First the model, then the rail.

Most weeks in agentic commerce are about a new rail: a network shipping a way for an agent to pay. This week the story sat one layer down. Before an agent can spend on your behalf, it has to know you, and the thing that gives it that knowledge is starting to have a name. Meanwhile the fight over which rail the money settles on stopped being a crypto argument and became a bank one. Here is what landed, and what it looks like from the operations side.

The layer under the rail gets a name

NVIDIA's payments lead set out the case this week for what the company calls payment foundation models: transformer architecture, the approach behind large language models, applied to tabular transaction data to produce embeddings that encode a customer's past, present, and likely next move [1]. The point for agents is direct. An agent can only act for you if it knows you, and those embeddings are where that knowledge comes from, which is why NVIDIA describes payment models and agentic commerce as a flywheel feeding each other [1]. Stripe, Mastercard, and Revolut have each built models of this kind on NVIDIA's platform, and PayPal is rolling agentic features out to 19 million small and medium businesses using open Nemotron models to keep the search step affordable [2].

Meta calls it the next tier

The demand-side framing hardened too. Meta's chief data officer described agentic commerce as the next tier of business and said stablecoins are now assumed inside Meta's own infrastructure, run under wallet abstraction so the customer never sees an address or a gas fee [3]. That is the quiet part of the shift. The settlement layer is being designed to disappear, so the buyer feels a normal checkout while a stablecoin moves underneath it [3].

Wall Street pushes back, and builds its own rail

Not everyone wants that stablecoin underneath the checkout. Bloomberg reported on 13 July that large banks are mounting a coordinated pushback against the stablecoin boom, wary of deposits leaving the banking system for tokens they do not issue [4]. Their answer is not only lobbying. JPMorgan, Citi, Bank of America, and Wells Fargo, working through The Clearing House alongside a long list of other lenders, are building a shared tokenized deposit network aimed at a first-half 2027 launch, where a bank deposit is recorded on a shared ledger rather than a siloed one, carrying the same credit risk and regulatory treatment while settling around the clock [5]. The bet is that if the banks own the tokenized settlement layer, there is less room for a stablecoin issuer, or a retail central bank digital currency, in the institutional stack [5].

Read from the rails

Put the two threads together and the shape is clearer than either headline. The model layer decides whether an agent should spend, and the settlement fight decides where the money lands and who owns the record afterwards. Both are questions an operations team already recognises, because they are the same two questions behind every authorisation and every reconciliation: was this allowed, and can we prove it later.

A payment foundation model is, underneath the branding, an identity-and-behaviour substrate. Ten years in payments operations taught me that the value of knowing a customer shows up not at the sale but at the dispute, when you need to say why a payment looked normal and the model has to survive being questioned. The bank versus stablecoin contest is really a fight over the settlement and dispute path, and whoever owns that owns the repair queue. Invisible rails are pleasant until something breaks, and then someone has to trace a payment the customer never saw through a token the customer never touched. The networks and the banks are building the happy path in public. The exception path is still where this gets decided, and it is still mostly unbuilt.

A payment model decides whether the agent should spend. The settlement layer decides who owns the record when it goes wrong. Both are operations questions wearing new names.