👋 Welcome back to another edition of Fintech Radar!
If you’re new, here is a breakdown of what you can expect from each issue.
If you missed our recent editions, you can catch up here. Some previous issues you might want to check out if you’re new include “A Deep Dive Into The Cash App’s Growth Machine”, “The Future Of Payment Initiation”, and “Current: Doing It Differently”.
Your ad could be HERE!
Fintech Radar is a must-read for fintech founders, operators, and investors. If that’s your target audience, placing an ad right HERE is a cost-effective way to reach them.
If this sounds like a good fit for your brand, head to our sponsorship page for more details and to secure your advertising slot. Prices start at $100 per issue!
If you have any questions, reply to this email and ask away!
🤖 Anthropic wants to build its own payments stack, The Information
🏃 The Rundown: The Information reported last week that Anthropic is working out how much of its own billing, payments and fraud infrastructure to build rather than buy. Job listings on the Billing Platform team cover pricing primitives, payments infrastructure, entitlements, revenue pipelines and reconciliation, and talk about evaluating build vs buy while leaning heavily on providers such as Stripe. Other roles cover real-time risk scoring and promo code abuse, and the hiring runs into treasury, where Anthropic wants to consolidate its banking relationships and run cash management on its own software. An Anthropic spokesperson said Stripe “has been a strong partner to Anthropic for years, and we continue to work with them across our business.”
🥡 Takeaway: Anthropic’s run rate went from roughly $9 billion at the end of last year to $47 billion in May and past $65 billion by the end of July, on Bloomberg’s reporting. At that size, metered token consumption priced differently per model, cache reads and writes, rate limits, entitlements and enterprise agreements with committed spend stop being a configuration problem and become worth building real in-house muscle around. Each of those is a place where a rounding error could compound into a big number, and where the pricing logic sits close enough to the product that waiting on a vendor’s roadmap carries its own cost. It’s a classic story of build vs buy.
There is also a decent chance the metered billing infrastructure underneath them is starting to feel the strain of the insane growth. Anthropic’s usage billing runs on Metronome, which processes billions of usage events a month for OpenAI, Anthropic, Databricks and NVIDIA. Stripe bought Metronome for a reported $1 billion, closing in January. Adyen then paid $335 million for Orb, the equivalent engine behind Vercel, Replit and Supabase, and Salesforce took m3ter in between. Inside six months every independent usage-metering specialist of any size was bought by a company that earns on the flow it measures, which leaves Lago and a few open-source projects as the remaining alternatives. Meaning the acquisition route is closing too. Anyone who wanted to buy their way to owning the meter has just watched the targets come off the board, which obviously means build is the only option if you want to get closer to the payments metal.
OpenAI has likewise started to play with its payment routing. It brought Adyen in to process ChatGPT subscription sales while the metered token billing stayed put. The commodity half went out to tender and the lower-cost provider prevailed, while the more complex stuff stayed with Stripe. Subscriptions are a solved problem any large acquirer will price aggressively to win. Metering token spend across models and contract structures is not, especially at the scale and increasing level of complexity the labs are starting to see.
It’s super interesting actually reflecting on what Stripe’s last two years have looked like in terms of acquisitions and how that’s positioned them in the token economy. It bought the meter the labs bill on, then paid seven billion dollars and change for OpenRouter and the routing layer where token spend aggregates across hundreds of models. I (like many others) was sceptical of that price in August and I still think developer defaults are more portable than the price implies. But the logic underneath it holds a little more water now. If the labs take more control of their own payments, the direct processing relationship is the first thing that’ll go, and Stripe has spent a great deal of money making sure it still holds a position in the token economy when that happens.
The thing fintech outsiders always ask is why a company with a line item this size does not simply bring payments in-house, and the honest answer is nearly always that Stripe is better at it than you are and the difference is not worth the headcount. The part people say less often is that most 10x engineers do not join somewhere like Anthropic to go and work on billing. Payments is a career cul-de-sac in a building full of research problems, so it never gets the good people. This might be one of the rare cases where the depth and the complexity genuinely warrant taking more of the flow yourself, and that taken with Anthropic being the ultimate software factory means this might be the moment the assumption changes. These are still only job ads, and plenty of build-versus-buy reviews end in keep buying. If Anthropic does go ahead though, how far they take it is the part I’ll be keenly watching.
💳 India readies agentic payments on UPI, Reuters
🏃 The Rundown: Reuters reported last week that India is preparing a framework letting AI agents make small UPI payments without the account holder approving each transaction. The Unified Agent Protocol is expected to be unveiled at the Global Fintech Fest in Mumbai and would build on two mechanisms already in UPI: UPI Circle, which lets a primary account holder delegate payment authority to a secondary user, and Reserve Pay, which lets customers block funds for multiple debits. Banks currently cap those blocks at 10,000 rupees for up to 90 days, and the sources said that limit may be revisited for agentic use. NPCI is expected to give merchants direct integration, with customer-set spending rules, audit trails and identity checks, and plans a liability framework. NPCI has not confirmed any of it. UPI handled 24.51 billion transactions in August.
🥡 Takeaway: Every agentic payments framework that has shipped so far has been built by a company. Visa has one, Mastercard has one, Stripe and OpenAI wrote the Agentic Commerce Protocol between them, Google has AP2, and Rain assembled an alliance around its own. India would be doing it at the level of the rail itself, which is a completely different proposition. A protocol that arrives as national infrastructure is not a product decision somebody can reverse next quarter, and every bank and merchant already connected to UPI inherits it whether or not they had a view. For better or for worse.
It also quietly answers a question a lot of people have been stuck on. Every bank, merchant and fintech in India has spent the past year being asked what their agentic payments strategy is, and from here the answer is that the rail has one and you will be plugged into it. There is still an enormous amount of connective work between a protocol and an actual consumer transaction, and a layer of providers will make a good living doing exactly that. The difference is that it stops being a strategy question and becomes an integration project, which is a much better problem to have.
The other thing it does is route around everybody currently trying to steer agentic commerce from the outside. The alliances, the consortiums and the protocol working groups are all trying to become the standard by getting enough of the industry to agree with them, which is slow, political, and historically ends in three competing standards and a decade of adapters. NPCI does not have to win that argument. It runs the rail, and setting the default inside the world’s largest real-time payments network is a materially different lever to pull than convincing a room full of competitors to sign something.
Liability is the unsolved problem, and NPCI committing to write a framework for it goes further than any of the private schemes have. Card networks spent sixty years building chargeback law, and the whole edifice rests on a human having authorised something. When an agent buys the wrong thing at the wrong price, working out whether that lands on the customer, the bank, the merchant or whoever wrote the agent is genuinely unresolved everywhere in the world right now. The detail is not public yet, so I would not get carried away, but somebody at least has it on the list.
The delegation model is elegant and also a slight fudge. UPI Circle was designed for a son paying for his mother’s groceries, and reusing a human-delegation primitive for software is a clever way to avoid building something new. The fraud profile is not the same though. A human delegate has a phone, a face and a bank relationship. An agent has an API key and can be running in a data centre in another country. The 10,000 rupee cap on blocked funds, about $105, is doing a lot of load-bearing work in the meantime, and it is the right place to start: groceries, utilities, top-ups, the stuff where being wrong costs somebody a few dollars.
NPCI has form here too. The Razorpay and OpenAI pilot last October already put UPI payments inside ChatGPT using Reserve Pay, with Axis Bank on the banking side and BigBasket as the merchant, and the Unified Agent Protocol reads like the generalised version of that pilot. On scale, if even one percent of August’s volume went agentic you would be looking at roughly 245 million agent-initiated transactions a month, which is more than most Western markets do in total across all payment types. Set against that, I would still hold some caution: a protocol unveiled at a conference and a protocol that every bank in India actually supports are separated by a couple of years of integration work.
🏦 TabaPay raises $155M and buys itself a bank, The Paypers
🏃 The Rundown: TabaPay announced last week that it has closed $155 million in strategic growth financing led by FTV Capital, combining primary capital with a secondary transaction, and that it intends to acquire Transact Bank, an OCC-chartered and FDIC-insured bank based in Denver. Transact will be renamed TabaBank and sit alongside TabaPay under a newly formed holding company, TabaHoldings. The deal is expected to close in the fourth quarter subject to regulatory approval, which means both the OCC and the Federal Reserve. TabaPay says it will keep working with its existing network of more than 20 partner banks across the US and Canada.
🥡 Takeaway: Five weeks ago I wrote about Increase buying Twin City Bank. Now TabaPay is buying Transact, funded by a growth round assembled largely for that purpose. Both of those are the acquisition route, and what makes it interesting is that the other route opened up at the same time. Not that long ago a fintech looking at either one was staring at a very long and arduous climb.
The OCC’s own numbers on the de novo side are genuinely startling. It has taken 40 charter applications in the past 18 months, against 48 across the whole of the 14 years from 2011 to 2024, three of which produced none at all. Twenty-one of the 40 have been approved. On the agency’s own account a full-service national bank received final approval and opened its doors for the first time in five years, which was Erebor, the Palmer Luckey and Founders Fund vehicle. Revolut cleared the OCC on Thursday and now waits on the FDIC and the Fed. Comptroller Jonathan Gould has spent the year saying in public that America and the OCC are open for business, and the file backs him up.
Wise and bunq being knocked back looks like counter-evidence and mostly is not. The OCC turned them down over capital documentation, management experience in the product being underwritten and supervisory record, which is what it has asked of anyone applying for decades. The door is open, the bar sits about where it always sat, and decisions are landing inside 120 days rather than disappearing into a drawer for three years. That turns the charter question into a real strategic choice rather than taking whatever you can get. Buying gets you a going concern with a date on it, along with the BSA and AML programme somebody else designed and a core system somebody chose in a decade you would rather not think about. Applying gets you a clean build and a wait (which to be fair is getting shorter).
The line about continuing to work with more than 20 partner banks is the diplomatic version of a slightly awkward conversation. Every sponsor bank in that network read a press release last week announcing that their customer is becoming a competitor. That may be entirely fine, since most of those relationships exist for specific programme and geographic reasons rather than sentiment, but it is a dynamic worth watching in a market where sponsor banks have been getting choosier about who they work with.
I would not treat Q4 as a date so much as an aspiration. The OCC piece is a change of control on an existing charter, which is the more routine half. Standing up TabaHoldings as a bank holding company puts the Federal Reserve in the middle of it, and the Fed has its own view about payments companies owning banks that it has not been shy about historically. It’s probably sensible to assume this slips.
💸 X moves US creator payouts off Stripe, TechCrunch
🏃 The Rundown: X said last week that all US creator payouts, covering both the Original Content Rewards Program and creator subscriptions, will be paid through X Money rather than Stripe. Payouts are instant, replacing a fortnightly cycle with a $30 minimum threshold. X told TechCrunch the change is required for creators in the US, while creators elsewhere stay on Stripe. Users need to be Premium or Premium Plus subscribers and live in a state where X Payments LLC holds the necessary licences. X Money launched at the end of July and bundles a bank account, peer-to-peer transfers and a Visa debit card, with interest and cashback on balances. The old Creator Revenue Sharing programme closes on 7 September, with Original Content Rewards opening the day after.
🥡 Takeaway: The payout is the acquisition channel, and it is the cheapest one in consumer fintech. Nobody wakes up wanting another bank account. They open one because that is where money they are already owed happens to be sitting. X has a captive population of US creators with earnings waiting for them and one route to collecting it, which converts at a rate no amount of paid marketing gets close to. eBay did exactly this to its sellers with PayPal in the early 2000s, and it worked well enough that eBay went and bought the company.
Instant payouts sound like Elon is being a generous billionaire, but they cost X almost nothing, because the money lands inside an X Money balance rather than in the creator’s bank. Then the interest and cashback features go to work keeping it there. Paying somebody the moment they earn it, into an account you operate, is a very cheap way to buy a deposit base, and deposits are the thing that makes the rest of a consumer finance product possible.
Making it compulsory is where this might earn the ire of creators. There is no opt-out, and X confirmed to TechCrunch that the change is required for creators in the US, so anyone perfectly happy with a Stripe transfer into their existing bank now has an X Money account whether they wanted a new financial relationship with X Money or not. Musk has been promising the everything app for three years and has arrived with the everything app making it mandatory.
State-by-state licensing is what might slow this down for Musk. X Payments LLC holds money transmitter licences in 41 states plus the District of Columbia, where full US coverage runs to 49 states plus DC, Montana being the one state with no money transmission regime at all. So payouts have been made compulsory through a product X cannot yet offer everywhere it has creators. New York is the sharpest version of the problem. X told customers there that their rate drops to 0.00% from 1 October after the Department of Financial Services informed it that it is not currently permitted to pay interest on stored-value accounts in the state, and it is offering a $300 bonus instead, which X’s own FAQ is careful to say is not interest. Getting to full coverage is a multi-year administrative grind that a lot of companies have underestimated.
⚙️ Deutsche Bank rips out fifteen core systems, FF News
🏃 The Rundown: Deutsche Bank signed a global licensing agreement with Thought Machine last week, adopting the Vault Core platform as the technology foundation for its Private Bank’s retail modernisation in Germany. The bank runs 15 core banking systems today and plans to consolidate onto two by 2028, with Vault Core the first of them. Development is under way, testing is planned towards the end of this year, and products migrate progressively from 2027 with old and new systems running in parallel through the transition. GFT Technologies is the system integration partner. The Private Bank plans to invest around €600 million in technology, operations and AI by 2028 and expects roughly €300 million in annual run-rate savings. Thought Machine says Deutsche Bank takes it to 20 tier-one banks.
🥡 Takeaway: Fifteen core systems down to two is the sort of sentence that reads tidily in a press release and consumes most of a decade in practice. Core migrations are where bank transformation programmes go to die, and the industry has enough cautionary tales that any board approving one is doing so with its eyes open. Deutsche Bank has structured this about as sensibly as you can: parallel running rather than a big-bang cutover, a controlled German rollout first, and products moving across progressively from 2027 instead of over a weekend.
On the numbers, €600 million spent to save €300 million a year gets you a payback inside roughly two years once you are through the other side. That is why boards sign these off. The €600 million is the number that is known though, and the €300 million is the number that is hoped for, and the gap between those two is where these programmes usually get “interesting”.
For the vendor market, Thought Machine reaching 20 tier-one banks matters more than this single deal. The standard answer from a large bank for the past decade has been that cloud-native cores are fine for challengers but their own estate is too complex and too systemic to touch. Deutsche Bank is about as systemically important as European banking gets, and it has just said otherwise in public. That makes the conversation harder for Temenos, FIS and Fiserv, and it makes it easier for every CIO who has been trying to get the same business case through their own investment committee.
💰 Felix raises $200M to move beyond remittances, Bloomberg
🏃 The Rundown: Felix announced $200 million last week, split between $87 million of equity led by Andreessen Horowitz and a $113 million credit facility from General Catalyst’s Customer Value Fund. QED Investors, Castle Island Ventures, Switch Ventures, Contour Venture Partners and Endeavor Catalyst also took part. The Miami company says the round values it at roughly three times its last one. It moves money from the US to Latin America through WhatsApp, settling in USDC.
🥡 Takeaway: The interface is the product. There is no app to download and no onboarding funnel to abandon, because the transfer happens as a message in a WhatsApp thread the family is already using. In the corridor Felix serves, that removes the genuinely hard part of remittances, which was never the FX or the settlement. It was getting somebody’s uncle to install a financial app and complete identity verification on a phone with 4% battery.
The stablecoin leg is invisible, which is the only version of consumer stablecoins that has worked anywhere. Nobody sending $300 to Guatemala knows that USDC is involved and none of them would care if you told them. Compare that with three years of consumer crypto wallets asking people to understand what they were holding, and the lesson is fairly clear about where the technology belongs.
Setting aside $113 million to lend against says the plan is to stop taking a cut of a transfer and start holding assets, because remittance margins compress the moment a competitor turns up and lending margins behave differently. Underwriting a population with thin US credit files is a considerably harder business than moving their money, and it is the thing that will decide whether the tripled valuation makes sense in the long run (read as the next funding round).
💰 Apate.AI raises $8.15M to waste scammers’ time, FinTech Global
🏃 The Rundown: Sydney-based Apate.AI announced an oversubscribed $8.15 million seed round last week led by Lobby Capital, with OIF Ventures, Investible, Concept Ventures and Baobab Ventures joining. The company builds conversational AI agents that pose as victims, engage scammers across voice and text, and extract intelligence from the conversation. Commonwealth Bank is a customer, and Apate says it has run more than 2.5 million autonomous conversations with threat actors, producing over 250,000 pieces of intelligence.
🥡 Takeaway: Dali Kaafar’s research at Macquarie University turned into a fleet of bots that string scammers along with wrong card numbers and trouble logging into online banking, and some of those conversations have reportedly run for two hours. Every hour a scammer spends on a bot is an hour they are not spending on somebody’s grandmother, which beats detecting the same scam after the money has gone.
A seed-stage company with Commonwealth Bank in production is unusual. Getting through a tier-one bank’s procurement, security review and vendor risk process is brutal, and most companies this size are still doing unpaid proofs of concept by the time they get there. Apate is well past it, with more than 2.5 million autonomous conversations run and 250,000 pieces of intelligence handed back into customers’ disruption processes. That reference is worth more than the $8.15 million, because it is the thing that makes the next ten bank conversations possible.
The segment underneath it is getting more important too. AI has moved the economics decisively towards the attacker, since a voice clone now needs seconds of audio, and the American Bankers Association put voice fraud in banking up around 30% in 2025, with roughly one in every 750 calls into bank contact centres flagged as potentially fraudulent. Detection on its own struggles against volume that is close to free to produce. The thing that does not scale for a scammer is time spent on a live conversation, which is exactly the cost Apate goes after. Add Australia’s Scam Prevention Framework, the UK reimbursement rules and PSD3 in Europe all moving the bill onto the institution rather than the victim, and this is a category that did not exist as a purchasable item three years ago.
🎧 The $100B Niches Hiding Inside Payments, The a16z Show (September 3)
Max Levchin joins Alex Rampell and Erik Torenberg to walk through twenty-five years of payments, from the Affirm origin story and the pajama problem to why he thinks agents will end up choosing how you pay rather than what you buy.
🎧 What is Missing From Instant Bank Payments, Fintech One-on-One (September 3)
Arpit Goel talks to Peter Renton about what still does not work in instant bank payments and why the merchants who should want them most have been slow to move. Add this one to your playlist.

❤️ Show Some Love For FR
📧 Feel free to reach out if you want to connect. I’m @alantsen on Twitter, or you can D.M. me directly by clicking the button below ↴
Ps. If you like what I’m doing with FR, please share it on your social disinformation network of choice. I’d also appreciate it if you forwarded this newsletter to a friend who might enjoy it.
🙏 What did you think of this week’s issue of FR?
I Love It! ◌ I Like It ◌ Not Bad ◌ I Don’t Like It ◌ It’s Awful





The Apate.AI stat on wasting scammers' time for two hours is hilarious! Turning the economics upside down by making the fraudster burn human/bot hours is way more effective than just flagging the transaction after the fact 👏👏👏