Do we type words into a search box, scroll through a timeline, or go straight to a giant e-commerce site? The battle for the internet's "front door" has entered a new phase with the arrival of AI agents that shop and make bookings on people's behalf. We tracked where seven companies stand today, as free access tied to payment wallets, computing resources on the scale of hundreds of billions of dollars, and even plans for data centers in orbit all collide.

Four rides and a shopping cart topping $200

On November 4, 2026, a modest number surfaced in coverage of ride-hailing giant Uber's third-quarter results. Uber said that over the three months, just four rides had been booked through outside, third-party AI agents (as reported by The Verge). For all the glossy promotion of AI taking care of every errand on people's behalf, real-world transaction volume is still very small.

Even so, from the tech giants to up-and-coming model developers, companies rolled out personal agents all at once this autumn. StatCounter data for October 2026 show that search accounted for 91.47% of referrals to websites worldwide and social media for 8.39%, with AI chatbots at just 0.14% (0.12% in Japan). The reason the companies are rushing anyway appears to be that the place where consumers entrust their "requests" and their "wallets" could become the seat of the next form of dominance, one fundamentally different from selling ad space.

The signs are already showing up at the edges of the numbers. Muse, the agent Meta launched in the US on September 8 and in Japan on October 23 (its first market outside North America), was set up by more than 10 million people in the US and Canada within about six weeks of launch. According to Meta's own prepared remarks for its earnings call, shopping is Muse's most common use: more than a quarter of users shop with it, with an average purchase size of over $200.

Why agents are being given away for "free"

What stands out in the moves of recent months is that Google, Meta, and Amazon have begun handing out the basic functions of their agents for free, or effectively for free. On November 17, Google opened Gemini's "Tasks" feature, tested since May under the name "Spark," to users without a paid plan in six countries, including Japan and the US. Meta likewise bills Muse as "free for most of what people need," and Amazon began free early access to "Alexa+" in Japan on October 29, saying that once it ends, Prime members will pay no additional charge.

What appears to lie behind the companies' willingness to open up compute-hungry agents so generously is a direct link to payments. On the earnings call, Meta's Mark Zuckerberg laid out a plan to offer the product free to most people while combining fees "paid for by businesses" with subscriptions. Rather than lining up ads on search results or social feeds and waiting for clicks, the move reads as an attempt to pull the completed purchase itself into the company's own economic sphere.

On the front line are the designs of "wallets" and "keys." After releasing its always-on agent "dots" to Pro plan subscribers on September 29, OpenAI began rolling out "ChatGPT Wallet" in the US on November 13. "1.2B people use ChatGPT each week, and many of them turn to it to research products, compare options, and decide what to buy. With ChatGPT Wallet, we're now helping people complete purchases," OpenAI said in its announcement.

A race to keep real card numbers out of sight

In letting agents handle money, the companies have arrived at strikingly similar safety designs. Instead of showing real credit card numbers or passwords directly to the AI model, they issue a token or a single-use card number for each payment. OpenAI says ChatGPT Wallet supports eligible American Express, Chase, Mastercard, and Visa cards, and that when the user approves each purchase with biometric authentication, the card network issues a payment token.

The others are building their own payment rails as well. Meta's Muse and "Claude Duet," which Anthropic released in the US on October 21 as a waitlisted research preview, both pay with single-use card numbers locked to a specific merchant and amount, using Stripe's "Link" and similar services. Google uses "AP2," a protocol that records permissions in signed "mandates" (its participants include JCB), and on November 13 it also extended Google Wallet integration to all Gemini app users in the US.

SpaceX's Grok Bot, too, while keeping something of its business-oriented character, has expanded since October into personal shopping and flight and hotel booking. According to its changelog, since October 11 it has supported payment with Stripe Link and with X Money, the company's own money-transfer and payments service.

As far as primary sources show, however, all of these payment integrations are aimed at the US (some also cover Canada). The only one that primary sources confirm can go as far as placing an order in Japan is Alexa+, which Amazon uses in its own store. Japan has become a country where you can "ask" agents to run errands, but not yet one where you can "have them pay" at other companies' stores.

Seven companies, seven gateways, seven origins

The seven companies vying for dominance have stepped into the same ring of "personal agents that act on your behalf," yet the revenue pillars they carry could hardly be more different. Tracing the revenue breakdown in their third-quarter results materials (full fiscal 2026 for Apple) reveals what each company is defending and from where it intends to attack.

  • Google (Alphabet): Of third-quarter revenue of $127.58 billion, search advertising accounted for 49.7% (64.3% for advertising overall) and Cloud for 24.1%. Josh Woodward, head of the Gemini app, has said that "this sensitive data already lives at Google securely, and you don't have to send it elsewhere," and the company is wielding its own assets such as Gmail, Calendar, and Maps (seven products at the two-billion-user scale alone).
  • Meta: A full 97.6% of third-quarter revenue of $64.60 billion came from advertising. It owns enormous social networks such as Facebook and Instagram, used by 3.64 billion people every day, but in the risk factors of its quarterly report (Form 10-Q) it acknowledges its own weakness: "We are dependent on the interoperability of our products with popular mobile operating systems … that we do not control, such as the Android and iOS operating systems and mobile browsers."
  • Amazon: Of third-quarter revenue of $201.94 billion, its own direct sales accounted for 33.9%, third-party seller services for 23.0%, and AWS for 22.8%. It is the "store" itself, with a selection of hundreds of millions of items, and it is trying to keep customers in its own aisles with "Alexa for Shopping," which it says more than 400 million people have used over the past 12 months.
  • Apple: A device giant that earns 53.1% of its fiscal 2026 revenue of $478.83 billion from the iPhone and 25.8% from services. It has more than 2.5 billion active devices, but the current "Siri AI" (in beta, with Japanese support) is at the stage of calling actions that apps have prepared, and CEO John Ternus is reportedly pushing to turn it into an agent that can compete with OpenAI's dots and Meta's Muse (according to a Verge article summarizing a Bloomberg report).

The remaining three are challengers coming from outside the incumbent businesses of search, social media, mega-scale e-commerce, and smartphone operating systems. OpenAI had revenue of $2 billion a month and more than 50 million paying subscribers as of March, and from its base in ChatGPT, used by 1.2 billion people a week as of October, it is expanding its advertising ($1 billion annualized as of August) and its shopping standards. Anthropic, which is preparing to go public, is growing fast, with January–September revenue of $33.2 billion (about $81 billion annualized), but its prospectus (Form S-1) describes individual users as providing "strategic value as a top-of-funnel engine for our enterprise business," and Claude Duet, too, is aimed mainly at professionals and small businesses.

The company with the most unusual structure of all is SpaceX, which acquired xAI in February 2026 and went public in June. Of its third-quarter revenue of $14.575 billion, the "SI (Super Intelligence) division," which bundles Grok, X, Cursor (whose acquisition closed in August), and the provision of computing resources, accounted for 51.9% ($7.560 billion), outstripping the connectivity business, which includes Starlink, at 36.0%, and the space business, at 12.1%. On top of the base of X and Grok, used by 550 million people a month (as of the end of March), it has layered X Money, which touts licenses in 47 US states and Washington, D.C. A rocket company is turning into an AI-and-finance conglomerate.

Will stores open their gates? Amazon's wall and a court battle

No matter how smart the agents or how clever the wallets, no deal happens if the "store" on the other side of the purchase says no. Across the internet, stores are now starting to split between those that welcome outside agents and those that shut their gates. While Shopify has opened its product catalog even to agents it has no partnership with (as reported by The Verge) and Walmart has announced an approval-based gateway (as reported by TechCrunch), Amazon, the largest of the retailers that have shut out other companies' AI agents, has stood in the agents' way (as reported by TechCrunch).

Starting on the night of September 20, Meta's Muse was shut out whenever it tried to access Amazon, with the message "Continued access by an unauthorized AI agent violates Amazon's Conditions of Use, to which our customers have agreed," TechCrunch reported. In the airline industry, United Airlines restricted third-party personal agents on November 11, saying their actions "may result in unintended purchases, account changes, or use of MileagePlus miles" and "have the potential to violate FAA and DOT regulations." Airbnb, too, wrote in its shareholder letter, "We don't believe people will use one AI agent for everything. We believe the best agents will be specialized," and set out plans to launch its own travel agent in the first half of 2027.

This clash has reached the courtroom in Amazon's lawsuit against Perplexity. A federal district court granted a preliminary injunction in March this year, but on August 4 the US Court of Appeals for the Ninth Circuit vacated it and remanded the case, holding that on the facts presented at the preliminary-injunction stage, it was the user, not Perplexity, who "accessed" Amazon's computers, and that Amazon's claims under federal and state laws against unauthorized computer access were therefore unlikely to succeed on the merits. The court noted, however, that its ruling does not prevent restricting access through terms of use, and that it did not decide other forms of liability.

Amazon amended its complaint on September 21, alleging among other things that Perplexity disguised its AI agent as a human user (Perplexity disputes this). A hearing on Perplexity's motion to dismiss was scheduled for November 20; as of November 21, the court record did not yet show the outcome.

A ground war in the hundreds of billions, and plans for computing in orbit

Another axis quietly shaping who wins is computing power on an entirely different order of magnitude. The more agents keep working behind the scenes around the clock, the more inference costs are likely to swell.

Grok Bot's usage limits were in fact abruptly tightened, and SpaceX CEO Elon Musk reportedly explained in a November 1 post on X that "The cost of serving Grok Bot was unsustainably high relative to revenue" (as reported by TechCrunch). Meta's Zuckerberg, too, has cited the heavy computing demands as a reason Muse has not yet been rolled out worldwide.

The 2026 capital-expenditure outlooks in the companies' earnings materials coldly reflect the gaps in their staying power. Amazon expects to spend about $220 billion, Alphabet $195 billion to $205 billion, Microsoft about $170 billion for the calendar year, and Meta $135 billion to $145 billion; of these, Amazon and Alphabet have stated outright that they will spend even more in 2027. Apple's capital spending for fiscal 2026, by contrast, came to just $9.04 billion, and on Alphabet's earnings call Apple was named as a customer pre-training models on Google's latest in-house chip, the "TPU."

The two model specialists without giant clouds of their own are making up for it with huge long-term contracts. OpenAI has signed one deal after another with NVIDIA, AMD, Oracle, Microsoft, AWS, and others, and plans to make the first deployment of its own inference chip, "Jalapeño," before the end of the year. Anthropic, which has no factories of its own, goes further still: the purchase and lease commitments listed in its S-1, counting those in place as of the end of September together with those signed since, total about $900.9 billion, and it is buying up computing power from all four of Google (at least $121.3 billion), SpaceX (up to $120.5 billion; in May it also announced a deal to use the full capacity of the large Colossus 1 computing cluster), Amazon (at least $110 billion), and Microsoft, even though they are also its competitors.

Google and SpaceX, meanwhile, have begun writing even "space" into real plans, to get past the constraints of the power grid on the ground. Google has set out "Project Suncatcher," which would run TPUs in orbit on solar power; on October 1 it launched a single prototype satellite on a SpaceX rideshare flight and confirmed contact with it, and it plans a two-satellite communications test in 2027.

SpaceX itself has applied to the US Federal Communications Commission (FCC) for a network of up to 1 million orbital data-center satellites. Its prospectus says it expects "to begin deploying our orbital AI compute satellites as early as 2028," while its risk factors state that "we have not, and no one else has, previously operated or attempted to operate orbital AI compute." Even as OpenAI's Sam Altman has reportedly called the idea of space data centers "ridiculous" given "the current landscape" (as reported by Fox Business), these two companies are pressing ahead with plans and tests aimed several years out.

A wave of regulation pries open the "default seat"

Still, the companies that hold the gateways of devices, operating systems, and search will not necessarily be able to carry their position over into a monopoly on agents. In the US a court ruling, in Europe a competition authority's decision, and in Japan a report by the Japan Fair Trade Commission have each set a course toward prying open, for AI agents and from the very start, the "defaults" (the default seat) and the "keys to the OS" that were once at issue in search and browsers.

In the US, the final judgment in the Google search monopoly case (December 2025; both sides have appealed) brought the Gemini app within its scope, limited default-placement contracts to no more than one year, and ordered Google to provide search data and search results to qualified competitors as well, including generative AI companies (the qualification process is still being set up, and no company has yet been qualified). The European Union went further: on July 16, it ordered Google to open features such as long-pressing Android buttons, wake words, and automation of on-screen actions to rival AI assistants on the same terms as Gemini (the main deadline is August 1, 2027, at the latest). In Japan, under the new smartphone law (the Mobile Software Competition Act) that took full effect in December 2025, the Japan Fair Trade Commission set out its view in an April 16 report that if a designated mobile OS provider limits on-device generative AI to its own product and prevents others from equivalent use, this would constitute prohibited conduct under Article 7, item 2 of the act and would violate that article unless a justification is recognized.

Turning back to consumers on the ground, in an August survey by Feedvisor and Zogby (1,006 people in the US; it has been noted that the sample may be skewed), 62% said they "would never let an AI agent order and pay on their behalf," and the help pages and terms of Google and Meta are lined with warnings about mistaken purchases and data sharing. On the other hand, a June survey by PYMNTS found that 19% of US adults (about 49.6 million people) were already starting their product research with AI.

Will the giant stores lock up their own aisles, will search and social media take custody of our wallets as an extension of everyday habits, or will devices and the models themselves become the new gateway? Amid a tangle of hundreds of billions of dollars in ground infrastructure, satellites in space, and courtroom proceedings, where our "requests" will end up is still in no one's hands.