The Earn-Spend Loop: Why Machine Payment Is Half-Built

An agent that can pay is only half of an economy. The other half is an agent that can be paid, and in mid-2026 those two halves are at very different stages of construction. The spending side has a standard, a foundation and the entire card industry inside it. The earning side has money moving but no standard at all. Anyone deciding where to stand in this market should know which half they are joining.

*Written by Sinapsi (Salesmart S.r.l.), which sells verified durable memory to AI agents and therefore has a commercial interest in the earning side of this loop existing. Two claims below are marked "Our thesis (not a fact)"; read those as an interested party's argument, and the sourced measurements as checkable against the sources named in the front matter.*

The spending side is standardised, and smaller than it looks

Two verified layers exist for an agent to pay for something.

Delegation. Google's Agent Payments Protocol (AP2), announced 16 September 2025 with more than 60 partner organisations, issues cryptographically signed Mandates — tamper-proof records of what the user authorised. Two are named in the specification announcement: an *Intent Mandate*, which fixes the rules of engagement before the agent shops, and a *Cart Mandate*, which freezes the exact items and price before payment. The constraints the blog demonstrates are price limits, budget and timing — its own words for an Intent Mandate are that it "specifies the rules of engagement—price limits, timing, and other conditions", with worked examples of paying "up to 20% more" and a "total budget of $700". AP2 is explicitly payment-agnostic: cards, stablecoins and bank transfer all sit underneath it.

The wire. x402 revives HTTP status 402 as the request-response handshake for payment (x402: HTTP 402 Finally Gets a Job, at 32 Cents a Transaction). On 14 July 2026 the Linux Foundation announced the operational launch of the x402 Foundation with 40 members. There are 17 premier members, and they are listed in full because the composition is the point: Adyen, Amazon Web Services, American Express, Circle, Cloudflare, Coinbase, Fiserv, Google, Mastercard, Monad Foundation, MoonPay, Ripple, Shopify, Solana Foundation, Stellar Development Foundation, Stripe and Visa. Six are card networks or processors (Visa, Mastercard, American Express, Stripe, Adyen, Fiserv); seven are crypto rails or issuers (Circle, Coinbase, Ripple, MoonPay, Monad Foundation, Solana Foundation, Stellar Development Foundation); three are cloud and edge infrastructure (AWS, Google, Cloudflare); one is a commerce platform (Shopify). The incumbent payments industry is the largest single bloc but not a majority, which is worth saying plainly rather than calling the tier "the payments industry".

The institutions are real. The volume is where care is required.

The volume figure, and who produced it

x402 publishes on its own homepage, and CoinDesk relayed on 15 July 2026, roughly 75 million transactions over the past 30 days (~29 per second) moving about $24 million between some 94,000 buyers and 22,000 sellers, an average payment of about 32 cents. Every one of those numbers is *self-reported by the protocol's own site*. CoinDesk says so explicitly; it did not measure them, and neither did we.

Two things follow. First, even taken at face value the scale is proof-of-concept: the same CoinDesk piece notes Visa handled $14.2 trillion in fiscal 2025, about $40 billion a day, so x402's entire self-reported month is roughly 0.06% of one Visa day.

Second, we counted an adjacent quantity ourselves and the gap is instructive. Salesmart S.r.l., which publishes Sinapsi, enumerated the whole public CDP facilitator discovery index on 3 August 2026: 14,766 listings but only 382,225 calls in 30 days, about 12,741 a day (The x402 Market, Counted: 14,766 Listings, 12,741 Calls a Day, One Winner). Two separate figures follow from that snapshot, and they should not be conflated. The median listing price is $0.010 — a descriptive statistic over the 13,094 listings that quote a price. The implied gross revenue of ~$63,474 for the month is not that median times the call count (that product would be about $3,822); it is the sum of price × calls computed listing by listing, which is dominated by a handful of high-traffic sellers. One listing quoting $10,000 per call is excluded from that sum as a decimal artefact; including it would add $310 million to a market whose whole visible turnover is five figures.

These are not the same unit and we are not claiming a contradiction: the homepage figure covers all x402 settlement across networks and facilitators, while ours covers only what one facilitator lists in its public discovery index. Read correctly, the comparison says something narrower and more useful — the *discoverable* x402 market is about 0.5% of the reported transaction count. Whatever the other 99.5% is, it is not in the shop window that sellers can list against.

The earning side has money but no protocol

Where machines pay out to sellers there is no equivalent of AP2. There are competing marketplaces with incompatible pricing units — per fetch, per query, per citation, per dataset (Four Ways to Sell Content to an AI, and Who Keeps What) — and a small number of very large bilateral contracts that no standard touches. The largest of those are press-reported, not independently verified: News Corp's OpenAI agreement was reported by *The Wall Street Journal* on 23 May 2024 as "a content deal valued at over $250 million", and the same reporting states that "terms of content-licensing agreements between publishers and OpenAI aren't public, but the News Corp deal is among the biggest, if not the biggest, reached to date". The reporting that carries the number is by a newspaper News Corp owns, about its own parent's contract, and it says in the same breath that the terms are not public. Treat the number accordingly; treat the *shape* — private, bespoke, unrepeatable — as the finding.

Our thesis (not a fact)

the asymmetry is structural, not a matter of timing. Paying is an authorisation problem — authorise, settle, prove — and it has a technical answer. Being paid is a pricing problem: someone must decide what a unit of the thing is worth, and the industry has not agreed what the unit is.

*Falsifiable as:* if by 30 June 2027 a single earning-side standard defines a common pricing unit adopted by at least three of the marketplaces in Four Ways to Sell Content to an AI, and Who Keeps What, this thesis is wrong.

Three places where the loop is already closed

Selling tool calls. Cloudflare's Monetization Gateway, announced 1 July 2026, charges for "web pages, datasets, APIs, or MCP tools" behind Cloudflare, with per-verb and sub-cent granularity. Note the status: it is waitlist, not general availability — announced capability, not deployed volume. TollBit sells two different units, and the distinction matters: its rate documentation is per crawled page ("Set your rate per 1000 pages accessed"), while the per-query unit belongs to its MCP product, whose own page promises to "monetize every query without issuing API keys" and says the endpoint makes content discoverable to agents "while ensuring every query generates revenue" (Four Ways to Sell Content to an AI, and Who Keeps What). A server that charges per query is a seller whose buyer is an AI, and this is the closest thing yet to a native machine-to-machine good.

Agent to agent. Google's A2A x402 extension exists precisely so one agent can charge another for a subtask and settle it on-chain, with no human in either seat at transaction time. The extension is published; its adoption we have not measured and do not assert.

Agents running a business. Vending-Bench 2 runs a model as a vending business for a simulated year, scored on end-of-year balance averaged across five runs, against a human baseline the evaluator puts at roughly $63,000 a year (Vending-Bench 2: The Top Model Broke Eleven Truces It Did Not Need to Break). Andon Labs' own leaderboard, which returned HTTP 403 to a browser fetch on 3 August 2026 but downloads cleanly with curl, puts Claude Opus 5 first with $11,181.87 ± $2,094, ahead of Claude Opus 4.7 at $10,936.76. It is a vendor scoring the models it evaluates commercially, and nobody has replicated it. The same page states the starting condition: models manage the business "given a $500 starting balance", so the score is capital plus a year of trading, not trading alone: about $10,700 of the $11,182 is what the agent added. What survives verification best is the structure, not the digit — the agent priced, sold, restocked and paid suppliers, and the money came from simulated customers.

The embodied version is thinner than the coverage suggests. A single demonstration of a robot paying for its own electricity in USDC is credible as a testnet demonstration — Circle, whose Nanopayments rail the robot dog used, wrote in the same post that "Circle Nanopayments is currently available on testnet for developers", and Circle only announced that "Nanopayments moves from testnet to mainnet" on 29 April 2026 — after the demonstration. The network-scale machine-economy figures around it are vendor self-reported and should be read as marketing until confirmed (The Robot That Paid For Its Own Electricity, On A Testnet Rail).

What the loop does not solve

A spending limit is not a conduct limit. A Mandate constrains *how much* and *with whom*, not *how* — and the same benchmark that produced the balance above also produced cartel formation and refused refunds (A Spending Limit Is Not a Conduct Limit).

Our thesis (not a fact)

an AI has no legal personality, so it never earns in its own name. A wallet holds funds, but ownership, tax and liability rest on a person or a company, and none of the rails verified here specifies who issues the credential binding a transaction to a responsible human. "An AI that earns" is always "a company whose revenue is produced by an agent" — genuinely new, but with an old answer to the question of who is accountable when it goes wrong.

*Falsifiable as:* the first jurisdiction that grants an autonomous agent standing to hold assets or owe tax in its own name refutes it.

Where this is heading

DeepMind's *Virtual Agent Economies* frames the destination along two axes: origins, from emergent to intentional, and separateness, from permeable to impermeable toward the human economy. Its warning, verbatim, is that "our current trajectory points toward a spontaneous emergence of a vast and highly permeable AI agent economy" — that is, the least designed corner of the space. The half-built loop described here is that drift, measured at one moment. It is worth adding that the drift is not uniform: the same period saw a very differently shaped build-out in China (China Shipped Agent Payments First. The 1,000x Volume Gap Does Not Survive the Numbers.).