AI shopping agents can now find a product, compare options and complete checkout for you, but almost nobody involved in the transaction fully trusts them yet. Agents such as Meta’s Muse and Instinct plug into payment processors like Stripe to pay once a user gives the go-ahead. Card networks are building ways to settle disputes over purchases an agent made, and online stores are torn between blocking these agents and courting them. Meanwhile, on the infrastructure side, SpaceX’s AI unit has turned its spare computing capacity into a fast-growing cloud business. Here is what is happening on both fronts and what it means for anyone thinking about letting software spend their money.
Shoppers still want to press the buy button
Consumer appetite for fully autonomous checkout is low. In a consumer survey by ACI and YouGov, only 7% of fashion shoppers said they were willing to let an AI agent complete a purchase without human approval. Most people use agents for discovery, comparison and product recommendations, then keep the final confirmation for themselves.
Whether to hand a card to an agent comes down to personal risk tolerance. Once financial authority moves to autonomous software, though, merchants, payment processors, AI developers and shoppers all face security and privacy questions none of them has had to answer before. With shopping agents still at an early, experimental stage, holding back on sharing card details for now looks like a reasonable choice.
Card networks turn chat logs into evidence
Visa and Mastercard still have to settle transactions that an agent starts. The hardest part is the chargeback, the process a cardholder uses to dispute a charge and reverse it. If an agent misreads a request and buys something the user never wanted, it is not yet clear who carries the legal or financial liability.
Mastercard’s answer is a protocol called Verifiable Intent. It works in three steps:
- It captures the conversation between the shopper and the agent.
- It stores that record so it can be audited later.
- When a charge is disputed, the transcript serves as evidence of whether the agent followed the user’s exact instructions.
In agent commerce, then, what the user actually asked for becomes the key piece of proof. Shoppers can help themselves by keeping their own instructions and chat history whenever they delegate a purchase.
▲ Chat logs as evidence in payment disputes
Merchants block agents but still want the traffic
For more than 30 years, online stores have built fraud defenses designed to spot and stop automated bots. Shopping agents break that model: they behave like bots, yet they act on legitimate authority delegated by a real customer.
Amazon has blocked Meta’s Muse from its marketplace, citing concerns that an undisclosed third party could scrape user data and browsing histories without permission. A deeper worry is first-party behavioral data. When an outside agent skips the browsing and goes straight to the purchase, the store loses its own record of what the customer looked at and weighed.
Retailers are not rejecting AI outright. They want their products to show up in AI chatbot recommendations, while steering shoppers toward first-party agents that run on their own websites. The strategy appears aimed at capturing AI-driven sales without handing valuable data to someone else’s agent.
Credit cards, not crypto, carry agent payments for now
Crypto developers want to give agents their own digital wallets so programs can pay each other directly. Circle has released the Circle Agent Stack and command-line tools that let agents create and manage their own wallets, and it points to open standards such as x402 for instant, low-cost payments between agents with no human in the loop. Circle says agents are overwhelmingly choosing the USDC digital currency to settle those payments.
Payment industry experts see a different picture at the checkout counter. Merchants are not prioritizing crypto because everyday shoppers are not asking for it.
| Credit cards | Crypto wallets | |
|---|---|---|
| Current use | Handle the vast majority of consumer purchases | Little demand from consumers |
| Main backers | Card networks, processors, merchants | Crypto developers, Circle |
| Near-term role | Primary focus of agent commerce | Machine-to-machine payments between agents |
For consumer shopping agents, the near-term build-out looks set to run on credit cards and existing payment gateways.
SpaceX’s AI unit becomes a cloud provider
About a year ago, executives proposed renting out spare data center capacity to outside customers. Elon Musk rejected the idea at first, calling cloud computing unsophisticated and arguing that acting as a hosting utility would not suit an ambitious AI lab. He changed his mind earlier this year after seeing how badly Anthropic needed more compute.
The company started by offering older NVIDIA GPUs that were not being used for its own model training. The customer list then grew beyond Anthropic to include Reflection AI and Google. Anthropic made an initial commitment of $45 billion, and Reuters reported that Anthropic’s IPO prospectus shows a further $40 billion commitment. Earlier this summer, the company also held preliminary talks with Microsoft about access to its compute.
The buildout relied on several unusual moves:
| Area | Approach |
|---|---|
| Speed | 200,000 GPUs brought online in 122 days |
| Power | Gas turbines ordered years before sites were final, giving 400 megawatts of on-site power |
| People | SpaceX engineers on six-to-eight-week site rotations; Tesla engineers on cooling and mechanical design |
| Automation | ABB industrial robot arms coiling and dressing network cables |
| Construction | Prefabricated modular data center blocks shipped to the site |
Generating its own power let the company switch on hardware without waiting for grid approval, which was the key to its speed. The approach came with tradeoffs. Redundant backup power was deliberately left out to save time, which led to reliability problems, including a major outage a few weeks ago.
The bigger open question is how long the capacity will stay available. Musk framed leasing as a temporary way to earn money from idle hardware until it is needed to train Grok models. Companies that want stable compute under multi-year contracts face the risk that xAI could pull capacity back for its own work.
▲ A rapidly built AI data center
What to check before an agent pays for you
Shopping agents are convenient, but the rules for payments and disputes are still being written. A few practical guidelines follow from where things stand today:
- Let the agent search, compare and recommend, but confirm the final purchase yourself.
- Keep your instructions and the chat history so you have evidence if you need to dispute a charge.
- When a retailer offers its own shopping agent, consider using it, since you keep direct access to that store’s customer support and warranty coverage.
- Expect third-party agents to be blocked on some stores.
For businesses renting AI compute, the lesson from SpaceX’s pivot is to ask a newer cloud provider whether it plans to reclaim leased hardware for its own training before signing a long-term deal.