A card payment feels instantaneous and is not. The authorisation a cardholder experiences takes a second; the settlement between issuer, network and acquirer takes a day or more, and against that delay sit prefunded balances doing nothing. That idle balance is the cost, and it is what a settlement project is buying back. Everything else — the interface, the branding, the token — is downstream.
Why settlement is the expensive half
Between the moment a payment is agreed and the moment value moves, somebody is carrying exposure and somebody is holding capital against it. Lengthen that window and the capital grows; shorten it and the capital is released. That is the whole economic argument, and it is why these announcements read as plumbing and are priced as treasury.
When a bank moved its entire card programme to blockchain settlement using its own stablecoin, the word worth noticing was entire. Card settlement pilots have been announced for years, almost always confined to a corridor, a product line or a merchant category. Moving a whole programme means the settlement path has stopped being an experiment running beside the production one.
What a ledger actually adds
The ECB's framing, set out in October 2026, rests the case on two properties rather than on the technology in general. Programmability lets settlement be made conditional on rules that execute automatically — payment on delivery, release on a date. Atomicity makes the legs of a transaction settle together or not at all, which removes the exposure that exists between one leg completing and the other.
Neither property requires abandoning the structure of money. The same presentation set out three ways to put central bank money on chain — native issuance, a bridge between the existing settlement system and DLT platforms, or a private intermediary issuing fully backed settlement tokens — and all three preserve the two-tier shape in which public settlement assets sit above private ones at one to one.
This is the part most commentary skips. The interesting designs are conservative about what money is and aggressive about how it moves. A project that proposes to change both at once is usually proposing to change neither, slowly.
Settlement in markets that are not payments
The same word governs other areas we cover, and the pattern repeats: whoever settles carries the risk, so the rules follow the settlement rather than the trade.
| Area | What settles | What breaks without it |
|---|---|---|
| Card payments | Issuer, network and acquirer positions | Prefunded balances sit idle against the delay |
| Securities | Delivery against payment | One leg completes and the other does not |
| Compute markets | Payment for a job that ran | A seller takes payment for work done on cheaper hardware |
| Machine payments | A per-request transfer with no account | An agent cannot buy anything it has no subscription for |
| Interbank transfer | A claim moving between institutions | A tokenised deposit at one bank is not one at another |
The compute row is the clearest illustration that settlement is only half a mechanism. Settlement without attestation lets a seller take payment for work done on cheaper hardware than was sold; attestation without settlement leaves an invoicing relationship, which is what the market was meant to replace. The networks whose utilisation moved shipped both.
The record is part of the settlement
When we examined three production deployments where one autonomous process paid another per request, without a human approving the transfer, the technical part turned out to be the easy part. Settlement worked. What broke in all three cases was the accounting boundary: nobody could say afterwards which agent had authorised what, because the audit trail lived in application logs rather than in the payment record.
That is a general lesson rather than an agent-specific one. A settlement system that moves value correctly and cannot reconstruct who decided to move it has solved the part that was already working. The record tying a payment to the decision that caused it, and a path from that record into ordinary accounting, is what an organisation needs before it will permit any of this — and it is not a protocol feature.
Questions worth asking of any settlement claim
- What settles, between whom, and how long does it currently take?
- What capital is held against that delay today? That number is the size of the prize.
- Does the new path settle both legs atomically, or does it replace one exposure with another?
- Who is the counterparty while the transaction is pending, and are they supervised?
- Can you reconstruct afterwards who authorised the payment, from the payment record rather than from logs?
- Is the first phase the one customers see, or is it interbank? A project starting with the visible half is usually a demonstration.
Questions
- What is the difference between authorisation and settlement?
- Authorisation confirms a payment will happen and takes about a second. Settlement is when value actually moves between institutions and takes a day or more. The delay between them is what prefunded balances are held against, and it is what settlement projects are trying to shorten.
- What does putting settlement on a ledger actually buy?
- Two things, in the framing the ECB uses: programmability, so settlement can be made conditional on rules that execute automatically, and atomicity, so both legs of a transaction settle together or not at all. Neither requires changing what money is.
- Why is attestation discussed alongside settlement in compute markets?
- Because either half alone produces a market nobody serious will use. Settlement without attestation lets a seller take payment for work done on cheaper hardware; attestation without settlement leaves an ordinary invoicing relationship.
- What usually goes wrong in machine payments?
- Not the transfer. In three production deployments we examined, settlement worked and the accounting boundary failed: nobody could say afterwards which agent authorised what, because the audit trail sat in application logs rather than in the payment record.