
[12:07] What's the Coasean reading of AI?
[12:10] In the near term, the within firm effect
[12:12] is the most obvious. Companies have
[12:14] shared context and systems of record and
[12:16] aligned incentives and all of that
[12:18] allows for easier coordination with AI.
[12:21] You're probably feeling this in all your
[12:23] own companies already.
[12:25] But in the medium term, external markets
[12:27] are likely to get even more efficient.
[12:30] Agents are already great at discovery.
[12:33] They make it trivially easy to integrate
[12:35] a new piece of software. They make
[12:37] contracting much more straightforward.
[12:40] And as we'll discuss in a minute, they
[12:41] can transact much more frequently and
[12:44] for smaller amounts than humans can.
[12:47] Together, all of this should bring
[12:49] coordination costs between firms down by
[12:52] quite a bit. Now, some of these changes
[12:54] will take a little while to play out,
[12:56] but some of them are already here.
[12:59] So, on that, with AI, we expect fewer
[13:02] people per firm,
[13:03] more output per firm, just more firms,
[13:07] and more coordination happening through
[13:09] market-like mechanisms.
[13:12] Coase would have liked that idea. He
[13:14] always thought firms were an inefficient
[13:16] setup. The more we can do via markets,
[13:19] the better.
[13:20] >> An indie hacker from another era. Yeah.
[13:22] Thanks, Emily.
[13:30] >> Okay.
[13:31] So, that was the explosion in business
[13:34] dynamism that we're seeing.
[13:37] The second trend I want to talk about is
[13:39] how commerce itself
[13:41] is becoming agentic.
[13:44] At Stripe, we think about this in
[13:45] increasing levels of autonomy in the
[13:48] purchasing flow from simple help all the
[13:50] way up to significant agency.
[13:52] You might be wondering, where have we
[13:53] actually gotten to in 2026?
[13:56] Well,
[13:57] level one is already here. This is
[13:59] software schlepping through forms on
[14:01] your behalf.
[14:02] The in-app checkout experience we're
[14:03] doing with Meta is actually a good
[14:05] example this. So, maybe you know, you
[14:06] find something, maybe in an ad, you
[14:09] express interest, and the agent it has
[14:11] your details already, and it can
[14:13] complete the checkout for you. Super
[14:16] handy, really convenient, but not
[14:18] exactly science fiction. I mean, you
[14:19] mightn't even think of this as agentic,
[14:21] even though strictly speaking the
[14:22] software is doing the purchasing for
[14:25] you. It is your agent.
[14:27] Level two is the shift from plain old
[14:29] keyword search that we've had for
[14:31] decades to a shopping assistant that can
[14:33] actually reason within constraints and
[14:35] find products accordingly.
[14:38] Think about when you do some shopping on
[14:40] ChatGPT. Imagine, totally
[14:42] hypothetically, you say, "I need a
[14:44] birthday gift for my brother. He's 38
[14:47] and has kind of weirdly at this late age
[14:49] gotten very into calisthenics, but he
[14:51] already actually has a lot of
[14:53] calisthenics gear. And so, what's a
[14:54] non-obvious gift that I could get for
[14:56] him for under $100?"
[14:59] Hypothetically. Um,
[15:01] it's not obvious, you know, what
[15:03] keywords you would put in to get these
[15:05] results.
[15:07] Wayfair does something similar. You can
[15:09] describe a room or a style or a feeling,
[15:12] and the agent rummages through the
[15:13] catalog for you.
[15:15] And you're still in control here. You're
[15:16] making the buying decision. You don't
[15:17] have some awful vase you didn't want
[15:19] showing up unexpected.
[15:21] But you have better ways to find what
[15:23] you're actually looking for.
[15:25] Most people say they already shop this
[15:27] way. I mean, you've probably all already
[15:29] done this.
[15:31] But what does it take to get to levels
[15:32] three, four, and five where agents are
[15:35] making and executing purchasing
[15:37] decisions with real autonomy?
[15:40] Well, one way to peer into the future is
[15:42] Open Claw. And there, the demand for
[15:44] autonomous commerce is really palpable.
[15:46] This is the cumulative downloads of
[15:48] payment-related skills on Claw, about
[15:50] 125,000 in 12 weeks.
[15:53] And this is despite the fact that Open
[15:54] Claw is still pretty hard to use for
[15:56] regular folks.
[15:57] So, the question isn't whether there's
[15:59] demand, there is.
[16:01] It's how do we get what's already live
[16:02] at the frontier to go mainstream?
[16:07] Will talked yesterday about how we need
[16:08] the economic infrastructure for AI.
[16:11] Agents need to be able to pay,
[16:12] businesses need to be able to accept
[16:13] payments from those agents, and the
[16:14] whole thing needs a trust layer.
[16:17] And Stripe is working hard at getting
[16:18] all of this deployed.
[16:21] There's actually one corner of commerce
[16:23] where things are already moving quickly,
[16:25] which is software buying from software.
[16:27] Let me show you what I mean.
[16:32] So,
[16:35] previously, you know, yesterday you saw
[16:37] buying stuff, but here just imagine we
[16:39] have an agent that we want to help us do
[16:41] some research. So, I have a question
[16:42] I've been wondering.
[16:46] Hey Claude, how is AI demand affecting
[16:48] commodity prices and supply and demand
[16:50] for different energy sources?
[16:57] You guys are probably still typing to
[16:58] your Claude, but you can just talk to
[16:59] it. Um, so,
[17:03] what's going on here? You've probably
[17:04] heard so much about how, you know, the
[17:07] AI build-out is this massive capex boom.
[17:09] We're building all these new data
[17:10] centers. And at various points in the
[17:12] past, it's actually been power
[17:14] constrained. Uh, and so, you know, we
[17:17] need to plug these data centers into
[17:18] something, but
[17:20] you know, for many, many years we
[17:21] haven't actually expanded the US grid,
[17:23] and now we're adding all this new
[17:24] demand. And the electricity grid is a
[17:26] market. It's got a, you know, supply and
[17:28] demand equilibrium. And so, just when
[17:30] you have this
[17:32] equilibrium that's existed for many
[17:33] years, and then you plonk, you know,
[17:35] this new demand comes along and
[17:36] cannonballs into the pool, just what
[17:38] happens? And so, I asked Claude to go
[17:41] research this, and it's going off and
[17:44] it's like finding things. And here,
[17:46] okay, it said, "It's picture is
[17:48] striking. It is very I'm going to need
[17:50] some commodity and equity data to ground
[17:51] this analysis. Off advantage has what I
[17:53] need. I'd like to buy this stuff. Total
[17:55] is 4 cents. Yes, I thought it's within
[17:57] our budget."
[18:03] Okay, and it is
[18:05] blanching. Um
[18:08] So, what did you see here? The agent
[18:10] analyzed my question. It looked for the
[18:12] relevant sources. It found a paid
[18:15] source. And now, it's off buying and
[18:18] downloading that a data autonomously.
[18:20] And you're probably used to your AI
[18:22] doing lots of thinking and building, but
[18:24] then it's asking you to carry out the
[18:26] grunt work. It wants you to do the
[18:27] deployment or the checkout flow or the
[18:28] sign-up. But where we're rapidly headed
[18:30] is the agent doing that work for you.
[18:34] In the demo yesterday, Will used the
[18:36] Link CLI to pay the API reviewer. Here,
[18:38] as you can see from some of the Tempo
[18:40] requests at the top, we're actually
[18:41] using the Tempo CLI because my agent has
[18:43] a stablecoin wallet. You know, machine
[18:45] payments can use fiat, but for these
[18:47] tiny purchases, you know, for micro
[18:49] payments, you need a different type of
[18:51] infrastructure. You need stablecoins,
[18:52] which have near-zero transaction costs,
[18:54] making them viable for the first time.
[18:56] So, it's blanching away here. Um See,
[18:59] this is where you really need a fellow
[19:01] with a guitar.
[19:02] Um
[19:05] Just
[19:06] wrong timing.
[19:10] But while and we did splurge for fast
[19:12] mode uh for you guys, but it's still
[19:14] blanching away there. And so I just
[19:16] thought I'd show you some cool
[19:18] tabs.
[19:19] Uh
[19:20] while we're racing. This is from works
[19:23] in progress stripes magazine about
[19:25] progress. You've probably seen us at the
[19:27] cafe out there and everything and just
[19:29] to the discussion of all these new
[19:31] energy sources. We have this cool
[19:32] article about how Britain made a lot of
[19:35] progress and then forgot it all in
[19:37] nuclear.
[19:38] What else do we have? We have the
[19:39] Guinness tank the first real world
[19:41] application of AI where an AI agent
[19:44] called every pub in Ireland to
[19:46] have a real time tracking of the cost of
[19:48] a Guinness.
[19:50] So finally something useful.
[19:55] It's still Okay, it's right Oh, okay,
[19:57] we're done. It's So
[20:00] Claude has given me my output. It's open
[20:02] in my browser. And so you see here
[20:04] again, I gave it a single prompt here
[20:07] just my kind of one word question and it
[20:09] spat out this house this report and
[20:11] what's it saying? US data center
[20:12] electricity demand is projected to
[20:15] nearly triple by 2080.
[20:17] Hyperscaler capex expenditure has surged
[20:19] 62%
[20:21] um
[20:23] natural gas prices have surged 104% you
[20:26] know natural gas is actually picking up
[20:28] a lot of this So anyway, super
[20:29] interesting. We don't have time I would
[20:31] love to just actually read all this. I'm
[20:33] not going to read all this in front of
[20:34] you.
[20:35] But I will be interested to read it
[20:36] later. You guys will probably be
[20:37] interested to read it later. And so what
[20:39] I can actually do is go back to my
[20:41] Claude and say
[20:43] Hey Claude, publish and sell this
[20:45] report. Price it as you see fit for
[20:48] other agents and humans to find and buy
[20:50] it.
[20:57] Great. So it's off working and you know
[20:59] you saw our vibe deploying yesterday and
[21:01] so it's going to go off and make a
[21:03] website where any of you can buy us and
[21:05] actually now that I say that, I should
[21:07] maybe check.
[21:10] Check the licensing terms for this Alpha
[21:12] Vantage data set. Do I actually have the
[21:14] rights to commercially redistribute the
[21:16] final report?
[21:22] Who, okay. Yes, I checked the terms of
[21:24] service, blah blah blah blah blah, we're
[21:26] fine. Um
[21:35] Okay, so it's doing the thing there. But
[21:36] while we wait for the reports to get
[21:38] published, what should you take away
[21:39] from this?
[21:40] Well, agentic commerce, again, it is
[21:42] here.
[21:44] And we think there'll be a really big
[21:46] first-mover advantage or an early-mover
[21:47] advantage. It's one of the reasons we're
[21:49] moving so quickly at Stripe to enable
[21:51] you to do this. If your product or your
[21:53] platform can possibly support
[21:55] machine-to-machine payments, we think
[21:57] you should build for it now. And still
[22:00] going. It's blanching.
[22:05] And again, previously, you had to, you
[22:07] know, you probably had to go get API
[22:08] keys or go poke around in the Versel
[22:10] interface or anything like this. Again,
[22:12] now, thanks to Stripe projects, it can
[22:15] orchestrate all of this for you.
[22:18] We have any other good tabs here. These
[22:20] are some of the companies that already
[22:21] support um agentic commerce. This is at
[22:24] Parallel and uh browser-based for
[22:27] uh
[22:28] agentic web browsing. You have uh Postal
[22:31] Form, which will your agent can mail a
[22:33] letter for you. So, if you want to some
[22:36] compatibility between the new way of
[22:37] working and the old way of working.
[22:38] Okay.
[22:40] Here we go. Uh it is live at
[22:41] johnsreport.versel.app.
[22:43] So, if I just open that, you see here,
[22:45] you can go to johnsreport.versel.app.
[22:48] You can click
[22:50] purchase report, and I see I'm getting a
[22:52] link confirmation there. That's great.
[22:53] That's all working. So, I check out. But
[22:55] also, if I go to LLMs
[22:58] .txt,
[23:01] you see here, it also constructed an
[23:03] LLM.txt for us with instructions, you
[23:05] know, for how with a single temporal
[23:07] request agents can um
[23:10] can buy the product. So, I would uh
[23:12] welcome you. Indeed, I would encourage
[23:14] you. I would beseech you to please buy
[23:16] my report for the uh princely sum of $5.
[23:20] Uh you know, that will help me with my
[23:22] my token budget.
[23:24] Ah.
[23:25] There we go. Um
[23:27] that is it. And agents
[23:29] Yes.
[23:32] Can we Yeah, thank you.
[23:35] I give them the the confetti. These are
[23:37] Gen Zs. They need it. There we go. Yeah.
[23:39] Um
[23:41] So, and that's agent commerce live
[23:43] today. You can go check it out.
[23:47] And it raises a really interesting
[23:48] question, which is my third topic.
[23:52] In a world where intelligence can do all
[23:54] of that
[23:55] and everyone has access to that kind of
[23:57] intelligence,
[23:59] what actually becomes more valuable?
[24:02] There's an old rule in economics. When
[24:04] something gets cheap, its complements
[24:06] get more valuable.
[24:08] So, when containerized shipping
[24:10] collapsed the cost of moving goods,
[24:12] ports that could handle the ships became
[24:15] much more valuable.
[24:17] The first radio spectrum auctions in the
[24:19] 1990s raised hundreds of millions of
[24:21] dollars. But then mobile phones got
[24:23] cheap, loads of people had them, and the
[24:25] same airwaves were suddenly worth orders
[24:28] of magnitude more. Governments started
[24:29] auctioning them off for tens of billions
[24:31] of dollars. So, these things have joint
[24:33] demand curves.
[24:34] And so, one question we should be asking
[24:36] about AI is what are AI's complements?
[24:40] What are the complements to
[24:40] intelligence? What becomes more valuable
[24:43] as intelligence becomes cheaper?
[24:46] Some of the answers are obvious.
[24:48] For example, you can see this effect
[24:49] very clearly in chips. You know, GPUs
[24:53] were really useful before AI.
[24:56] But it's clear from Nvidia's uh
[24:58] deliveries and market cap that chips
[25:00] have become much more useful recently.
[25:02] You know, they were previously a
[25:03] majority gaming and now you see the
[25:06] computing networking segment take off.
[25:09] Same goes for energy.
[25:11] Power is more valuable if you have
[25:13] intelligence to plug it into. Nuclear
[25:15] power is undergoing a renaissance
[25:17] largely because we just we need nuclear
[25:19] power to power data centers.
[25:21] In the meantime, we're going to need a
[25:22] lot of gas turbines as you can read
[25:24] about in the report.
[25:26] Uh you can see the order volumes taking
[25:28] off here. It's also reflected in the
[25:30] market value of the handful of companies
[25:31] that make them. This is Siemens Energy,
[25:33] one of the big gas turbine makers.
[25:37] But less discussed complement to AI is
[25:39] proprietary data which gets much more
[25:42] valuable when you can let super
[25:43] intelligent agents reason over it.
[25:47] One way you can tell data's getting more
[25:48] valuable is that companies that used to
[25:49] give it away for free
[25:51] have stopped doing that.
[25:53] This chart shows the percentage of
[25:54] various parts of the internet that have
[25:55] been shut off to AI crawling and instead
[25:58] those companies are starting to monetize
[26:00] it.
[26:01] Reddit has always had a ton of data, you
[26:03] know, all the comments in the subreddits
[26:04] were always there. But before AI, it was
[26:06] a dormant asset on the balance sheets
[26:08] and today their non-ad revenue which
[26:11] comes from
[26:12] data agreements is $35 million a
[26:14] quarter.
[26:16] We see the same dynamic with our own
[26:17] data. Take Stripe Radar.
[26:19] So, Stripe Radar has always been able to
[26:21] reason across Stripe's entire corpus of
[26:23] data, but in recent years as the AI
[26:25] models have gotten better, the
[26:26] underlying data is then more valuable.
[26:31] Network effect is another one we should
[26:33] talk about.
[26:34] Buyers and sellers still need places to
[26:35] meet and if switching costs are low, the
[26:37] network effects are even more important.
[26:40] To understand this better,
[26:41] we took a look at the public take rate
[26:43] for 10 top marketplaces. What you see is
[26:46] take rates flat for a few years and then
[26:48] you get this bump up during the pandemic
[26:50] and then a steady increase over the past
[26:52] 3 years as marketplaces see higher
[26:53] returns to better AI techniques.
[26:59] The last compliment I want to call out
[27:00] is just
[27:01] companies that have figured out the
[27:03] complex interactions between software
[27:05] systems and real-world execution. I
[27:08] think we'll see that that is an enduring
[27:10] moat.
[27:11] Take John Deere.
[27:13] If I asked you to name an AI
[27:14] beneficiary,
[27:16] John Deere might not be the first name
[27:18] that you'd think of.
[27:19] But they've spent years integrating GPS
[27:21] guidance and machine vision and sensor
[27:23] arrays into their equipment.
[27:24] And the defensible part, the moat here,
[27:26] is not the AI. Lots of people could
[27:27] build the AI.
[27:29] It's having the tractors in the fields
[27:32] across 130 countries.
[27:35] So, those are five things we think
[27:37] probably increase in value alongside AI
[27:40] and therefore create even more durable
[27:42] competitive advantages in the years
[27:44] ahead.
[27:46] And for all of you, AI should change how
[27:48] you think about your own competitive
[27:50] advantages.
[27:51] You might previously have built the best
[27:54] software in your space, but you might be
[27:55] finding that software is not the name of
[27:57] the game anymore. But what you do have
[27:58] is powerful proprietary data,
[28:00] interesting network effects, real-world
[28:03] operations and tools that took a decade
[28:05] to get right,
[28:06] all sorts of advantages that hold their
[28:08] value or even become more valuable in a
[28:10] world of abundant intelligence.
[28:12] So, as we wrap up, will you indulge me
[28:14] in just a little more economic history?
[28:16] I can't
[28:17] I can't resist.
[28:19] In 1882, Thomas Edison lit up 82
[28:22] customers in Lower Manhattan using six
[28:25] dynamos. Finally, electricity in
[28:27] Manhattan.
[28:28] And for decades after, even as
[28:30] electricity adoption grew, productivity
[28:33] growth fairly blunted, even slowed down
[28:35] as the railroad investment boom started
[28:37] to wear off.
[28:39] And the economists were confused
[28:40] initially. I mean, we had this awesome
[28:42] technology and electricity. Why wasn't
[28:43] it showing up in the statistics?
[28:46] And the problem wasn't the technology,
[28:47] the electricity did work.
[28:50] It was the economy had to digest it.
[28:52] You see, factories, they'd been built
[28:54] around steam, the shafts and the belts
[28:56] and the floor plans, it was all wrong
[28:58] for electricity. And it wasn't until we
[29:00] redesigned factories from scratch that
[29:02] the productivity gains finally appeared.
[29:05] And people sometimes forget, but this
[29:07] took a full 30 years from 1882 all the
[29:09] way to the late 1910s. And then, in the
[29:12] 1910s, in that single decade, the growth
[29:15] rate of output per worker more than
[29:17] doubled. But there was this big lag.
[29:21] We saw the same phenomenon again with
[29:22] the birth of computing. In fact,
[29:24] economists have since dubbed this whole
[29:26] phenomenon the Solow Paradox after
[29:27] Robert Solow's 1987 quip that computers
[29:31] were everywhere to be seen except in the
[29:33] productivity statistics.
[29:36] And he was right.
[29:37] And they weren't to be seen, and they
[29:38] wouldn't be until the mid-1990s.
[29:43] Transformative technology looks for a
[29:45] long time like it's not doing much. If
[29:48] you're looking at the economic gauges,
[29:49] you're going to sit there and you're
[29:50] tapping the gauges, is this thing on?
[29:53] I think this is what we're actually
[29:54] watching in real time with AI.
[29:58] You see the seeds of this in the
[30:00] phenomena I already mentioned. The
[30:01] minimum efficient size of a serious
[30:03] business is collapsing. Solopreneurs are
[30:06] scaling to seven figures and beyond.
[30:08] Agents are buying from agents. Companies
[30:10] are launching globally from day one. But
[30:11] none of that fits in the old model.
[30:14] These changes might not yield
[30:15] productivity dividends tomorrow because
[30:17] we have to digest them.
[30:20] But they're the early indicators
[30:22] of an economy that's re-platforming
[30:24] itself.
[30:26] The businesses you're all building now,
[30:28] they're not a footnote to AI history,
[30:29] they are the AI history. They are the
[30:31] story of the economic and productivity
[30:34] gains.
[30:36] Electrification took 30 years to
[30:38] reorganize the economy.
[30:40] But I suspect we won't need to wait
[30:41] anywhere as long as that for AI.
[30:44] Thank you so much for being here. I hope
[30:46] you are getting a ton out of sessions.
[30:48] Enjoy the rest of your morning and I
[30:50] will see you back here for our final
[30:51] fireside with Daniel Gross and Nat
[30:53] Friedman this afternoon.
