Summary

Stripe’s cohort data pushes against the software-commoditization thesis: first-half signups grew 50% YoY, the median 2026 cohort generates 50% more revenue than 2025’s (which was 70% above 2024’s), and the new SaaS-platform cohort is 103% larger despite the narrative that vertical platforms would struggle. Gaybrick says the commoditization thesis is “plausible” and that “we’re either in the singularity or creeping towards the singularity,” but “we are seeing the exact opposite right now”—more software created, monetized faster, with ElevenLabs using 14 Stripe products.
Stripe’s internal agent fleet, “Stripe Minions,” went from ~1,200 one-shot PRs per week at its January/February blog post to 7,000 last week—roughly 30% of all Stripe PRs—and Minion PR share is one of the company’s most important metrics. The design principle: “you’re not going to iterate, you’re not going into planning mode, you’re just going to say this is what I want, go do it.” The org answer is flatter, smaller teams: Stripe Projects was mostly built by a PM and one senior engineer in a few weeks, and that engineer now orchestrates 16 agents.
Gaybrick’s contrarian resource-allocation call: most companies treat agentic efficiency as opex reduction; Stripe’s belief is “I’m being a little cheeky, but build everything,” because “optimizing your cost structure is going short on your own future potential.” Evidence for Jevons over layoffs: internal AI tool KAI, built by two people, lifted seller productivity 20%—and the conclusion was “we need a lot more sellers”—while global tax filing shipped in a third of the time US filing took, at greater complexity.
Agentic commerce hasn’t had its “Claude Opus 4.5 Cambrian explosion moment” because primitives are missing—hence the machine payments protocol with Tempo, where a service returns a 402 response saying “here’s how you buy me,” and the Link agent wallet CLI atop 400 million Link users. Gaybrick thinks checkout pages could go away, at least in the best-case scenario, for humans as well as agents. The most exciting area is B2B: agents provisioning Vercel or Browserbase via Stripe Projects. David George says this already shapes a16z dev-tool theses: “assume the agents are going to be the shoppers.”
Micropayments, a failed idea since the dawn of the internet, may finally work because agents plus stablecoins remove the old friction. The old squeeze—per-article pricing loses to both subscriptions and ads—breaks when agents act as “little hummingbirds going around the internet just slurping up a little data here… doing a little compute over here” on ephemeral budgets (“your budget is $15, go”), without creating a monthly account for every service.
The stablecoin thesis is geopolitical plumbing: national rails like UPI (~86% of sub-$5 payments in India versus single-digit percentages on US cards) and Pix work locally, “but you need a Schelling point for the global economy,” which crypto rails could provide. Concretes: Stripe reaches around 60 countries in fiat but around 150 in stablecoins; Felix Pago is now between 5% and 10% of remittances along the US-Mexico corridor after a few years. Tempo aims to provide a payments-specific chain with first-class privacy, stable throughput and fees, with DoorDash among its collaborators.
George estimates Stripe is over $2 trillion in volume; Gaybrick sees tokens blurring into money. Attacks on general-purpose token consumers like Cursor and Replit are “very reminiscent of what we see in terms of people trying to steal money from Stripe users,” creating a mandate to make token↔dollar movement as safe as dollar↔euro. On fraud, Gaybrick thinks Cursor was Stripe’s first user in this area and says something like one in six free-trial users were abusive; Stripe’s signals now help ElevenLabs block 2,000 free-trial abusers daily.
The strategy in one line, via Clerk’s Colin: “win all the startups and then win them again”—startups are canaries for the next opportunity and hold Stripe to the highest standards (enterprise CSAT praises Stripe reporting; startups call it “garbage”), pulling Stripe toward nearly half the Fortune 500. Taste scales through top-down repetition, product use and simulation: PII-free, growth-randomized synthetic accounts with live-feeling disputes and refunds. George’s investor frame is that founders are the asset class that finds the next product areas; Gaybrick’s closing ambition is to “build the next Stripe inside of Stripe.”

Deep dive

1. Stripe inverted: from payments company to friction-killing, agency-granting platform

Gaybrick’s framing: Stripe has “inverted our value proposition” from a payments company with add-ons to a multiproduct financial-infrastructure platform—roughly 25–30 headlining branded products, plus hundreds and thousands of features. George says the average AI company uses 11 Stripe products and cites 288 distinct product launches at Sessions; Gaybrick responds, “I believe you had 288.” The organizing test for everything: reduce friction and increase agency across whatever touches revenue and cash.
The AI-era fraud story as told: free-trial abuse “wasn’t really an issue pre-AI” because wasted compute was negligible, but software now has a cost structure. Gaybrick thinks Cursor was the first user Stripe worked with on this problem; he says something like one in six free-trial users were abusive, with some opening repeated accounts and possibly doing model distillation. Stripe “got in the bunker with them” and stood up a weekend pipeline using its foundation model, signals across the Stripe network, embeddings and a reasoning layer that explains why each signup looks abusive. ElevenLabs now blocks 2,000 abusers a day using Stripe signals.
On the agency side: Stripe Managed Payments stands in as merchant of record in long-tail markets—“the Kazakhstans” and across APAC—handling tax calculation and remittance so digital-goods AI companies can compliantly sell in 100-plus geographies without entities.

2. “Win all the startups and then win them again”

Gaybrick credits Colin at Clerk for the summary. Startups grow into tomorrow’s giants and are “canaries for what the next opportunity is”—but the subtler reason is standards: enterprise CSAT praises Stripe’s reporting (“way better than anything we get from any of our other” providers, whose incumbents they are accustomed to), while startups call the same reporting “garbage. You’ve got to fix this.”
Working with startups forever pulls Stripe upmarket—approaching, but not quite at, half of the Fortune 500, with Amazons and Microsofts included—with a different sales and post-sale activation motion but the same playbook: stay close to the user and be provably better on their metrics.

3. The cohort numbers: software creation is exploding, not commoditizing

First-half signups grew 50% YoY; the median 2026 cohort generates 50% more revenue than the comparable 2025 cohort, which generated 70% more than the comparable 2024 cohort. Gaybrick’s two drivers: AI widens the opportunity landscape (“you couldn’t build a Suno four years ago… you couldn’t build a Higgsfield”) while agentic coding lowers the cost of building.
Against the nine-month-old narrative that vertical SaaS platforms would struggle, Stripe’s new SaaS-platform cohort is 103% larger in 2026 than in 2025, and Stripe Billing usage grows faster than Stripe overall because it skews toward new software-company creation.
On the commoditization hypothesis, hedges intact: it’s “plausible”—“we’re either in the singularity or creeping towards the singularity”—“but we are seeing the exact opposite right now.” More software is being created and adopted faster than before, with more Stripe products per customer: ElevenLabs uses 14 (“sort of wish it were 11… actually I’m glad it’s 14”).

4. There’s no one left to copy—so make Stripe a platform for founders internally

Stripe deliberately borrowed its institutional mechanics—Google-like OKRs, a Microsoft-shaped sales org, Apple’s DRI culture, product-quality standards and reviews, and day-to-day mechanisms from Alan Mulally—but now “there’s just no one to copy… when a single engineer can do what two teams of engineers could do two years ago.” The theory: Stripe must become internally what it’s always been externally—a platform for founders. Examples include Metronome’s Scott Woody running Metronome and billing; Henri leading most of crypto; Zach leading OpenUSD and Bridge; Privy’s A.A. leading crypto engineering; and Lemon Squeezy’s J.R. Farr leading Treasury.
Stripe Minions are one-shot agents—“you’re not going to iterate, you’re not going into planning mode, you’re just going to say this is what I want, go do it”—running through CI/CD and testing before human review. From ~1,200 PRs per week at the January/February blog post to 7,000 last week, about 30% of all PRs, Minion PR share is tracked as one of Stripe’s most important internal metrics.
The org answer is “flatter,” with smaller teams and engineers operating as engineer-PM-designer hybrids. No dogmatic team size: a line-managed team of eight is “just doing 3× more… the team is doing three things rather than one,” so manager-to-IC ratios may hold while output multiplies. The senior engineer associated with Stripe Projects now orchestrates 16 agents from one screen; the project itself was mostly built by a PM and that engineer in a few weeks, with a couple of others contributing.
The bottleneck has moved to the back office: “we are merging so much more code than last year and it is stressing every system”—seller systems, pricing pages and training sellers fast enough—so the optimization target is the full critical path from user ask to product in users’ hands.

5. Build everything: “optimizing cost structure is going short on your own future potential”

Many companies read agentic efficiency as a way to shrink opex—layoffs and cost cuts. Stripe’s belief: “I’m being a little cheeky, but build everything,” because “the best way to optimize your cost structure is to grow more” against “reams of user asks going back many years that are unmet.”
Jevons in action: KAI, the internal knowledge-AI tool built by two people about six months ago, is at 83% weekly actives and about 60% daily actives, with seller productivity up 20%—“we’re not saying ‘wow, we need fewer sellers.’ We’re saying ‘we need a lot more sellers.’” Payback on sellers just improved.
The compression proof points: global filing—more complex than US filing—was built in a third of the time; spend management, which might have been roadmapped for two years from now, was picked up by one engineer as a project (Gaybrick wrote the first commit). His analogy, “a bit reductionist”: postwar consumer goods and toys exploded after injection molding got good at melting plastic evenly and injecting it into metal castings. Stripe’s code production now works similarly, with molds, templates and AGENTS.md files in every repo: “You can hand off so much of the work to agents and they just get it done.”
One hedge kept: this is still “a good moment for company leaders to create some resourcing back pressure,” since that pressure drives more efficient and more expensive use of new tools—but he expects budgeting eventually to normalize toward “doing more with more.” George’s investor framing is that founders are the asset class in late-stage venture: they find the next product areas, making future revenue opportunities a reason to reinvest rather than treat cost optimization as the only option. His translation: cost-cutting is finite, like returning cash instead of reinvesting it.

6. Agentic commerce is pre-Cambrian—and checkout pages may disappear

“We talked about the Claude Opus 4.5 Cambrian explosion moment—we haven’t had that in agentic commerce”: no canonical, repeated use cases yet because primitives are missing. Stripe’s answer is the Machine Payments Protocol created with Tempo, where a service can return a 402 response that “just says, here’s how you buy me.” Agents crawling checkout forms via browser automation is, per George, “the skeuomorphic version—there’s going to be a native version,” and Gaybrick agrees.
Gaybrick’s non-speculative, best-case call is that “checkout pages shouldn’t exist for humans”: users should be able to say “buy it” on a product display page. George adds that this can apply to humans too, and Gaybrick agrees; Stripe has launched the Link Agent Wallet and Link CLI, with humans in the loop for credential use, alongside Link’s 400 million users.
The most exciting area is B2B: Stripe Projects lets an agent adopt Vercel for hosting or Browserbase without visiting the provider’s website. George says this already informs a16z investing: “assume the agents are going to be the shoppers in the future—is this the one the agents will want to pick?” A demo showed an agent picking up a Browserbase session, researching on ESPN.com and filling out Gaybrick’s NCAA bracket. (“How’d it do?” “I don’t actually know… probably better.”)
Why micropayments work now after decades of discussion: the old objection—per-article pricing gets “squeezed between the subscriber business model and the free business model”—was “probably true in the past,” but agents on complex tasks can be “little hummingbirds going around the internet just slurping up a little data here… doing a little compute over here.” Nobody wants a $9.99/month account per service; the model is ephemeral consumption—“agent, discover the services, your budget is $15, go”—and agents can move dollars into stablecoins or use a shared balance without minding the back-and-forth.

7. Stablecoins as the global Schelling point; Tempo as the payments-only chain

The infrastructure-nerd case: nationalized rails are excellent locally—roughly 86% of sub-$5 payments in India ride UPI versus single-digit percentages on US cards; Pix has grown meteorically in Brazil—“but you need a Schelling point for the global economy… crypto rails solve this political problem.” Hedge preserved: universal stablecoin use “might be one of those should-work-this-way, may-never-work-this-way problems,” which Stripe is “resolutely focused on trying to change.”
Stablecoins are native balances in Stripe Treasury alongside USD/EUR/GBP; Stripe reaches around 60 countries in fiat but around 150 in stablecoins—“you can build a powerful AI company with two engineers in Thailand just as well as in the US.” Felix Pago, built on stablecoins, is now between 5% and 10% of remittances along the US-Mexico corridor after a few years. George contrasts that with the longer timelines involved in building efficient fiat-remittance businesses such as Wise and Western Union.
Tempo’s rationale: “there was no obvious need for another blockchain outside of payments.” A payments-specific chain aims to make privacy a first-class primitive, preserve throughput during trading events and avoid transaction-fee spikes and floating gas fees. It is still early, is working with DoorDash, and is being made the default—but not the only—blockchain in Stripe.

8. Tokens are becoming money—and taste scales through simulation

George frames Stripe as over $2 trillion in volume, with the estimate hedged as “I think.” Gaybrick says, “I’ve always loved the word token… it just sounds like an approximation of money. And increasingly that’s just what it is.” Attacks on general-purpose token consumers like Cursor and Replit are “very reminiscent of what we see in terms of people trying to steal money from Stripe users,” so Stripe now claims a mandate—its internally preferred word over “rights”—to make moving between tokens and dollars “as seamless and safe as moving between dollars and euros.”
The labor framing: many long-running agentic tasks will replace services—“close my books… I used to pay a human in dollars for that; now I pay fewer humans, or a human, in tokens.” Two token-management problems follow: the opex side, tokens spent building, and the product-efficacy side, which models to shift between in-product. Stripe wants to help with both, tied into Treasury and spend management.
George’s closing question asks how Stripe maintains its widely praised taste when AI writes more of the code. In the exchange, Gaybrick says the cynical view is that “taste” is how people justify future value; George sharpens that into “no one actually has taste and we’re all justifying our future value.” Gaybrick says the stronger claim that models will never have taste is probably not true.
Gaybrick’s actual answer is that product quality is core both because users deserve carefully crafted tools and because teams enjoy building products that feel great. Scale it by repeating quality top-down “over and over and over again,” using the product, and simulating usage. Stripe says it can point at an account and request something that looks like it; George specifies a PII-free, growth-randomized replica that “feels live,” with synthetic disputes and refunds, so EMs experience what users experience and can declare, “our next sprint is going to be elevating quality.” The analogies are aviation simulation and NVIDIA simulating chip performance instead of waiting for the fab. George closes by citing the ambition to “build the next Stripe inside of Stripe.”

$NU: is Nubank Capital One in 1994 or Capital One in 2006? | Vanshap Capital Ep. 25 - DYLAN IS HERE, LIVE! | Dylan Patel & Jordan Nanos