How Stripe Quietly Became More Important to AI Than Any AI Company

Levi Strauss didn’t go to California to find gold. He went to sell pants to the people who did. By 1853, his dry goods business was generating more reliable revenue than virtually any of the miners he was supplying — because he got paid regardless of whether they found anything. The miners bore the risk. Levi Strauss collected the margin.

This is not a novel observation about the AI industry. The “picks and shovels” analogy has been floating around venture capital circles since at least 2022. But most of the people who make it name the wrong companies. They point to NVIDIA, which is a real answer but an obvious one. They don’t talk about Stripe.

They should.

The Revenue Model Hidden in Plain Sight

Every AI company that charges for API access runs through Stripe. OpenAI does. Anthropic does. Mistral does. The hundreds of AI startups charging monthly subscription fees for their products do. When you pay for ChatGPT Plus, that $20 doesn’t go directly from your credit card to OpenAI’s bank account — it goes through Stripe’s payment infrastructure, and Stripe takes 2.9% plus $0.30.

That doesn’t sound like much until you do the math. OpenAI is estimated to have been approaching $4 billion in annualized revenue in early 2026. If even 60% of that flows through Stripe (and the actual figure may be higher), that’s roughly $70 million a year to Stripe from a single customer, on a product Stripe’s engineers don’t have to maintain or support. Scale that across hundreds of AI companies and you begin to see the shape of the business.

Stripe’s actual revenue figures are private — the company is still not public as of mid-2026 — but industry estimates put them somewhere between $14 billion and $18 billion in annual payment volume processed. The company earns its cut from all of it. It does not care whether any individual AI company succeeds or fails. It cares that the AI industry collectively processes more transactions next year than this year.

The Stack That AI Runs On

Stripe is one layer. The full stack of AI infrastructure dependencies is worth tracing, because the pattern repeats at every level.

Compute: AWS, Microsoft Azure, and Google Cloud together supply probably 85-90% of the GPU compute that AI companies rent for training and inference. They do not develop frontier models. They sell the electricity, cooling, networking, and hardware to the people who do. AWS’s Q4 2025 revenue was $29.6 billion — up 19% year-over-year — driven substantially by AI workload demand. Amazon’s risk in this trade is essentially nil: AI companies pay upfront for reserved capacity, and the cost of switching cloud providers mid-training-run is prohibitive.

Data storage and databases: MongoDB, Supabase (Postgres under the hood), Pinecone, and Weaviate are doing significant business selling vector database services to AI applications that need to store and retrieve embeddings. None of these companies are building frontier AI. They’re selling the equivalent of file cabinets to people building AI applications.

Communication and deployment: Twilio handles SMS, voice, and email for AI applications that need to reach users. Vercel hosts the web frontends. Cloudflare handles CDN and edge routing. None of them need AI to succeed — they need AI applications to reach users.

The pattern at every layer: pick a company that provides essential infrastructure, find the AI use case, observe that the infrastructure company doesn’t care which AI model wins. They just need traffic.

Why This Is More Durable Than Model Companies

The model companies are in a peculiar position. OpenAI, Anthropic, Google DeepMind, and Mistral are in an arms race that requires them to spend enormous capital — training runs that cost $100 million or more — in order to produce a product that competitors might replicate within months. The capabilities of GPT-4 in 2023 are now available in open-source models that anyone can run for free. The capabilities of o3 in 2025 will eventually be commoditized too. This is just how software technology progresses.

Infrastructure companies don’t have this problem. Stripe does not face a scenario where a competitor releases an open-source Stripe and suddenly payment processing is free. The value Stripe provides — regulatory compliance in dozens of jurisdictions, fraud detection trained on trillions of transactions, SDKs for every major programming language, relationships with banks and card networks — cannot be replicated by a well-funded team in twelve months. The moat is different in kind from the moat a model company builds.

This doesn’t mean model companies are bad businesses. OpenAI is almost certainly going to be a very large company. But the risk profile is categorically different. If OpenAI makes a strategic error — if the model quality stagnates, if regulatory pressure hits, if a competitor out-executes them — the business can deteriorate rapidly. If Stripe makes a strategic error, the business deteriorates slowly, because switching payment processors is painful enough that customers stay even when annoyed. Stripe’s worst-case scenario is considerably less catastrophic than OpenAI’s worst-case scenario.

The Historical Parallel Is Closer Than People Admit

The 1990s internet boom produced a similar structure. The companies that became enduringly valuable weren’t necessarily the ones building the most exciting applications — they were the ones that became structural dependencies for everyone building applications.

Cisco routers carried internet traffic regardless of which dot-com was sending it. Sun Microsystems sold servers to half the companies in the valley. Oracle sold databases. None of these companies needed any individual internet company to succeed — they needed the internet itself to grow. When the dot-com bust came in 2000-2001, Cisco lost 90% of its market cap, which was bad. But it didn’t go to zero, and it recovered, because the internet didn’t stop. The demand for networking equipment didn’t collapse when Pets.com did.

The parallel breaks in one place: cloud computing has changed the capital structure. In 1999, Cisco was selling hardware that companies bought and owned. Today, AWS is renting compute time. This means AWS is more exposed to customer retention than Cisco was — if an AI company goes bankrupt, AWS loses the recurring revenue. But the portfolio effect matters: AWS has thousands of AI customers, not one. The failure of individual companies doesn’t materially affect the infrastructure providers.

What Nobody Talks About: The Data Layer

There’s a layer of the AI infrastructure stack that gets even less attention than Stripe: the data layer.

Training large models requires enormous amounts of high-quality text. The companies that have been licensing that text — Reuters, The New York Times (which settled its lawsuit with OpenAI in early 2025 for a reported $220 million), Associated Press, major academic publishers — are in a structurally enviable position. They own the raw material. They didn’t build anything new. They simply had the content that AI companies need, and they’re charging for access to it.

This is even more like the Levi Strauss analogy than Stripe is. These companies aren’t in the AI business. They’re in the content business, and AI companies need their content. Whether GPT-5 or Claude 5 or Gemini 2 ultimately wins the model competition is irrelevant — whoever wins will have paid licensing fees to Reuters.

The content licensing market is early and chaotic. Prices vary enormously. Standards for what constitutes “fair use” for training data are being litigated in courts across multiple countries simultaneously (as of mid-2026, no final binding precedent has been established in any major jurisdiction). But the direction of travel is clear: content owners are getting paid, and the amounts are going up.

The Network Effects That Don’t Get Discussed

Stripe’s position is stronger than the revenue math alone suggests, because of what the revenue math doesn’t capture: switching costs and network effects.

A company that processes payments through Stripe has integrated Stripe’s SDK into its codebase, trained its finance team on Stripe’s dashboard, built its fraud workflows on Stripe’s Radar system, and likely configured its subscription billing logic using Stripe’s Billing product. Switching to a competitor — Braintree, Adyen, even a direct payment processor — requires rebuilding all of this, re-negotiating rates, retraining staff, and accepting some period of payment processing instability during the transition. For a company where payment processing is mission-critical infrastructure (and for an AI API company, it is), that risk is prohibitive. Stripe knows this. Its pricing reflects it.

The network effect is subtler but real: Stripe’s fraud detection improves with transaction volume across the entire merchant base. A fraud pattern identified in one Stripe customer’s payment stream gets flagged across all of them. An AI company using Stripe benefits from fraud detection data from millions of other merchants who have nothing to do with AI. This is a genuine network effect — the more merchants use Stripe, the better Stripe’s product gets for all of them — and it compounds over time in ways that simple market share data doesn’t capture.

The Investor’s Takeaway (And Its Limits)

The “infrastructure over models” thesis has real merit. As a framework for thinking about where durable value accrues in the AI transition, it’s more defensible than chasing whatever the hottest model company is at any given moment.

But it has limits. Infrastructure businesses trade at lower multiples than growth-stage AI companies, which means you don’t get the upside if you’re right about the thesis — you get stability and reasonable returns, not ten-times appreciation in two years. NVIDIA has been the exception, not the rule: the GPU scarcity that drove its extraordinary stock performance from 2023 to 2025 was a specific, temporary bottleneck condition. Infrastructure companies in general don’t produce moonshot returns; they produce compounding returns. That’s excellent if you’re patient and it’s catastrophic for anyone expecting the AI boom to turn their infrastructure bet into a venture-style outcome.

The other limit: infrastructure companies have their own risks. Stripe faces regulatory risk in every jurisdiction it operates — and payment regulation is a domain where governments move unpredictably. AWS faces antitrust scrutiny that, if it resulted in structural remedies, could fundamentally change the competitive dynamics. MongoDB faces the possibility that a well-resourced competitor (say, Google with AlloyDB) undercuts them on price in the vector database market.

None of these risks are negligible. They’re just different in kind from “our best researcher just got poached by a competitor who can outbid us on compute for the next training run.”

The AI gold rush is real. The miners are exciting to watch. But Levi Strauss built a company that’s still running 170 years later. The miners are mostly names in history books.