For most of the modern era, a payment system has had to answer three questions. Who is paying. Who is being paid. How do we know the money is good. The plumbing that grew up around those questions (cards, correspondent banking, ACH, SWIFT) assumed a particular shape of payer: usually a person or a long-lived business, with a billing address, a credit history, and the patience to wait two days for a wire to clear.
Two things are breaking that shape at once. Stablecoins finally have the volume and the regulatory cover to behave as a settlement asset, not a speculation. At the same time, AI agents are showing up as a new class of payer: short-lived, programmatic, indifferent to the relationship machinery the rails were built around. Either shift would force a rewrite on its own. Together they change what a payment rail is even for.
Stablecoins as a settlement asset
The stablecoin pitch used to be about access. Move dollars on the weekend. Route around correspondent friction. Get an FX rate that didn't bake in three intermediaries. The pitch in 2026 is settlement. APAC volumes tripled year over year. MAS and HKMA both have frameworks that license stablecoin issuers as payment institutions. Tempo, Solana, and Base are running institutional flow, not retail speculation.
The technology caught up
A decade ago Bitcoin cleared seven transactions per second and Ethereum cleared fifteen. A wire transfer was faster than a stablecoin transfer in any meaningful payment context. A few things changed since then:
- Throughput. Modern payment-grade chains run into the thousands of transactions per second sustained, with peaks an order of magnitude higher. Tempo is built for institutional flow. L2s like Base inherit Ethereum's trust assumptions while clearing at L2 economics.
- Finality. Confirmation windows that ran ten minutes on Bitcoin or a minute on Ethereum dropped to sub-second on the chains payment institutions are actually using.
- Cost. Per-transfer fees fell by three orders of magnitude. The marginal cost of a stablecoin transfer is now well below the marginal cost of an ACH debit or a card authorisation.
The throughput envelope for stablecoin settlement now sits above the throughput needs of most card networks. Capacity is no longer the bottleneck.
The regulators caught up
The 2010s were a decade of prohibition. The 2020s are a decade of licensing. The frameworks now in force:
- MiCA (EU). Regulated stablecoin regime fully in force across the European Union since 2024. E-money tokens and asset-referenced tokens both have a license path.
- GENIUS Act (US). Federal framework for payment stablecoin issuers, signed in 2025. Issuers operate with bank-like oversight and 1:1 reserve requirements.
- MAS and HKMA. Singapore and Hong Kong license stablecoin issuers as payment institutions. The two jurisdictions where most APAC institutional flow originates are now both regulated rails.
What follows from licensing is that the controls a payment institution actually needs (issuer attestations, regulated reserves, KYT, travel rule) live inside the rail rather than being bolted on by a vendor afterwards. Once a settlement asset clears in seconds, costs a fraction of a cent, and satisfies the controls a compliance officer can actually sign off on, the question of which rail the agentic stack runs on stops being interesting.
Agents are a new class of payer
A card network was designed for someone with a name, a billing address, and a thirty-day cycle. An agent has none of that. It's spun up to do a task, calls a few services, and disappears thirty seconds later. There's no credit history behind it, just a budget. The merchant isn't a relationship; it's a quote on a resource page.
This is in production already. Protocols like x402 and MPP turn HTTP 402 into a working handshake: the agent calls a resource, the server returns a quote, the agent settles, the resource clears. The unit of value is the call, not the account. The whole exchange might last 400 milliseconds.
The grain breaks the existing rails. A credit card charge for $0.001 is uneconomic before the network fee. A bank wire is uneconomic by three orders of magnitude. Agentic commerce only closes on rails where the marginal cost of a transfer is below the marginal value of the call being paid for. That's never been true of fiat infrastructure. The stablecoin rail just described is the first one where it can be.
Velocity changes the unit economics
T+2 used to be a settlement window. It was also a working capital problem, a reconciliation problem, and a financing line on the balance sheet. Compressing settlement from days to seconds does more than make a transaction feel fast. It removes whole categories of cost from the business sitting on top of the rail.
A payment institution that settles internally on a stablecoin rail doesn't need to pre-fund counterparty accounts to handle weekend traffic. A merchant clearing at T+0 has no reason to keep a factoring relationship just to bridge cash flow. The treasury team can stop modelling settlement timing as a variable independent of volume.
For agents the effect is sharper. An agent with a dollar of budget making a hundred calls can't afford to wait on any one of them. Without T+0 the workload simply doesn't exist.
What this forces businesses to rebuild
A few architectural decisions stop being optional once an agent shows up at the door:
- Identity at the transaction, not the account. A payer that exists for thirty seconds will not fit a row in a customer table. Identity becomes a property of the transaction (wallet, issuer, admission contract), and the KYC stack has to learn to resolve at that layer.
- Pricing on the resource. A 402 response with a quote replaces the API-key handshake. Without it the agent has no way to know what to pay. With it, the resource itself is the price list.
- Compliance inside the rail. Bolted-on KYT vendors lag by hours and break under agentic call volume. The decision has to resolve before settlement, not after.
- A treasury that thinks in stablecoins. Reserves, hedging, reconciliation: all of it changes shape when the settlement asset is a stablecoin. The finance team has to learn the new instrument the way it learned cards in the 2000s.
The fintechs we work with hit all four of these in their first quarter on the rail. Usually in that order.
What 2026 to 2028 looks like
The next two years won't be about whether stablecoins are a real payment rail. That question is closed. They'll be about which institutions move first, and whether the rails those institutions pick are compliance-native or assembled out of generic onchain SDKs with a vendor stack glued on top.
The first movers tend to find the same things in roughly the same order. Engineering cost is dominated by compliance, not by chain integration. The country-by-country corridor playbook of the last decade doesn't survive contact with rails that are licensed at the rail layer. The reconciliation team that used to need a quarter to close the books needs a week, and most of them are needed somewhere else.
The shift won't feel dramatic from inside any single company. A treasury starts settling a corridor in stablecoins. The pricing on a previously API-key-gated endpoint quietly switches to per-call. A compliance officer signs off on a control that resolves before the chain even accepts the payment. None of those moments looks revolutionary on its own. The cumulative effect is.