Did money move?
Yes. Facilitator-linked USDC transfers are visible.
A huge count carried little money.
A transfer can be smaller than one cent and repeat thousands of times. Count and amount must be read side by side.
18.8Mtransfer records across the 30-day sample
266.6KUSDC written on those transfers
91.6% of the records were below 0.01 USDC. They carried only 13.9% of the amount. The headline count mostly describes repetition, not spending power.
One network made the count look enormous.
Solana supplied 99.8% of the transfer records but 94.0% of the USDC amount. The gap matters.
Remove Solana and Base still shows 45,918 records carrying 16,100 USDC. Activity survives the removal test, but the giant count does not. The other network held 0.2% of records and 6.0% of amount.
More records did not mean more buyers.
On 2026-09-17, the two-network total reached 18,847 records. Keep the wallet and seller observations beside that number.
18,847transfer records that day
1,021 USDCUSDC transferred that day
690paying-wallet observations
449seller observations
30 days earlier · 559,920 records2026-09-17 · 18,847 records
The latest seven days had 95.2% fewer records than the seven days before.This is record activity, not a customer count.
That is about 27 records for each paying-wallet observation on the latest day. One wallet can appear more than once across networks, so even 690 is not a count of people.
7known payment helpers matched on Base
2known payment helpers matched on Solana
The line stops here.This edition does not split the records by individual payment helper. We cannot honestly say whether one helper produced most of the activity.
What sellers offer is not what buyers used.
The service list shows machine-readable prices: software can see the price, pay and continue. It is a shelf, not a receipt book.
15,487services with a listed price
1,937seller domains
40,113listed payment choices
86.6%of choices use USDC
The middle listed USDC price is 0.010 USDC, and 34.6% of priced USDC choices are below one cent. Tiny prices fit data lookups and model calls. They do not prove those services were bought.
“Can it pay?” is the easy question.
The harder questions are who allowed the choice and what happens when the choice is wrong. Payment routes answer them differently.
Did software act?
Yes in named trials. The anonymous wallet record cannot show it.
Did AI choose?
Not visible in the wider onchain record.
Did the owner allow it?
Some systems record limits or signed instructions. Coverage is uneven.
Can a mistake be fixed?
Card routes have disputes. A confirmed chain transfer has no built-in reversal.
| Payment system | Who or what pays | What allows it | How money moves | If it goes wrong | Still unclear |
|---|---|---|---|---|---|
| Visa Intelligent Commerce | Named agent | Customer-set limits | Card network | Card dispute process | A poor product choice may not count as a payment error |
| Mastercard Agent Pay | Registered agent | Recorded instructions | Card network | Card dispute process | The agent still needs clear authority to choose |
| x402 | Paying wallet | Wallet approval | Blockchain | No built-in reversal | Wallet approval does not prove that AI was allowed to choose |
| Stripe MPP | Seller's account rules | Limited-use payment token for cards | Card or stablecoin | Depends on payment route | Controls differ with the payment route |
| Google AP2 | Signed identity | Signed buying instructions | Uses another payment route | Depends on payment route | The legal effect may differ by country |
Two purchases prove the route works. No more.
These named card-network cases are bounded examples. They show a controlled agent purchase completing under stated rules. They do not measure market-wide use.
Europe
Visa
AI agents completed purchases at independent merchants under cardholder-set limits.
Read the official accountGermany
Mastercard
An authenticated Agent Pay transaction booked and paid for a real experience.
Read the official accountA successful payment does not prove that the product choice was good. A refund or card dispute may fix a payment problem; it may not fix a poor decision the owner had already allowed.
Short answers before you trust the headline.
Does one transfer record equal one user?
No. One wallet can transfer thousands of times, and the same wallet may appear on more than one network.
Does a wallet signature prove AI chose the purchase?
No. It proves that a wallet key approved the transfer. A person, a fixed script or an AI system could have caused it.
Why are so many transfers tiny?
The chain does not say why. Small data purchases are one possible explanation. Testing, repeated software activity or other transfers through the same helper could produce the same pattern.
Can a blockchain transfer be reversed?
The confirmed transfer cannot be cancelled. The receiver can send money back, or a separate agreement can provide a refund, but the original record remains.