Unfiltered: the metering bet in fintech

In the August column for The Paypers, Dwayne Gefferie, Payments Strategist, analyses the latest fintech acquisitions of AI companies and why metered pricing is currently an uncertain strategy.

In the last few editions of this column, I’ve discussed how the data intelligence gap is growing, that sovereignty is now the biggest focus for European companies, and how reduced interchange fees are prompting payments companies to find margins too. 

Stripe's largest acquisition in its entire history is, surprisingly, not a payments company. 

On 19 August, Stripe told investors that OpenRouter is joining its business, calling this move the biggest deal it has ever done. Bloomberg had estimated the price above USD 7 billion two days earlier, while Axios reported the deal is north of USD 8 billion, mostly in stock. OpenRouter routes API calls across more than 400 AI models for around 8 million users, and it raised at a USD 1.3 billion valuation about 82 days earlier, so the price increased fivefold in under three months. 

Although most coverage said this is Stripe buying its way into AI, I think that's backwards. Stripe paid a record price to defend the only thing it has ever sold: the meter sitting where value occurs, a meter that has now moved. 

Payments spent thirty years perfecting a fee on a discrete event (e.g., a single financial action or data exchange): a purchase happens, a fee is taken, and that’s it. AI-native businesses don't have these discrete events; they deal with continuous consumption, which turns the commercial question from ‘did this authorise’ into ‘who tracked and measured the actions/transactions?’ 

To understand fintech deals, you must follow the cheques

Stripe's own announcement says the company now thinks about a developer's intelligence workflow the way it does about a revenue one. It groups OpenRouter with Bridge, Privy, and Metronome (the usage-billing engine Stripe acquired in January 2026), and says Radar, built to catch card fraud, is now working on token fraud. 

When Metronome’s acquisition went public, Patrick Collison, Stripe’s CEO, was blunter. He said metered pricing is the native business model for the AI era, and considers the change at least as big as the arrival of SaaS. Metronome already bills OpenAI, Anthropic, and NVIDIA, so with its acquisition, Stripe first bought the invoice, then the revenue pipeline it measures. 

Adyen had a similar, less publicised approach. On 1 July, it closed on Talon.One and Orb, the first two acquisitions in a 20-year history, a roughly USD 1.2 billion expense in one go. Orb is usage-based billing, and Adyen's decision lies in the industry shift to consumption pricing, driven partly by AI. On the H1 call, company management said Orb lets them pick up early-stage AI-native companies. 

Another aspect almost nobody wrote about is that Adyen is pulling capital spending from 2027 into H2 2026 to secure compute and storage, and lock in pricing. This takes full-year capex to about 7% of net revenue from 5% in H1 2026. The interim CFO called this a deliberate move, amid unprecedented demand. Now queuing for GPUs behind companies committing USD 725 billion to infrastructure in 2026, a payment processor’s cost base is starting to resemble a cloud provider's. 

Why fintech is changing from money movement to measurement

Card economics assume a behaviour: someone taps, a message travels, a fee applies to the basket, and interchange, scheme fees, the acquirer's markup – all depend on this behaviour. If you remove the behaviour, the model fails to apply. 

Moving money when an agent calls a model 400 times an hour, when a customer’s bill is a function of tokens, or when a platform charges per seat, API call, and gigabyte on a single invoice is trivial. The challenge with these use cases is determining what's owed. Pricing power belongs to those addressing the challenge, which is why serious buyers sought out measurement first. Interchange has always been the intermediary between two parties who needed to agree on what happened. 

Stripe stated OpenRouter's token consumption is growing at 9% a week in 2026, a figure they provided, not an audit. If you roughly take this at face value, it’s pretty clear why a billing company acquired a router at a record price. 

Merchants didn't ask for these payment developments

Now, the part I'd rather say here than on a stage is that almost none of what’s happening serves the merchants who pay the fintech industry's bills. 

Can you show me the restaurant owner, or the hotel group, the pharmacy chain, the football stadium, the tyre shop asking for token-level billing? They have baskets and staff who need the terminal to work on a Saturday night. Their payments-related needs centre on authorisation rates, chargebacks, and the cost of premium card acceptance, which is the case since 2015. 

Volume data backs this up:

• Adyen's in-person volume grew 28% in H1 against 23% online. 
• Block's US Square volume hit its strongest rate since 2023, with food and beverage up 20%. 
• Global Payments is getting more merchant locations out of each salesperson.
• Shift4 spent Q3 2026 signing up hotels, jewellers, and waterparks. 

Although the commerce economy is growing, the fintech industry is targeting a customer segment with enormous valuation and a small transaction count. 

Agentic AI numbers make things worse. In the UK, 58% of online merchants believe AI agents have already arrived on their platforms, with 3% of transactions involving one. Of the top hundred retailers, 15% say their payment stack is ready for agents, and half are spending on agentic AI regardless. 

OpenAI switched off native checkout inside ChatGPT in March and sent buyers back to merchant sites, and while this market move said something, the industry answered by forming another standards body. 

Can the revenue meter survive a bad year?

Interchange is boring because it's tied to people buying groceries. However, boring has carried it through multiple recessions, a pandemic, 21 years of litigation, and a European price cap. 

Metering revenue depends on AI capital spending, which increasingly runs on debt and is the collateral for these developments. In 2026, Morgan Stanley has global AI-related debt issuance on track for about USD 570 billion (more than double from 2025) and sees a USD 1.5 trillion gap between projected data centre spend through 2028 and what hyperscalers can fund from operating cash. 

As AI inference gets cheaper, the same task requires fewer units each year. When AI budgets normalise, usage-based revenue drops faster than subscription revenue, because it has no baseline. The pricing model that scales beautifully can’t protect the business if usage or value decreases, which is the part the usage-pricing evangelists never address. 

Competition is another worrisome aspect. Foundation model providers continue shipping the features that made routers valuable in the first place: automatic fallback, cost-optimised model selection, and multi-model access inside a single API. So, Stripe might have paid its largest-ever amount for a layer with 18 months of independent life. A CNBC July investigation also found that Chinese-origin models account for 46% of US enterprise token usage on OpenRouter, so a US payments company now has a traffic pattern with a political dimension it didn't have in Q2 2026. 

I still think Stripe and Adyen are directionally right about where the fee goes, but they're a cycle early, and merchants are subsidising the wait. 

The AI gap for fintechs: billing, not intelligence

The firms that everyone says are behind on AI aren't behind on models, but on what they can charge for. 

On 6 August, Fiserv reported a 5% decrease in organic revenue and adjusted EPS down 26%. It cut full-year EPS guidance from USD 8 to between USD 7.20 and 7.40, and saw shares fall over 11% premarket. Worldline cut its revenue outlook in the same period, pointing to the introduction of generative AI across the organisation. Global Payments GAAP revenue increased 68.6% with the Worldpay acquisition, with diluted EPS of USD 0.05, and a H1 2026 net loss of USD 1.79 billion. 

These companies’ revenue depends on somebody else's card volume at a moment when every posted price is being handed to distribution. Wise cut its cross-border rate to 50 basis points, the lowest it has ever charged, and plans one or two more cuts. PayPal's take rate fell 7 basis points, to 1.61%. Mastercard paid out USD 6.0 billion in rebates and incentives in Q3 2026, up 22%, against 8% currency-neutral growth in the payment network itself. Nobody wins that by charging less per swipe. 

What I'm inferring

This is my personal belief here, not a fact. When Visa announced record volume and 2,600 job cuts (about 7% of staff) the same day, USD 563 million in severance landed in its Q3 2026 earnings. Block cut over 4,000 people in February, and the stock rose 24%, while Mastercard trimmed 4%. 

In part, all three justified their move as a consequence of AI. Block has proof: code changes per engineer increased 150% since January, but Mizuho's Dan Dolev told the Wall Street Journal that Block's cuts probably weren't AI at all. 

I think both things are true. While AI is doing real work inside these companies, AI is also the most flattering explanation available for a cost base attached to a business model currently repriced publicly. As the industry pays well for this explanation, expect to keep hearing it. 

The market is spending its scarcest capital on a customer that doesn't exist at scale yet, funded by a boom with no end date yet in sight. Meanwhile, the merchants generating the volume keep asking for the same three things they asked for a decade ago: authorisation rates, chargebacks, and the cost of premium card acceptance. 

About the author

Dwayne Gefferie is a payments strategist, data scientist, and advisor with over 22 years of experience in the global payments industry. Unfiltered is his monthly column for The Paypers.