- Fintech Brainfood
- Posts
- Two New Protocols That Teach AI How to Pay - x402 & ATXP
Two New Protocols That Teach AI How to Pay - x402 & ATXP
Plus; Tether to raise at $500bn, Zerohash raises $104m and why Apple Pay is iterating nicely, but not innovating at all.
Welcome to Fintech Brainfood, the weekly deep dive into Fintech news, events, and analysis. You can subscribe by hitting the button below, and you can get in touch by hitting reply to the email (or subscribing then replying)
Weekly Rant 📣
Two New Protocols That Teach AI How to Pay - x402 & ATXP.
ChatGPT has 700 million users. 99.9% have spent exactly $0 through their AI assistant.
Not because they don't want to spend money. Because the internet has no native way for AI to pay for things.
When your AI hits a paywall, it shrugs and asks you to handle it. When it wants to buy something, it generates a shopping list and sends you to Amazon. The most sophisticated AI in history is reduced to a helpful research assistant because it can't hold a credit card.
This week, that started to change. Cloudflare and Coinbase resurrected HTTP 402—a 30-year-old ghost in the internet's code with their new x402 protocol. Meanwhile, an ex-Stripe team quietly launched ATXP, and Google pushed forward with AP2. Each protocol tackles a different piece of how AI handles money, but together they're racing to solve the same fundamental problem.
The internet was born without payments built in.
Now we're retrofitting it for an age where your AI assistant might need to buy from other agents while you sleep.
And I’m now 90% confident that will happen via stablecoins.
The Problems Everyone's Trying to Solve
All of the emerging protocols from Google, Cloudflare and the Stripe-alum are trying to solve slightly different problems
Problem | Protocol | Use Case |
How does any bot access any paid resource online? | x402 | Your AI wants to access a paywalled article |
How does an Agent say a human told them to complete a payment online? | Agent to Pay (AP2) | You tell your AI to buy a vintage N64 game when it appears in stock |
How do MCP servers advertise they want to be paid for tool calls | Agent Transaction Protocol (ATXP) | Your flight booking MCP server wants to get paid for queries. |
x402 is baked into HTTP: An MCP Server could use x402 or it could use AXTP. It helps with discovery and payment initiation for any online resource. Because the payment is triggered via HTTP, anything that uses HTTP could trigger a payment request.
AXTP helps any MCP server say I want to get paid. Any agent could easily pay if the agent is compatible. Because the payment is triggered at each tool call, the MCP server could bill per tool call.
Google A2P helps with the authorization problem (somewhat): As I wrote last week “A2P is a standard to cryptographically prove that a user intended to buy an item and provides a framework for how to build this into a secure, auditable process that works with any payments rail.” Any agent can pay for anything with its intent mandates
Nobody's solving the wallet problem yet: Who’s the human who controls the money? How do you know it’s them? The real value the card networks add is the ability to authorize a payment. A card is essentially a representation that you are you, and you’ve found a secure way to approve the movement of money.
Visa Intelligent Commerce and Mastercard Agent Pay help by giving agents tokens
But who will be the wallet? Will Apple Pay live in ChatGPT? Or will ChatGPT build their own?
It’s land grab time, and nobody seems to be focussed on the higher ground. That said, x402 matters, and you should pay attention.
Why x402 Matters
x402 did something clever: it went straight for the internet's jugular - HTTP itself.
The Hypertext Protocol (HTTP) runs everything. Every API call, every webpage, it’s even the backbone for apps. By reviving the 402 status code, Cloudflare and Coinbase are embedding payments into the fabric of the web. Every server that speaks HTTP can demand payment, and has a handshake to make it happen.
x402 is a handshake protocol. It lets machines agree on price and payment method, but doesn't actually move the money. That's handled by existing rails (cards, stablecoins, whatever)
Here’s how x402 works
Imagine your AI tries to access a page that is paywalled, or wants to pay another AI for a service.

PS. You should read more Net Interest. Marc is amazing.
You say “I want to do this thing.” Which is an HTTP request to something like a paywalled article on Net Interest
The server says “pay me then.” The server responds with “HTTP 402 Payment Required”
Your device or AI agrees to pay. Requesting the x402 resource (the article itself and someone to pay).
The payment facilitator enables the payment. (e.g. a PSP or blockchain) collecting credentials and instructing a payment.
The server hands you your article and confirmation payment completed
We don’t know quite yet how ATXP will work (the docs are quite sparse), but its clear there’s an intent to make internet or AI native payments a thing. And nobody wants to miss that. It feels close enough to have FOMO.
The Opportunity for AI payments is enormous.
ChatGPT's 700m users represent a larger user base than Apple Pay globally (at 650m).
Imagine if every ChatGPT user had $10 in credits that their AI could spend autonomously. That's $7 billion in immediate purchasing power, waiting to be unlocked (and a significant stablecoin yield for somebody).
Yet the true prize is to fix what Marc Andreessen (Netscape founder and a16z founder) described the original sin of the internet
the original sin was we couldn’t actually build economics, which is to say money, into the core of the internet and so therefore advertising became the primary business model
Payments didn’t work in the browser. They didn’t work with banks or payment networks.
We tried very hard to build payments into the browser. It was not possible…We made a huge mistake. We tried to work with the banks and we tried to work with the credit card companies…it was sort of the classic kind of single point of failure bottleneck, or at least in this case, two points of failure. Visa and MasterCard essentially had a duopoly at the time, and so they were just literally, if they did not want you to be in the switch, they did not want you to be able to do transactions, you just simply weren’t going to do it.
But where that left us was with an internet who’s business model revolves around data, ads, subscriptions and then commerce is layered in. Card payments still run on ISO8583 (from 1987), even the “new” ISO20022 was created in 2002. Since then we’ve layered on tap-to-pay, and APIs to make it simpler.
But we never fixed the bare metal.
Now we’re at a crossroads where new standards like MCP and A2P are emerging, and there’s a window to fix that original sin. Perhaps x402 or ATXP are ways to do that.
Stablecoins are internet native payments infrastructure.
Stablecoins are quietly becoming entrenched as the payment method for AI. As I was writing this, Cloudflare dropped this tweet.
Cloudflare introduces NET Dollar, a new U.S. dollar-backed stablecoin that will enable instant, secure transactions for the agentic web.
— Cloudflare (@Cloudflare)
1:57 PM • Sep 25, 2025
Read that again. The company that handles 20% of internet traffic is launching a stablecoin. And for some good reasons.
Stablecoins don’t carry the legacy of traditional payment methods and give more space for innovation (albeit at the expense of not having as much distribution).
Stablecoins are a natural choice to begin with because:
Cards are expensive and not internet native. PayPal, Adyen and Stripe have launched various tools to give AI Agents virtual cards. But this isn’t standardized (although the card networks are working to change that). The model also breaks when you’re trying to do fine-grained, microtransaction billing but each transaction costs at least $0.20.
They offer yield to issuers (for now). As an MCP server, tech company or nonbank you can issue a stablecoin and any balance held by wallets becomes yield you get to collect.
Stablecoins have verifiability built in. Its very easy to tell a stablecoin payment happened and the recipient received their funds. That’s harder in tradfi, where layers and layers of infrastructure separate the payment instruction, from the evidence of settlement.
There’s always some tendency for the tech world to be excited about whatever the thing is, but ATXP, x402 and stablecoins are valuable on merit.
(I also noted, buried deep in the ATXP docs, that ATXP uses Privy to manage wallets, and base (from Coinbase) to facilitate payments. )
I doubt stablecoins are the only answer for agentic commerce and payments. I’ve spent too long in payments to think that way. Payments methods don’t die, they just stop growing. But I do think for new agent to agent payment flows, stablecoins will be the default way to pay.
How real is it and should you do something?
The biggest signal I can give you is how everyone I speak to wants to figure out agentic commerce, agent-to-agent payments, and how everyone agrees
It’s not solved
Now is the time to solve it
So, at an absolute minimum, this is a “must understand.” MCP is now a default way for AI to understand how to work with other APIs, resources and AI’s. It happened incredibly fast. And ask anyone who’s worked in the internet and payments for long enough, and they’ll tell you there’s an HTTP 402-shaped hole in what AI needs to be able to transact online.
Coinbase and Cloudflare have both done a good job supporting those existing standards, and were prominent on the recent Google agentic payments protocol A2P.
If you’re a bank or nonbank with consumers, you can probably let this play out and adopt it when it hits scale. Unless of course payment facilitation is a large part of your business.
If you zoom out, there’s something bigger going on. There’s a battle for the business model of the AI era. While the internet era was defined by ads, subscriptions, and then made the legacy payments world kinda work online, the AI era is being born into a world where MCP servers and stablecoins exist (and the latter is regulated).
Questions left to answer.
We’re in land grab mode.
There’s protocols al over the place. But there’s two big things missing I can see.
Who’s going to solve human <> agent identity? It’s wild to me that there’s no embedded Apple Wallet inside of ChatGPT or similar. Maybe they’re working on something like that, or maybe ChatGPT wants to own that thing. Visa and Mastercard have published token specs, and if they do one thing well, it’s that security model between human and payment credential (as well as the rules around what happens when something goes wrong.
Who’s going to build the content / AI marketplace? The biggest unlock for the internet was adtech and ad marketplaces. The fact that advertisers can bid for eyeballs has created insane levels of efficiency. How will AI do that? (Ben Thompson wrote a great piece on this recently)
We need an EMVCo spec for how agents identify, but that’s rail-agnostic. We need a marketplace for agents and resources like tools or content. (Interestingly, ATXP maps nicely to AI tools, and x402 maps to content)
I suspect all of this is being worked on by someone, somewhere and will get announced.
And I can’t look away.
A new Internet or a New Chapter in Payments?
In 1995, we couldn't build payments into the internet because the banks, payment networks and gatekeepers had all of the distribution, and much more security.
In 2025, the gates are open. The incumbents and the new players are innovating at incredible speed. But we're not building one solution - we're building five competing ones.
And the biggest battle - who controls the wallet - hasn't even started.
x402 gives the internet a way to demand payment. AP2 gives humans a way to authorize it. ATXP gives agents a way to pay each other.
But until someone builds the wallet that ties it all together, we're just building beautiful infrastructure for a ghost town.
The protocol war isn't about protocols.
It’s about which protocol fits your existing or desired business model.
And that makes it fascinating.
ST
PS. I’ll admit deep down I was skeptical of stablecoins being the default payment rail for AI, but these announcements, and writing this Rant has changed that.
4 Fintech Companies 💸
1. Gradient Labs - Customer Ops AI Agent for Finance
Gradient Labs agents can resolve complex queries like lost cards or fraud checks in natural language with customers. The customer satisfaction scores massively outperform traditional chatbots (in 80 to 98% range) and resolved 52% of queries automatically from day 1.
🧠These bots outperform humans and are used by London’s hottest startups like Sling Money, Nala, and Yonder. These companies didn’t choose ServiceNow, or a generic AI customer service solution (like Sierra), because financial services is both regulated and has nuanced complex workflows. I wonder if they’ll crack the US.
2. Selfin - The AI First Bank
Selfin helps consumers manage all of their banking products from a single hub and AI “optimizes your finances automatically.”
🧠 Is this PFM with an LLM bolted on or is there more to it than that? The website is… vague
3. uiAgent - The AI Accounting Team
uiAgent helps teams close month-end 85% faster, reconcile with banks 80% faster by automating manual data extractions, and help finance teams reach month-end close. It integrates with workday, Microsoft Dynamics, Oracle, SAP to produce audit ready documentation.
🧠This type of tool is a no-brainer for finance teams. I can’t tell if this is a feature of a next-gen bank account (e.g. Mercury, Ramp) or if its a standalone tool. I suspect for legacy companies, a standalone tool is net, more beneficial than replacing their primary bank partner. Which gives companies like uiAGent some interesting channel partnership opportunities. What would a Citi or US Bank be able to do for this kind of company?
4. Nestimate - Retirement Fiduciary as a Service
Nestimate helps financial advisors, employers, and insurance carriers implement and manage defined contribution retirement plans. It helps evaluate, select and manage the different market plans and provide the retirement income planning tools so employees can visualize their future incomes and goals.
🧠Everything is embedded. This is one of those companies that you’d almost never think of existing, and yet if they can crack distribution they take one of the hidden pains away from one of the hidden industries that has massive impact on all of us. I wish them well.
Things to know 👀
Crypto infrastructure startup Zerohash has raised $104 million in funding with backing from financial firms including Morgan Stanley and SoFi, Apollo, and led by Interactive Brokers. Shortly after the announcement, Morgan Stanley confirmed crypto assets will be available in E*Trade powered by Zerohash.
🧠Zerohash helps firms access crypto, tokenize assets and interact with stablecoins. This makes them a versatile player for companies like Apollo who might want to tokenize funds, or Morgan Stanley who might want to offer those, crypto assets or work with stablecoins.
🧠It’s a highly strategic round. Nearly every investor named is a possible future customer.
🧠Big institutions are picking dance partners. They see the revenue Robinhood and Coinbase are generating and want a piece of that pie.
🧠Zerohash survived the winter and has come back strong. There’s a wave of companies that were born just before the 2021 crypto bull run, and had to buckle down through the dark days of the FTX and Terra / Luna collapse.
🧠There will be more crypto winters ahead, but stablecoins and tokenization are more durable trends. As rates come down, the AI bubble pops, risk assets will inevitably fall. But if you’re tokenizing any assets and have institutional clients, you’re on much higher ground.
Tether by market cap would be the worlds second largest bank just behind JP Morgan. Their raise of $20bn at $500bn makes them as large as OpenAI, except instead of losing money, they’re actually profitable.
🧠 Tether is a wildly efficient business. They take your $1, give you 1 USDT, park your dollar at 5%, and keep 100% of the yield. Today’s $172B of Tether issued × 5% = $8.6B annually. With ~200 employees, gives them a margin of nearly 99%.
🧠 Their CEO announced that with the raise, they're expanding into:
AI development
Commodity trading
Energy infrastructure
Communications networks
Media companies
Now you want to cosplay as a tech conglomerate? This is Cyberpunk 2077 isn't it.
🧠 Why pivot to volatile businesses like energy trading and AI?
Rate cuts. Fed cuts crater Treasury yields. So they're diversifying before the golden goose dies
What else do you do with all of that profit? Collect $8.6bn annually with almost no cost, then put it to work and build an empire.
🧠And who's the lead advisor? Cantor Fitzgerald. The firm run by Howard Lutnick. The Commerce Secretary. Who also happens to:
Own 5% of Tether ($600M stake)
Custody over $80B in Treasury securities for Tether
Make tens of millions annually in custody fees
Have a son who interned at Tether in Switzerland
Do with that information what you will.
Good Reads 📚
BNPL will become a feature for every card if a bank chooses to opt-in. Citi, HSBC and Monzo will be early participants, potentially allowing card issuers to compete in BNPL. Behind the scenes they’ll need to use something like Visa’s flex credential to make it all work. The app will also allow conversion of points into dollars and do order tracking if you give it access to your emails. Nice little quality of life bumps.
Tweets of the week 🕊
New newsletter: THE END OF THINKING
Doomsday predictions about AI taking our jobs in the future shift focus away from where it belongs: in present reality, the age of screened tech has coincided with declines in literacy, reading, and writing.
We are deskilling ourselves.
— Derek Thompson (@DKThomp)
1:02 PM • Sep 24, 2025
That's all, folks. 👋
Remember, if you're enjoying this content, please do tell all your fintech friends to check it out and hit the subscribe button :)
Want more? I also run the Tokenized podcast and newsletter.
(1) All content and views expressed here are the authors' personal opinions and do not reflect the views of any of their employers or employees.
(2) All companies or assets mentioned by the author in which the author has a personal and/or financial interest are denoted with a *. None of the above constitutes investment advice, and you should seek independent advice before making any investment decisions.
(3) Any companies mentioned are top of mind and used for illustrative purposes only.
(4) A team of researchers has not rigorously fact-checked this. Please don't take it as gospel—strong opinions weakly held
(5) Citations may be missing, and I’ve done my best to cite, but I will always aim to update and correct the live version where possible. If I cited you and got the referencing wrong, please reach out