Amazon is defending $68.6 billion in ads. Visa and Mastercard wrote the rules a year before Meta showed up. And stablecoins get the traffic nobody else can carry.
On September 8, 2026, Meta launched Muse, the AI personal assistant that not only can chat with users like Claude or ChatGPT, but can actually act, shop, and pay on behalf of users. In just over three weeks, it had already been downloaded 5 million times. On September 20, just 12 days after launch, Amazon blocked its access, claiming that it violated the website's T&Cs.
The actual story follows the money: Amazon's ad business, at $68.6 billion in 2025, is now bigger than its entire subscription revenue line ($49.6 billion), Prime included (10-K). Meanwhile, Bank of America cites third-party shopping agents as the source of "a bear narrative that Amazon could see erosion of direct traffic and pressure on ad revenues" (article). Muse skips search results pages and bypasses the whole recommendation engine, which means no sponsored listing and no ad dollar for Amazon.
Amazon only wants its own shopping agent
Amazon already runs a shopping agent within its walled garden: Alexa for Shopping (formerly known as Rufus) drove nearly $12 billion in incremental sales last year. It has an agentic auto-buy feature, and it bills sponsored placements per click. A walled garden that runs its own taxi service has no reason to allow ride-hailing platforms to come in and pick up its residents.
Rivals moved fast, too. OpenAI launched its own personal agent Dots three weeks after Muse (September 29), and Google launched the new Gemini universal agent on October 8. Meta may have forced other players at the table to show their hands sooner than they wanted. But the race to build the agent is only half the story. The other half is who sets the rules the agent plays by, and that race started more than a year earlier.
Three bets on who controls the ride
Ride-hailing maps the three players neatly.
Meta wants Muse to be the Waymo of shopping: once you start the journey, no human intervention is required until you reach the destination (at least on the happy path). Muse can execute requests all the way to checkout using its VM, which runs automated browsing. For now, though, actual payment still requires explicit user authorization, so today's Muse is closer to driver-assist than full self-driving.
Amazon is running the only approved taxi and classified ads services in town. It blocks third-party taxis for operating without a license, while promoting its own service to businesses on every corner. It wants both businesses and customers to pay it alone, and the last thing it wants is unlicensed taxis coming in and depriving it of its ad revenue. It is also fighting in court to keep those taxis off its roads. Amazon won a preliminary injunction against Perplexity's Comet shopping agent, then lost it: in August, the Ninth Circuit ruled that an agent acting at a user's direction doesn't independently "access" Amazon's site, the user does (Amazon v. Perplexity AI). That is why Amazon is now citing its terms of service instead of anti-hacking law, and why the question of who is responsible when an agent buys the wrong thing is still open.
And then there are Visa and Mastercard, who are Uber in this case. They create rules and governance standards for both sides of the platform (and earn network fees on the transactions that flow through them), but they do not run the cars or hire any drivers. Unlike Uber, they don't own the customer either: the banks do. Visa Intelligent Commerce and Mastercard Agent Pay, both announced in April 2025, more than a year before Muse, are agentic commerce infrastructure and frameworks, including agent tokenization and authentication, that enable agentic payments for AI partners, financial institutions, and merchants. They don't touch the existing rulebook: the issuer is still the consumer's bank, and the acquirer is still the merchant's bank.
The blue ocean cards can't reach
Almost every one of the three bets still settles on card rails, and card rails have a floor. At the risk of overstretching the analogy, the blue ocean is the places that traditional rails don't reach yet: transactions too small and too frequent to carry economically, such as payments for APIs, compute, and data. Imagine a developer handing an AI coding agent a $50 budget: during a particular task, the agent buys extra cloud compute, pays per call for a market-data API, and pays per request to a testing tool, each for a fraction of a dollar. In fact, Cloudflare's Pay Per Crawl already charges AI crawlers as little as a cent per request, using the same HTTP 402 status code the x402 protocol is named after, with 50-plus major publishers signed on.
Be honest about the size of it, though. x402, the leading agent-to-agent payment protocol, has processed roughly $50 million in total since October 2025, and Adyen notes that much of that volume comes from testing and hackathons. It is a blue ocean, but it is still mostly empty.
Old AI had a moat. Agents don't.
If all of this feels new, the AI part isn't. The payments industry has been using AI for a long time now, just not the LLM chatbots that everyone talks about these days. Deep learning has been used to classify fraudulent transactions, flag AML alerts, and underwrite credit. According to Mastercard's own survey, 42% of issuers and 26% of acquirers have saved more than $5 million in fraud attempts over the past two years.
However, deep learning needs data to be useful, and a classifier is only as good as the proprietary data it trains on. That data sits with the two networks, the largest issuers, and whichever processor is the data exclusivity holder for a given merchant. Most fintechs don't have the scale to compete with the incumbent players in this space. AI agents have no such moat yet. Nobody has a decade of data on agent-initiated commerce, which is exactly why Meta, Amazon, Visa, and Mastercard are fighting over it in public.
A skeptic will say this is ordinary platform competition, not a payments story. But the rules being written now decide who gets to sell to agents on equal terms: an open standard lets any merchant or agent in, while a closed loop lets whoever owns the chokepoint decide, one cease-and-desist at a time.
What each player should build now
Which bet you back depends on what you already own.
Merchants: make sure your site is set up to be discovered by agents, and that your content can be read by and optimized for machines. Better yet, configure your site to handle machine traffic, and if you sell API calls, data, or compute, accept stablecoins for machine-to-machine payments. If you're not Amazon, build for the open standard.
PSPs: this will be another type of customer to route to the different payment rails. On the pay-in side, adopt a multi-rail strategy: expect cards, stablecoin wallets, and credits to coexist. The same goes for settlement. The newer product is verifying the agents themselves: offer merchants "know your agent" checks, the bot equivalent of KYC, so they can let trusted agents in instead of blocking all of them the way Amazon does. And on fraud and AML, buy the models rather than build them: the data moat belongs to the networks and the largest issuers.
Stablecoin operators: go after the blue ocean and don't compete for the Muse-style retail agentic commerce volume: cards will win, because card rails already have the acceptance, consumer protections, and chargebacks that shoppers rely on. The real opportunity is in the Evaluation Economy, where buyers pay for value as it is revealed, in small continuous payments that card rails can't carry economically. Stablecoins also have a structural edge there: an AI agent can't open a bank account, so when there's no human or company behind the payer, a stablecoin wallet is the only account it can hold. Yes, Mastercard's Agent Pay for Machines can already settle in stablecoins, but that's the Uber model again: the networks set the rules and leave the issuance, custody, liquidity, and conversion to someone else. That someone else is the opportunity. Finally, build the dispute and evidence layer alongside settlement: a verifiable record of what the agent was authorized to do and what it actually did. The courts are still deciding who is responsible when an agent acts, and that record is what everyone will reach for when something goes wrong.
About us
Aquanow is a global digital asset infrastructure and liquidity platform enabling institutions to launch and scale digital asset and stablecoin capabilities. Operating through regulated entities across multiple jurisdictions, Aquanow combines institutional-grade technology and global liquidity to help more than 450 organizations worldwide - including Visa, Emirates NBD and WonderFi - bring digital asset products and services to market.
.png)
