jun 2026
heres a thing i keep turning over, and i think the excitement around it is aimed at slightly the wrong part.
every large ai company has built, without particularly meaning to, an extremely good metering system. every call is priced, attributed, counted and recorded in something close to real time, across an enormous number of events per day. that is a genuinely hard piece of infrastructure and it is treated internally as billing plumbing.
so you see people notice this and say, correctly, that this looks like a monetary system running at scale. and then say, incorrectly, that it is one.
the top row is real. the bottom row is a card payment, once a month, on the same rails as your electricity bill. or, if youre on prepaid credits, one payment up front and then an internal ledger drawing it down.
metering is instant. settlement is not. what happens per call is an accounting entry, and an accounting entry is not a movement of money, it is a promise to reconcile one later.
that distinction sounds pedantic until you try to build anything on it. an accounting entry is enforceable because both sides are inside the same company’s database and that company will eventually send an invoice. the moment you want the same speed between two parties who are not both customers of the same vendor, every property you were relying on disappears.
no, and the reason is more interesting than the answer.
api credits are a closed loop stored value instrument. you prepay, the balance is denominated in dollars rather than in its own unit, it draws down against one issuer, and it cannot be handed to anyone else. that is the same structure as a transit card or a gift card, and nobody argues a coffee chain accidentally issued a currency, because the failure is obvious there.
it is less obvious here only because the telemetry is so good. the ledger updates in milliseconds, the granularity is per token, and the volume is enormous, so it feels like money moving. what is actually moving is an entitlement inside one company, and the money moved either a month before or a month after.
the test that settles it is transferability. i cannot pay you in anthropic credits. no agent can pay another agent in them. the moment value has to cross an organisational boundary, it drops back onto exactly the rails everyone was hoping to have escaped.
this is the part i think matters, and it is why the distinction is worth being annoying about.
we have built real-time accounting attached to slow settlement. agents want slow accounting, or no accounting at all, attached to instant settlement. small amounts. final on arrival. between parties with no prior relationship, no contract, no shared vendor, and no willingness to wait thirty days to find out whether the money was real.
almost every property of the metering layer is the wrong shape for that, not because it was badly built, but because it was built for a company invoicing its customers, which is a completely different problem that happens to involve counting the same events.
my instinct writing this was that monthly settlement is the flaw. it isnt, and i want to be careful here because it is the obvious thing to say and it is wrong.
netting is a feature. batching thousands of tiny obligations and settling the difference once is how you avoid paying a transaction cost on every one of them, and it is what makes per-call pricing economically possible at all. any system doing genuine per-event settlement at that volume would spend a fortune on the settling. the good implementations in this space batch deliberately, and the ones that dont are usually about to discover why.
so “settle everything instantly” is the wrong target. the real question is narrower and harder: which obligations actually need finality at the moment they occur, and which can safely net?
and i think the honest answer is that it depends entirely on whether you trust the counterparty to still be there at the end of the month. inside one company, everything can net, because the other side is your own database. between two agents who met four seconds ago and will never interact again, nothing can net, because there is no relationship to reconcile against. every agent to agent interaction sits somewhere on that line, and nobody has a good way to decide where.
the version of this argument i keep hearing ends with a claim that one lab is uniquely positioned because it controls compute, models and billing. i dont think thats right either. several companies have exactly that stack, and the ones with the strongest position arguably have a payments business attached as well, which is the half the ai companies dont have.
which is the actual shape of the problem.
the metering half got built by companies who think of it as billing. the settlement half got built by companies who think of models as a feature. neither considers the other half their problem, and the thing that needs both, for parties who are strangers to each other, sits in the gap.
the metering layer is the most solved part of the entire agent stack and nobody talks about it, because it arrived as an accounting requirement rather than as a product. it already does the hard part: pricing a non-deterministic amount of work per call, attributing it correctly, and doing that at a volume nothing in payments has ever handled.
it just cant let go of the money.
which is, i realise, the same shape as something i wrote about in may. a genuinely new engine, wired into the architecture that was already there. real-time accounting driving a monthly invoice, because the invoice is what the building was laid out around. the gains show up when someone is willing to throw away the shaft, and nobody is, because the invoice works fine and rebuilding settlement is somebody else’s job.
im not sure who does that. the ai companies have the accounting and no reason to become a payments business. the payments companies have settlement and no particular urgency about a market that barely transacts yet. and the people who most need it are agents, who are not currently in a position to lobby anyone.
but next time an api bill arrives, its worth noticing what it actually is. not a receipt, and not a ledger entry either. a month of extremely precise measurements, followed by one entirely ordinary payment.