Author: Simon Taylor, founder of Fintech Brainfood
Translated by: Jiahuan, ChainCatcher
Cloud computing and the mobile internet are no longer the main forces reshaping financial services. Simon Taylor argues that two new building blocks have taken over: intelligent tokens, the units AI models use to process and generate information, and value tokens, onchain representations of money or assets.
A decade ago, mobile-first design, cloud infrastructure and replacing bank branches with APIs were what made fintech feel distinct. Now those ideas have become baseline capabilities across the financial industry. Once a tool becomes standard, it stops defining the next wave.
Taylor says three shifts are now organizing new financial products and new investment bets. Decision-making is moving from spreadsheets to AI models, and those models run on tokens. Records that once lived inside a single company are being expressed as tokens on shared ledgers. And customers themselves are starting to become agents with wallets.
His framing is direct: money is token, and token is money. The companies that build around AI and tokens, he argues, are the ones most likely to produce the first trillion-dollar financial business built on modern technology.
The prize is enormous. Taylor writes that banking generated more net profit than any other industry in 2025, while financial services remain the world’s second-largest sector by market value. Boston Consulting Group estimates fintech accounts for only about 4% of that pool. JPMorgan is on track to become the first bank to reach a $1 trillion market capitalization, yet no modern technology-based financial company is close to that size.
Fintech’s turning point
Taylor describes fintech as a category that, in a sense, “died of success.” It cut the cost structure of legacy banking, which had depended on branches, paper documents and phone-based service. The mobile internet put a computer in the customer’s pocket, while cloud computing gave firms cheaper and faster infrastructure to build on.
That shift let financial companies make money from customer segments that had once been hard to serve profitably. Nubank is one example. Built with a mobile-first approach, it has reached 140 million customers. According to its second-quarter earnings, the company spends $1 per month to serve each customer and generates $17.10 in revenue per customer on average.
Large banks look different. Taylor writes that a major bank earns roughly $350 per year from each retail customer, but its service costs are several times higher. Profit comes largely from cross-selling additional financial products and from the wider lending and investment franchise.
He does not argue that fintech winners are finished. Revolut, Nubank, Stripe, Ramp and Robinhood may still compound for years. His point is narrower: fintech as a standalone investment category may be ending, even if the winners inside it continue to grow and could become the largest companies the category has ever produced.
Venture data in the piece shows the same transition. Fintech’s share of global venture equity deals fell from 14.5% in 2021 to 12.3% in 2025. Over the same period, the share of fintech funding rounds going to AI companies rose from roughly 10% to between 70% and 80%, depending on the source. New venture capital is now being directed to AI first.
The three parts inside every financial product
Taylor reduces every financial product to three components: decisions, records and the way a product reaches the customer.
In the 19th-century branch model, a bank manager priced your loan based on personal judgment, wrote it into a paper ledger and served you through a physical location. A decision was a bet on risk. A record captured what had happened. Distribution depended on a building.
Since then, each layer has changed. Decision-making moved from individual judgment to spreadsheet-assisted analysis, and now toward collaboration between people and foundation models. Records moved from paper ledgers to databases and now toward tokens. Distribution moved from branches to phones and is now moving toward agents.
Intelligent tokens are the basic units used when AI participates in decision-making. They pay for computational work, but the output is not fully deterministic: the same question can produce different answers. Value tokens are the units used to represent money or asset value. A token such as USDC or OUSD, valued at $1, is backed 1:1 by dollars or dollar-equivalent reserves.
Taylor says the last two eras left three unresolved issues.
- Decision systems did not change as much as they seemed to. Machine learning, rules engines and credit scoring were added, but humans still wrote and adjusted the rules. Foundation models and agents let software take on deeper decision work.
- Records did not stay naturally consistent across institutions. A bank mainframe can maintain highly reliable records, but those records sit inside one institution and have to be copied and synchronized elsewhere. Stablecoins and tokenized real-world assets, or RWAs, let multiple parties share a single programmable ledger.
- Distribution still relied on people to operate interfaces. Better design reduced friction, but someone still had to press the button. A budgeting chart does not automatically make someone better at saving. An agent that is authorized to move idle cash into savings can act without waiting for the user to open the app.
That, in Taylor’s telling, is why the next era changes all three layers again.
Why “money is token, token is money” matters
One of the clearest signals, in his view, came from Stripe and Ramp.
On Aug. 19, Stripe agreed to acquire OpenRouter for a reported price slightly above $7 billion. OpenRouter is a gateway that routes AI requests across models. Taylor writes that it handles model calls involving 10 trillion tokens a day across more than 400 models.
On the same day, Ramp launched Router.com, which sends each request to the lowest-cost model that still meets the user’s quality threshold. Taylor’s point is that two of the strongest fintech operators arrived in the same week at the same conclusion: routing AI tokens is itself a financial services business.
a16z partner Martin Casado described the shift as “Token is the new dollar.” Taylor pushes that idea one step further. Not every token is a dollar, he writes, because dollars express a fixed amount while output generated through intelligent tokens is probabilistic. What routing platforms do is price each unit of AI computation in a deterministic unit such as dollars. That pricing happens millions of times per second.
He treats that as a financial market in its own right, one that has just been built by two payments companies. In his telling, the emerging intelligent economy rests on exactly that loop.
What large financial companies are already doing
Taylor says the fastest-growing scaled financial firms already understand the transition. Their recent product launches show the same pattern: AI models and agent workflows on the decision layer, tokenized assets on the record layer, and products opened up to agents and machines on the service layer.
He notes that Robinhood has moved more aggressively on tokenization than Nubank or Ramp, while Nubank has gone further on proprietary foundation models. Looked at together, though, the direction is clear.
Financial decisions are moving toward intelligent tokens
Assessing whether a customer will repay has always been a core source of financial advantage. Legacy institutions built that edge on two things: balance sheets large enough to absorb losses and decades of consumer credit performance data. But the actual work still depended on spreadsheets, credit committees and, at best, machine learning layered on top.
Revolut first scaled through cloud infrastructure and the mobile internet, which gave it a lower operating cost base. Then more than 80 million customers produced a data asset few peers could match. Using its own transaction history, the company trained a foundation model called PRAGMA to replace the rules and scoring systems associated with the fintech era.
Taylor says the results were material. Revolut improved high-risk loan identification by 130% and increased fraud detection recall by 65%, meaning the share of actual fraud caught by the system rose sharply. His line is blunt: no credit committee has delivered numbers like that.
Nubank pursued a similar path with nuFormer. According to the article, the model reads raw transaction data represented as tokens, removing the need for months of manual feature engineering.
The same shift is moving inside company workflows. Traditional automation breaks down when each process differs slightly from the last one. Customer onboarding and review, credit approval and month-end close still leave steps that people have to handle manually because fixed automation cannot cover every variation. Agents do not require exact sameness. They need to understand the contents and sequence of the work.
Taylor cites U.K. SME lender Allica as an example. Its agent can read free-form emails written by loan brokers, chase missing documents, call the decision engine and return a credit decision. Early in deployment, it handled end-to-end approval on half of cases by itself, taking an average of 12 minutes. The same process at a large bank often takes weeks.
That is why he says intelligent tokens are becoming a competitive advantage.
Financial records are moving toward value tokens
The second transition is in record-keeping. Core financial systems are starting to record asset value not as balances in private databases, but as value tokens. That matters most when institutions need to agree on ownership, transfer value or place assets into a broader market.
In the older model, lenders that originated consumer or business loans had to build expensive and complex back-office infrastructure if they wanted to package and sell those loans as securities. A single loan came with a paper contract and a matching internal accounting record.
Tokenization changes how easily each asset can move. Taylor uses Figure as the case study. The company records newly originated home equity loans with value tokens. Those are loans secured by the equity in a home, meaning the property’s value minus the outstanding mortgage balance. The loans then move into the Figure Connect marketplace, where capital markets investors can buy them.
That infrastructure is increasingly being opened to other lenders, including Figure’s competitors or businesses it may acquire in the future.
In the second quarter of 2026, Figure Connect posted $2.8 billion in quarterly volume, equal to 65% of the company’s total consumer loan marketplace volume. Adjusted net revenue growth was 95%, and adjusted EBITDA margin was 54.6%. Taylor adds the two figures to get a score of roughly 150, using it as a shorthand for strong growth paired with strong profitability.
Shared ledgers, he argues, also address a problem fintech never solved: the lack of a truly global financial company. Nubank is a giant in Brazil and strong in Mexico, but elsewhere it still looks more like a startup. Revolut is powerful in Europe, yet in most of its larger markets it ranks somewhere between fourth and seventh. Two obstacles have kept globally scaled financial platforms from emerging.
- Every new market requires local licensing, regulatory capital and legal staffing. In Mexico alone, Taylor says, a company needs about $100 million in regulatory capital before it has served a single customer.
- Money stops moving on the timetable of legacy rails. Fedwire does not run on Saturdays. CLS and T2 do not run over the weekend. U.S. equities trade only six and a half hours a day. An AI-driven business, by contrast, operates every hour.
Tokens are natively global and can run around the clock. One product can use a shared settlement network instead of integrating with 40 local banks. Taylor notes that Visa can now settle with U.S. issuers via USDC seven days a week, compared with the prior five-day schedule. Robinhood’s stock tokens, he adds, are already available in more than 120 countries.
His conclusion is practical: AI agents built for continuous trading only reach their full potential when the market itself can trade and settle continuously.
Agents are becoming the new front door to finance
Taylor compresses the third shift into one line: “I don’t want to use your agent. I want my agent to use your product.”
If AI can perform financial tasks effectively, then it becomes an economic actor. Agents need budgets, identity, permissions and a mechanism for holding or transferring value on someone else’s behalf. That creates a new type of financial customer. Services that were designed for humans holding phones are now being invoked directly by software acting for those humans.
Human users bring attention, which is why the consumer internet monetized discovery and purchase. Agents bring explicit tasks: get me X, my budget is Y, my constraints are Z. They send queries, receive structured results and make selections in milliseconds. A brand either appears in the model’s context window as one line of relevant information, or it does not appear at all.
Taylor says Robinhood is the first company to treat agents as customers at scale. Through MCP servers for trading and banking, it lets agents call account functions directly instead of guessing how to work through a webpage or app interface. So far, 100,000 users have granted permission-bounded accounts to agents.
He contrasts that with airlines that still block comparable traffic as bots, then wonder why customers say the website is broken.
Stripe’s Link product is described as filling in the payment layer. It gives an agent a one-time payment card that can only be used within an approved spending limit, allowing the agent to complete a purchase without holding the user’s original payment credentials.
For Taylor, that means a financial company can multiply the number of customers it serves because agents expand the surface area of demand.
What comes after fintech
Taylor returns at the end to market structure and scale. JPMorgan, he says, will likely reach a $1 trillion market value first because it has deposits, licenses and a balance sheet built to absorb shocks. Legacy moats are still real. They are simply no longer the only source of advantage.
From here, he says, every financial company can build an edge in three ways.
- Decision: risk choices move toward proprietary foundation models, while agents take on workflows that used to resist automation.
- Record: assets can move outside the institution, and money and markets can run globally on a 24/7 basis.
- Distribution: the customer becomes an agent.
He then points to several private companies that illustrate how current scale could turn into something much larger.
Stripe: valued at $159 billion, with what he describes as the strongest developer lock-in in fintech, rooted in deep technical integration and high switching costs. It also just bought a platform that prices AI tokens.
Revolut: reportedly raising capital at a $115 billion valuation, with more than 80 million customers and a foundation model for financial operations already in production. It continues to add licenses, including recent conditional approval in the U.S. and licenses in Australia and the UAE.
Ramp: valued at $44 billion, with annualized revenue above $1 billion and free cash flow now positive.
Taylor also cites Coatue’s Magnificent 8 list, which includes two fintech companies. Based on Coatue’s own data, once a company crosses a $100 billion valuation, the probability of reaching $1 trillion is roughly 31%.
He adds that AI-native finance firms should not be ignored. Hebbia says institutions using its products collectively manage $30 trillion in assets. Rogo’s valuation climbed from $750 million to $2 billion in just 16 weeks. If one of these companies reaches in finance the scale Cursor reached in programming, he writes, it will no longer describe itself as a fintech company.
A category name that no longer fits
Taylor closes by contrasting two waves. The first fintech wave put the bank in your pocket. The next wave reaches into the decisions, records and workflows that used to sit behind the screen.
Companies from the first wave may still lead that transition. But the category label they helped create no longer captures what comes next.
His final point is that fintech as a standalone investment category has run its course because the ideas that once powered it have become normal industry practice.