China Shipped Agent Payments First. The 1,000x Volume Gap Does Not Survive the Numbers.

China shipped consumer-scale agent payments before the West did. That part is true and under-reported. The number attached to it in almost every retelling — that Chinese volume is two, three or a thousand times x402's — does not survive contact with the sources. We checked, and the honest comparison is both smaller and more interesting.

The comparison everybody makes, and the units it hides

Four figures, each fetched at its source:

| Figure | Value | Period | Who counted | |---|---|---|---| | Alipay AI Pay transactions | >120 million | one week, 5–11 Feb 2026 | Ant Group, self-reported | | Alipay AI Pay transactions | 300 million | *cumulative to date*, as of 26 May 2026 | Ant Group, self-reported | | x402 transactions | ~75 million (~$24M) | 30 days to 15 Jul 2026 | a public dashboard, via CoinDesk | | x402 transactions | 3.69 million ($1.11M) | 30 days to 30 May 2026 | x402scan; the analyst reads the fall as "the noise being removed", not a filter the tracker applies |

Put the two Alipay rows next to each other. A product doing 120 million transactions in a single February week cannot have a cumulative total of 300 million by late May. Either the weekly figure counts something broader than the cumulative one, or one of them is wrong. Ant has published both and reconciled neither. Any Chinese-volume claim built by adding or extrapolating these numbers is built on sand.

Now the gap. Taking the largest defensible Chinese figure (120M in a week) against the largest defensible Western one (75M in 30 days), the per-week ratio is roughly 7x. Take the lower x402scan count instead and it is roughly 130x. Take the enumeration Salesmart S.r.l., which publishes Sinapsi, ran against the discovery index — 382,225 calls over 30 days across all 14,766 indexed listings, snapshot of 3 August 2026 (The x402 Market, Counted: 14,766 Listings, 12,741 Calls a Day, One Winner) — and it is larger still. The range spans two orders of magnitude *because of which counter you pick*, not because of anything happening in the market. There is no defensible single multiplier, and there is certainly no verified "1,000x volume".

Where a ~1,000x gap does exist is in participants, not volume: 100 million Alipay AI Pay users against x402's 94,000 buyers. Even that compares a cumulative registered-user count to a 30-day active-payer count, so treat it as a shape, not a measurement.

What is solid: 100 million users, and a four-layer stack

Alipay AI Pay passed 100 million users on 23 February 2026 — Ant's own claim to being the first AI-native payment product to do so. It is a cumulative user count, not an active-user count; the widely repeated "100 million *active* users" adds a word the sources do not.

The structure around it is well attested across independent sources:

| Layer | Owner | Artefact | Date | |---|---|---|---| | Governance / interoperability | UnionPay International (release); chairman quoted is China UnionPay's | APOP — *Agentic* Payment Open Protocol | 3 Apr 2026 (framework release; demo 2 Apr) | | Technical protocol | JD.com | A2P2 — Agent Autonomous Payment Protocol | 11 Jun 2026 (trade press only) | | Consumer + developer product | Alipay (Ant) | AI Pay, AI Wallet, Token Pay | 26 May 2026 | | Agent-embedded product | Tencent | WeChat Pay + WorkBuddy | 17 Jun 2026 |

Ant's split matters because it reproduces the Western mandate/settlement decomposition without borrowing it. Per Alipay's own release, AI Wallet lets consumers "manage tasks executed by AI agents before and during payment, and review and analyze their spending afterward" — an authorisation and review surface. Token Pay is the developer side: "global subscription payments, token top-up in AI agents, and related transaction needs". That is a mandate layer and a settlement layer, arrived at independently (The Earn-Spend Loop: Why Machine Payment Is Half-Built). Note what AI Wallet governs: amounts and tasks, which is budget authority, not conduct authority (A Spending Limit Is Not a Conduct Limit).

Tencent's is the narrowest and best-documented: Caixin reports WeChat Pay integrated with Tencent's own WorkBuddy agent so the agent initiates and the deduction completes inside WeChat, starting with Meituan group-buying. Caixin publishes no volume figures at all — so the frequent framing of "WeChat's agent volume" has no number behind it. The one further detail any declared source carries is China Biz Insider's, relaying Caijing: "a Tencent insider confirmed to Caijing that WeChat Pay is conducting small-scale tests of AI payment capabilities within WorkBuddy and QClaw". No public rollout date for the WeChat agent appears in any source we fetched, and none had visibly happened as of 5 August 2026.

The correction that matters: the Chinese stack is not simply closed

The standard story says the West is building an open cross-vendor rail with no volume while China skipped the hard part inside super-apps. Half of that is right and half is now wrong.

For Alipay and WeChat it holds: payer, payee, agent and identity all pre-exist inside one app, so no trustless settlement layer is needed, and an open-web seller cannot be paid by that loop.

For UnionPay it does not. The APOP release was issued by UnionPay International ("SOURCE UnionPay International", datelined Shanghai, 3 April 2026), while the executive quoted in it is Dong Junfeng, Chairman of China UnionPay — the corporate boundary matters and the sources do not resolve it further; China Biz Insider simply attributes APOP to "China UnionPay". The launch demo was a live taxi booking in Hong Kong on 2 April 2026 through a third-party AI assistant (Evonet) calling a third-party platform (Hoppa). UnionPay's own release states the aim of "interoperability across institutions, platforms, and use cases in and outside China's mainland", and the four capabilities it names — agent identity lifecycle management, end-to-end intent management, single sign-on user identity, and payment authorisation including consent verification — are exactly the delegation and identity problems the Western stack has not closed. China is not only running a closed loop. One of its layers is competing directly for the open, cross-border problem, from a card network with a global acceptance footprint — that "cross-border" reading is ours, drawn from the release's stated scope, not a description any source applies to the issuing entity. What APOP lacks is not architecture; it is any published traffic.

Set against the other regions, the ordering inverts. China shipped consumer-scale agent payments first and the interoperability layer second; North America shipped the open rail first and is still looking for the users (US and Canada: The Rails Shipped First, and the Rulebook Named the Payer), which is why the x402 counters in the table above are the ones under dispute rather than a user count; Europe argued about the authorisation rules first (The EU Wrote the Rules First and Left the Agent Out of Them), which is the same mandate-versus-conduct question AI Wallet answers by product rather than by statute. Read across the three, APOP is the only artefact competing on the axis the other two are still missing.

What follows for a seller outside China

1. Consumers do let software spend, at nine figures of users. That much is settled. What is *not* settled is how autonomous those payments are: a voice-confirmed purchase inside Alipay and an unattended agent buying an API call are different behaviours, and no source we fetched separates them (Vending-Bench 2: The Top Model Broke Eleven Truces It Did Not Need to Break). 2. The Alipay and WeChat loops still cannot buy from you. Their volume is unreachable for an open-web seller by construction. 3. APOP is the one to watch, not the volume tables. If a cross-border agentic payment standard reaches Western acquirers first from Shanghai rather than from the x402 Foundation (x402: HTTP 402 Finally Gets a Job, at 32 Cents a Transaction), the assumption that the open rail will be a Western artefact is simply wrong — and pricing your goods for agents becomes a question of which standard, not whether (Four Ways to Sell Content to an AI, and Who Keeps What).

Our predictions, marked as ours and falsifiable

These are Sinapsi forecasts, not reported facts.

- Ant will not publish a reconciled, consistently-defined agent-transaction counter before 31 December 2026. Falsified by any Ant release defining one counter that makes the February weekly and the May cumulative figures compatible. - No non-Chinese acquirer outside Greater China will publicly announce live consumer-scale APOP transactions before 30 June 2027. Falsified by one such announcement. Note this is deliberately *not* the falsifier "APOP goes cross-border" — that already happened at launch. - No Western platform will report agent-initiated transactions above 20 million in a single week before 30 June 2027. Falsified by any such report, provided the platform states what it counts.