Six months ago I wrote about the coming marriage of crypto and AI. They indeed got hitched. It is turning out to be a somewhat different relationship than I originally imagined. Some of the dreams they had were realised, others were dashed, and new ones unfolded.
Let’s start with money, an unromantic but practical matter in any marriage. In the original article I described x402 — the protocol that lets software pay for things natively on the internet in stablecoins — as a potential game-changer for internet commerce. (A refresher: x402 allows software to request an internet service and to receive a “402” message saying payment is required, automatically allowing a stablecoin payment to be made.) No credit card number is entered, nobody creates an account, and there is no Captcha asking a computer to prove it can identify all the bicycles in a photograph. One machine simply pays another machine.
The stampede of diverse industry members joining the x402 party has been notable. In July the x402 Foundation was launched under the Linux Foundation with 40 member organisations, including Visa, Mastercard, Stripe, Google, AWS, Cloudflare, Coinbase, and American Express.
On the face of it, this is a list of the usual suspects. But notice that it includes card associations such as Visa and Mastercard, whose entire business model rests on skimming a slice off every transaction. Yet they have signed up to steward a protocol whose central selling point is that it charges no protocol fee at all, leaving only blockchain gas costs, measured in fractions of a cent. But there is no way they can stay on the outside, where they have no influence. They have to be inside the room.
So how’s it going? According to blockchain analysis firm Chainalysis more than 138-million x402 transactions had taken place on Coinbase’s Base blockchain by the end of the third quarter. This is still tiny in the grand scheme of payment transactions; x402 is still a baby. The more interesting number is their size. More than 87% of the transactions were worth between one hundredth of a US cent and 10c. These are not people buying cars with bitcoin. They are the tiny kind of transactions that make sense to software but almost no sense to humans.
Imagine asking an AI agent whether you should invest in a particular company. It might buy 3c worth of company filings from one service, 5c of sentiment analysis from another, 8c of specialist industry data and perhaps 20c of inference from a powerful AI model. It could employ another agent to inspect patents and yet another to check regulatory filings. You give your agent $2; it spends 43c assembling the expertise and gives you the answer.
This is a different conception of internet commerce. The customer is software (likely an AI agent), the workers may be software, the procurement department is software, and the transactions can be so small that no human would bother making them.
Over-engineered
Credit cards are hopelessly over-engineered for this brave new world. Visa and Mastercard work well when I want to buy a $900 television. They have fraud protection, chargebacks, identity checks and enormous global clearing systems. They are useless when a computer in Cape Town wants to buy four thousandths of a dollar of data from a computer in California.
And so the payments industry and card associations have noticed. Visa now talks openly about the intersection of agentic commerce and stablecoins. Mastercard has launched Agent Pay for Machines. Stripe supports agentic commerce using conventional payments and stablecoins. AWS and Coinbase have integrated x402 so that publishers and API providers can charge agents for online resources.
Cloudflare (a gigantic internet security infrastructure provider that sits between most of us and the websites we visit) may have stumbled onto an especially important application. Content publishers are facing a Hobson’s choice when AI agents arrive — either allow machines to ingest their journalism, research, or data for free, or block them. Cloudflare’s Pay Per Crawl introduces another possibility. Let the AI agent in, but charge it. An agent encounters an article or database and is told access costs 0.7c. It decides whether the information is worth 0.7c, and if so, pays and proceeds.
Which brings us to the matter of identity. In my earlier column I framed the provenance problem as one of telling human-made content from machine-made content — was this post, this photograph, this news report written by a person? That framing now looks quaint. The urgent question is no longer whether a post was written by a human. It is whether the software agent knocking on a bank’s door, asking to move money, has a human standing behind it — and, when appropriate, which human. Without these provenance guarantees, the entire enterprise of AI-agent-facilitated micropayments would walk into a grumpy wall of regulators.
World — the iris-scanning blockchain project formerly known as Worldcoin — has pivoted hard in this direction. In March it launched AgentKit, which lets AI agents carry cryptographic proof that a unique human backs them, and plugged it into x402. On October 18, World extended this to banking and finance, pitching it as a way for businesses to check whether agents moving money and shopping are acting for a real person.
Then there is code security. My earlier optimism about AI being crypto’s security partner needs amending. I wrote enthusiastically that AI would hunt down the bugs that make smart contracts such a rich vein for the thieves prowling around the crypto space. It does. But it also exploits them more effectively than it finds them. A recent report from OpenAI found that AI is better at hacking than it is at patching — a concerning asymmetry.
The fourth strand I discussed, in which crypto would gift AI its hardened experience in decentralised compute, remains a dream unrealised. The vision was that crypto would teach AI how to pool thousands of idle machines into a distributed alternative to the hyperscaler oligopolies (Google and others), a problem that has long been solved by crypto’s decentralised architecture. But it turns out that decentralisation of AI compute is harder than imagined, especially for pre-training, which requires long, centralised, uninterrupted training runs.
So what will sustain this marriage if not all of the original promise remains? It is this: AI agents need the things that legal systems give persons — an identity, a reputation, a bank account, a way to pay and a way to be held responsible. Nobody in the regulated financial world was going to hand those to software quickly, and crypto already has them on the shelf.
The marriage is going well, all things considered. But as we all know, marriage is a long game with many obstacles along the way. To flog a now admittedly tired metaphor, we wish the newlyweds well, and may they bear many children.
The writer is visiting professor at the Centre for Graduate Studies, Rhodes University, and a partner at Bridge Capital as well as a columnist-at-large at Daily Maverick, Daily Friend and the FM. His new book, Checkmate: How Shoprite Checkers Rewrote the Rules and Won, is published by Jonathan Ball.
