Every assumption that made fintech different a decade ago- mobile-first, cloud, and an API instead of a branch- is now table stakes for every financial company on earth. When the assumptions win, the category stops being what’s next; it’s what’s now.

The next category is being built on three different shifts. Decisions are moving from:

Money is tokens. Tokens are money.

This essay is about what that does to finance, who's already doing it, and why the first trillion-dollar finance company will be an AI-first, token-first company.

The stakes are the biggest in the world.

Banking produced more net income than any other industry in 2025. Finance has the world's largest profit pool, the second largest industry by market cap, and BCG estimates fintech has reached roughly 4% of it. JP Morgan is about to become the first trillion-dollar bank. No company built on modern technology is close.

This is the map for what comes after fintech.

Fintech has been brutally effective at stripping costs out from the traditional branch, paper, and telephony model of banking. Mobile gave customers a computer in their pocket, and cloud gave companies a much faster, cheaper infrastructure to build on.

This made entire customer segments profitable for the first time. Mobile-first is how Nubank has reached 140 million customers, costing $1 a month to serve but delivering an average of $17.10 in revenue (per Q2 earnings). Compare that to a large bank, where a retail customer brings in roughly $350 a year, costs several times more to serve, and profits come from cross-sell or the wider balance sheet.

Revolut, Nubank, Stripe, Ramp, and Robinhood can keep compounding for years, and they likely will. An investment category can die while its winners become the largest companies it ever produced, driven by new tailwinds.

Fintech’s share of global venture equity deals fell from 14.5% in 2021 to 12.3% in 2025, while the share of fintech deals going to AI companies rose from around 10% to anywhere between 70 and 80% (depending on your source). The marginal VC dollar is now AI-first.

Every financial product is a decision, a record, and a way to reach you. In an 1800s branch, the manager used his decisions to price your loan, wrote it in a paper ledger, and the branch was how you found him. The decision was a bet. The record was a fact. Distribution was a building.

Since then, the decision moved from a person to people with spreadsheets to foundation models with people.

The record moved from paper to a database to a token.

Distribution moved from a branch, to a phone, to an agent.

An intelligence token is a unit of decision. It does work, and its output varies, because a model's answer is probabilistic. A value token is a unit of record. A dollar of USDC or OUSD is backed 1:1 by Dollars or dollar equivalents.

But in the last two eras there were three catches.

The next era will shift these three ingredients again.

On 19 August, Stripe agreed to buy OpenRouter, the gateway that routes 10 trillion tokens a day across 400+ models, for a reported $7bn and change. The same day, Ramp launched Router.com, which sends every request to the cheapest model that clears your quality bar. Two of the best-run companies in fintech decided in the same week that routing AI tokens is a financial services business.

Martin Casado at a16z calls this "tokens are the new dollars". I’d push one step further. Tokens aren’t dollars, and that's why they’re worth anything. A dollar is deterministic. An intelligence token is probabilistic. The router is where a probabilistic unit of work gets priced in a deterministic unit of value, millions of times a second.

That is a financial market, and it just got built by two payment companies.

The fastest-growing finance companies at scale understand this shift, and you can see it in their recent product launches. They’re all innovating on decisions (AI models and agentic workflows), the record (using value tokenization), and distributing to agents and machines.

Decision (AI models and agent workflows)

Record (Tokenization and Stables)

Distribution (Agents as customers)

Robinhood is leaning harder into tokenization than Nubank or Ramp, and Nubank is further ahead with its own foundation model. But taken in aggregate, the pattern emerges.

Underwriting a customer’s ability to repay has always been a competitive advantage in finance. Incumbents' historic advantage was a large balance sheet to absorb losses, and decades of historical data on how consumer loans perform. But ultimately this involved spreadsheets, committees, and at best, machine learning.

Revolut reached its initial scale at lower cost with cloud and mobile. From there, 80m+ customers gave it a dataset nobody else had. So it built PRAGMA, a foundation model trained on its own transaction history, to replace the rules and scores of the Fintech era. The result was a +130% boost in catching bad credit risk and +65% in fraud detection recall. A committee never got a number like that. Nubank did the same with nuFormer, which reads raw transaction tokens and skips months of manual feature engineering.

The same shift is reaching internal decisions.

When every workflow is slightly different, automation fails. Onboarding, underwriting, and month-end close still need humans to fill the gaps. An agent doesn’t need the workflow to be identical every time. It needs the workflow to be legible. Allica, a UK SME lender, runs an agent that takes an unstructured broker email, chases the missing documents, calls the decisioning engine, and returns a credit decision, and in its first production period it decided half of all cases end to end in an average of 12 minutes, on a journey that takes weeks at a big bank.

Intelligence tokens are becoming the competitive advantage.

The “core” of banking is shifting from traditional database balances to value tokens. This matters most when institutions need to agree on ownership, move value, or make an asset available to a wider market.

Historically, companies that made loans to consumers or businesses had to build costly, complex back-office operations to package those and resell them as securities. The loan existed as paper contracts and a record in their proprietary system of record.

Tokenization makes every asset more portable.

Take Figure, which originates home-equity loans as value tokens. Those loans are then made available on a marketplace (Figure Connect) for capital markets buyers. Increasingly, this becomes infrastructure other lenders (Figure competitors or future acquisitions) can use too. Figure Connect handled $2.8bn of quarterly volume in the second quarter of 2026, representing 65% of its consumer-loan marketplace volume. Its 95% adjusted net-revenue growth and 54.6% adjusted EBITDA margin give it a Rule-of score of about 150.**

A shared ledger also fixes the thing fintech never solved. There are no truly global fintech companies. Nubank is a giant in Brazil, big in Mexico and a startup everywhere else. Revolut is strong across Europe and fourth to seventh in most of its larger markets. Two barriers stopped us from getting a truly global hyperscaler.

A token is global and always-on by default, so one product runs on one rail instead of forty local banks. Visa now settles with US issuers in USDC seven days a week, against the five-business-day window it used before, and Robinhood’s stock tokens are live in 120+ countries.

AI agents built to trade 24/7 only work if there are markets that are open and settled 24/7.

If AI becomes useful at financial tasks, it starts behaving like an economic actor. The agent needs a budget, identity, permissions and a way to hold or move value on somebody’s behalf. That creates a new class of financial customer because a Distribution system built for people holding phones is now being called directly by the software working for them.

A human arrives with attention, which is why the consumer internet taxes the journey to a purchase. An agent arrives with a mandate. Get me X, budget Y, constraints Z. It queries, gets structured results and picks in milliseconds. Your brand becomes a line in a model’s context window, or it doesn’t.

Robinhood is the first at scale to treat the agent as a customer. It opened accounts to agents with trading and banking MCP servers, so the software doesn’t have to guess at a UI, and 100,000 users have already handed an agent a limited account. Compare that with the airline that blocks the same traffic as a bot and then wonders why its customer thinks the website is broken. Stripe’s Link is building the half that lets the agent pay, handing it a one-time card for an approved amount so it can spend without ever holding your credentials.

Agents multiply the number of customers a finance business can serve

JPMorgan will get to a trillion first, on deposits, licenses and a balance sheet that can absorb a shock. The old moats aren't dead, far from it, they just stopped being the only ones.

Every financial company now has three places to build an advantage.

The private company scoreboard is exciting because, for the first time, you can draw a line from here to a trillion.

Coatue’s Magnificent 8 holds two fintechs, and Coatue’s own numbers put the odds of $100bn to a trillion at around 31% once you’re over the line.

And don’t discount the new, AI native finance companies. Hebbia says firms with $30tn of AUM use it, and Rogo went from $750m to $2bn in sixteen weeks. If one of them reaches the scale Cursor reached in coding, it won’t call itself a fintech.

The first fintech wave put the bank in your pocket.

The next reaches into the decisions, records, and workflows that stayed behind the screen.

The companies from the first wave can lead that transition, but the category they created no longer describes the work ahead.

Fintech is dead as an investment category because its assumptions won.

ST.

Next time: The coming era is one where we finance the intelligence economy.