Things to know 👀
Stripe sent investors a letter this week, obtained by Axios, announcing OpenRouter as its largest-ever acquisition (more on that below) and declaring that January 1st marked the beginning of the singularity. Its definition is sober: a large inflection in long-run trends, like the rate of new company formation, that it decided to take seriously. The numbers back the swagger. H1 net revenue grew 41% year over year, free cash flow grew 43%, and by year-end more than ten Stripe products will each clear $100M in net revenue.
🧠 Stripe is growing net revenue and FCF almost 2x faster than Adyen YTD. They’ve been able to acquire with stock, while the overall share count has actually decreased over the last 3 years. The stock has also grown 2x faster than the S&P in that time.
🧠 This is reacceleration at a scale payments rarely sees. AI and crypto each more than doubled as a share of Stripe's revenue in a year. 88% of the Forbes AI 50 build on Stripe, and most of the remaining 12% are pre-revenue.
🧠 Stripe's book is the closest thing we have to an income statement for the AI economy, and right now it says the boom is showing up as real revenue.
🧠 New company formation inflected. New platforms are activating on Stripe Connect at twice last year's rate. AI made starting a business cheaper and faster, and people responded by starting more of them.
🧠 “Capital and intelligence are now the two flows running through every business,” and Stripe intends to run the rails for both. The letter argues that "intelligence is special: it is expensive, heterogeneous, and constantly changing," so businesses must reason about the cost and return of every unit of it, the way they already do with capital.
🧠 I am absolutely convinced the next decade of finance will be about how we finance the intelligence economy, the AI build-out, and the tokens it generates.
🧠 Then there’s OpenRouter
Stripe has signed a deal to buy OpenRouter for more than $7bn, per Bloomberg. If you’ve never used it, OpenRouter is an API that provides access to more than 400 AI models. Millions of developers use it to compare prices, switch providers, and route each request to the cheapest model capable of doing the job. It’s the place you go to use things like DeepSeek or Qwen as soon as they’re released.
In July 2026, the WSJ reported talks at around $10bn. The signed deal comes in above $7bn, still more than 5x the $1.3bn OpenRouter was valued at when it raised in May 2026, three months ago.
🧠 Stripe now has a 7.5% take rate on AI, instead of 2.5%. Stripe acquired Metronome, which helps most AI companies (like OpenRouter) bill for AI usage by the token and then collect payments. OpenRouter then meters each individual AI and adds a 5% margin over the top of the supplier (Neocloud’s) price.
🧠 Stripe already collects OpenRouter’s payments. OpenRouter has run its billing on Stripe since January 2026, so Stripe could see the target’s revenue in its own dashboard while it negotiated. It knocked roughly $3bn off the leaked price. At a reported ~$140m in annualized revenue, the leak implied a 70x multiple. The signed price is closer to 50x.
🧠 For AI companies, tokens are usually the biggest input cost. OpenRouter is where that spend gets managed. It routes each request to the best model for the job and takes a small percentage of the token flow as its fee, the same way Stripe takes a small percentage of payment flow.
🧠 Stripe now has both sides of the P&L for AI companies. Stripe Payments collects an AI company’s revenue (money in). Metronome, bought in December 2025 for a reported ~$1bn, meters the usage that drives that revenue (money in). OpenRouter routes the tokens all of it runs on (money out).
🧠 Patrick Collison called metered pricing “the native business model for the AI era” when he bought Metronome. Eight months later, he’s buying the marketplace where the token metering happens. Stripe built a $159bn company taking basis points on payment flows. OpenRouter takes a cut of token flows.
🧠 OpenRouter’s first funding round was less than two years ago. One breakdown on X puts founder proceeds near $3.8bn, with CapitalG’s Series B, less than four months old, set to return more than 5x on $113.5m deployed. For a16z it’s the second 100x exit in a week, after Cursor’s acquisition closed.
🧠 OpenRouter also sells inference. New data from Brex showed start-ups are increasingly buying their own inference for their AI usage. There’s a progression from buying AI from a lab, to using token routing, to buying inference. The median time from signing with an AI lab to renting your inference is now 5 months.
🧠 OpenRouter also sells that inference via MPP. The Machine Payments Protocol means your AI agent can buy tokens and inference via stablecoins. And with that, the Neocloud can get paid for that inference instantly.
🧠 If agents are going to transact with each other, what router would they use to buy all of those tools? Maybe an OpenRouter.
Hours after Stripe’s announcement, Ramp made Router.com generally available: one endpoint, every request routed to the cheapest model that clears your quality bar, free through 2026. Ramp says early users cut inference costs 40% on average.
🧠 Two of the most innovative companies in finance converged on the same idea. Financing the intelligence economy is the frontier and where growth will come from. There’s a massive lesson here for everyone in the industry.
🧠 The router is the purchasing decision. Every request that passes through one picks which model gets paid, at what price, for which job. Whoever owns routing owns model procurement.
🧠 Corporate AI spend has grown 20.7x since June 2025, per Ramp's own AI Index. Tokens went from rounding error to the fastest-growing line item on the P&L in 14 months.
🧠 Stripe is buying the flow. Tokens are the newest money in motion, and Stripe takes a cut of money in motion.
🧠 Ramp is selling the savings. Free through 2026 is the tell: the router is the loss leader, spend visibility and control are the product.
🧠 Ramp has spent seven years doing one thing: saving companies money. Tokens are a spend category like cards and SaaS before them, and Ramp's wedge fits the moment better than anyone's. That's the one I'd bet on.
🧠 Pricing power in AI just moved from whoever builds the best model to whoever decides which model gets the job most efficiently.
At the CFTC's first Innovation Advisory Committee meeting, Chairman Michael Selig laid out three fronts. Crypto: the agency is exploring a new class of Designated Contract Market (DCM) called a “crypto asset market,” giving registered and unregistered crypto exchanges a route to offer leveraged and margined trading under CFTC oversight, plus a legal pathway for onchain protocol developers.
Compute: The CFTC asked for comment on compute derivatives this week, including perpetual futures. Prediction markets: new rules are coming on retail protections, product governance, and incentives. The Senate's first procedural vote on CLARITY is September 15. If Congress stalls, Selig intends to propose the crypto rules anyway. They’d then be subject to a comment period.
🧠 We’re finally getting consumer protections! Selig conceded there are gaps. His answer is to write those protections into Parts 38 and 40 of the CFTC core principles every DCM has to meet. Part 38 covers retail protection requirements, and 40 the rules for listing contracts and limits on incentive programs
🧠 Prediction Markets, Perps and Compute forward curves will run on the same license. Kalshi is a DCM. Coinbase Derivatives is a DCM. Bitnomial used a DCM to launch leveraged spot crypto. Compute Derivatives would trade on DCMs. The DCM is becoming the wrapper for markets that sit outside securities law and offer margin trading.
🧠 Take the politics out of it, and this seems like a sensible path. Congress defined "commodity" in 1974 to include "services, rights, and interests," and Selig is reading it exactly as written. He cited a 1978 Senate report saying whether a market performs hedging or price discovery "should not be the determining factor" in CFTC jurisdiction.
🧠 Play it forward: We could see Hyperliquid-style perp venues onshore under a US derivatives license, GPU “forward curves” lenders can underwrite data center debt against, and sports contracts under one federal rulebook across all 50 states. Again, I like this. I’d rather have clear rules than have everything offshore.
🧠 Where the SEC has Howey, the CFTC has "rights and interests." That may turn out to be the more expansive phrase.
4 Companies 💸
1. Megawatt - Onchain yeld from solar and battery projects
Megawatt takes stablecoin deposits and uses them to build battery storage sites in Southeast Europe. The batteries earn three ways: charging when power is cheap and discharging into the evening peak, getting paid to hold the grid at 50Hz, and collecting capacity payments for sitting on standby. Revenue flows back on-chain, and depositors claim it daily.
🧠 Grid battery storage is a great asset to tokenize because the cash flows are metered, so you can pay investors what the battery actually earned instead of what a manager reports. And 2.4MW is exactly the size banks won't finance, so the capital gap is real. It’s also something we do not have nearly enough of. I just wonder about distribution. How does this project show up in every brokerage app?
2. Ellis - The “book of record” for private credit funds
Ellis pulls data from the systems a private credit fund already runs (fund admin, general ledger, loan accounting, banks) along with the loan tapes and compliance certificates that arrive as spreadsheets and PDFs, and reconciles it all into one book. AI agents then run the close, surface reconciliation breaks with a proposed fix attached, draft reports and forecast cash. Every number traces back to the row, cell, or GL account it came from.
🧠 Ellis isn’t the system of record here; the fund administrator still owns the actual book. Companies like Alter Domus, Citco, and Allvue are still the legal golden source. But the useful bit for finance teams is that every number traces back to the cell it came from, and that's the version of AI a finance team will actually sign off on. Private credit ops really is held together with spreadsheets and emailed PDFs. AI that understands this is super helpful. My question is, can’t the main models already do this quite well?
3. Henry - Deal prep for commercial real estate brokers
Henry lets a broker forward a deal to an email address, or paste in a rent roll, and comes back with an Excel underwriting model built on that firm's own pricing assumptions, a ranked set of comps from its past transactions, and a branded pitch deck. It also builds the buyer list, ordered by fit and by how well the firm already knows them. Edits sync between the model, the writeup, the deck and the data room.
🧠 This feels like the co-worker you forward a deal to and get back a pitch deck. And I imagine a lot of these brokers are small enough that they’re happy to let an AI tool in without worrying about their secret sauce leaking.
4. Soter - Crime and slashing cover for crypto institutions
Soter writes insurance for crypto exchanges, custodians and funds for crime, D&O and professional indemnity for the people running the firm. It also offers cover for slashing penalties and smart contract failure. It issued what it says is the first BTC-denominated crime policy, so a tier-one exchange's payout arrives in the asset it lost rather than USD.
🧠 The website says 41% of institutions now have crypto exposure. The problem with digital assets is they have net new risks that existing policies might not cover. Slashing is a great example; you get to see it on-chain, which makes it one of the few crypto risks you can underwrite and have a clean evidence chain for.
Good Reads 📚
Card transactions feel instant for the consumer, and they’re a modern marvel of network technology. But for merchants, banks, and everyone else in the payments chain, they can still be slow and expensive, especially where cross-border is involved. That’s because they run on correspondent banking networks and settlement services built in the 70s.
These settlement services don’t work weekends, but cards do. To cover those potential transactions, card issuers set aside large sums of money (up to their maximum daily volume) x the number of days the settlement networks are closed. On a public holiday, this could mean 4x the usual daily settlement volume. If that gets settled instantly, instead of being held just in case it's needed, it's funding growth, payroll, or product development. And stablecoins offer daily settlement with card networks.
🧠 This is the real killer app of stablecoin cards. Instant settlement. People are only just realizing this, but it will be 100x bigger in a years time.
Tweets of the week 🕊
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.
(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
