Amazon, Muse, and the Layer an Agent Can Skip

Two Amazon-is-dying stories landed in the same week, and they sound like one argument. Meta launched Muse on September 8. Amazon blocked it from amazon.com on September 20–21 after asking Meta to take the store out of the agent. Users who sent Muse shopping hit an anti-bot wall. The tape heard a fortress cracking: if Amazon has to lock the door, the storefront is already lost. The other story has been running all year. Companies are supposedly leaving AWS for CoreWeave and Nebius, because a new cloud that rents NVIDIA racks can offer the same service the old one does.

Both stories sort on a layer an agent or a GPU renter can skip. The MAG7 piece already argued that Amazon is the system of record in physical commerce and in a hyperscale cloud at the same time. The Muse note already argued that a personal agent sorts on habit plus checkout, and that Alexa is this category without the habit. This page is the skip test those two pieces imply and did not run: which layer did Muse actually threaten, and which layer do the neoclouds actually replace.

Amazon's own statement, as The Register printed it, is an opt-in rule borrowed from food delivery and online travel: third-party applications that buy from other businesses should operate openly and respect a provider that says no. The same write-up is honest about the cash register behind the rule. Amazon generated more than $68B of advertising revenue last year because shoppers still walk the store. An outside agent that compares, decides, and pays from a Stripe one-time card never sees the sponsored placement. That is a real threat to the interface and to the ads line that sits on it. It is not, by itself, a threat to the warehouse.

The ads line is the one the block is written to protect. It is also the one that was still accelerating in the last print — +26% on $19.8B, in the same quarter third-party seller services at $46.8B outsold AWS. A ranking that starts from the popup will call the storefront dying. A ranking that starts from those two lines will say the interface is the thing Amazon is willing to police, because the network and the transaction are still printing. The habit-plus-checkout note already put Amazon's learned-interface pillar on the weak side of the Muse cohort. The block is that cell defending itself. It is not a downgrade of the other four.

The grid does the work the popup cannot. Amazon is strong on the transaction, the record, the network and the data, and the learned-interface pillar is the one the framework marks as weak. That is the cell Muse attacks. Walmart's learned-interface pillar is marked the same way — an agent can shop Kroger, Target, or Amazon through one chat window — and Walmart's fortress is the store grid, not the app. Shopify is the name that invited the skip: Shop Pay is the rail Meta named on September 8, and the transaction-embedding pillar is the one Shopify wants an agent to call. CoreWeave and Nebius reach a durable mark on none of the five. That is the framework saying, without sentiment, that a rented GPU against a loan is not a cloud record, however large the backlog attached to it. The hyperscaler-build piece already grouped those two as borrowing to build for others. This page is the competitive claim that grouping implies: overflow is not displacement.

CoreWeave in one quarter is about a sixteenth of AWS. Nebius is about a seventy-second. Those ratios are not a smear. They are the scale at which 'a lot of companies are leaving AWS' would have to print if it were happening. What the neocloud prints actually show is the opposite customer: Microsoft, Meta, and the frontier labs buying spare training capacity from a specialist while they keep the identity plane, the object store, and the compliance perimeter where they already are. Microsoft declining a CoreWeave expansion and signing with another neocloud — the fact the CoreWeave analysis already carries — is the tell. Overflow is interchangeable. IAM is not. The Palantir piece ran this test on software: the agent has to call the ontology. The cloud version is the same sentence with a different noun. The training run can call CoreWeave. The account cannot.

Jassy told investors AWS is capacity-constrained through 2027. A capacity-constrained record leaking overflow to a specialist is how you would expect this cycle to look. It is not how a degrading record looks. A degrading record loses the contracted book, or it loses growth while the specialist takes the enterprise account. AWS grew 37% — the fastest in 18 quarters — and the 10-Q carried $496B of remaining performance obligations with a 6.4-year weighted-average life. That book doubled in two quarters. The build piece already warned that a material share of hyperscaler backlog is circular: Amazon funds Anthropic, Anthropic commits AWS spend. Circularity is a growth and ROIC fact. It is not evidence that Nebius replaced S3.

• Learned interface — the skippable layer. Alexa, the app, the search box, the sponsored module. Muse's browser is built to walk around that surface. The framework already rates the cell as weak on Amazon and on Walmart. Amazon's own Alexa for Shopping is the in-house version of the same skip. The block is Amazon refusing to let someone else's agent sit in a cell it already knows is not the moat.
• Transaction embedding — the layer Muse still has to call if the item ships from Amazon. 1-Click, Prime, FBA, the Buy Box. Automating the shopper does not remove the settlement. A Stripe one-time card can pay. It cannot put the SKU in a same-day node the seller does not run.
• System of record — two of them, and neither is a chat window. The store record is the order, the catalog, the last node that touched the SKU. The cloud record is S3, IAM, the account, the $496B book. An agent that reorders paper towels still clears the first. A training job that bursts onto CoreWeave still comes home to the second.
• Network effects — the marketplace a seller does not dual-run. Third-party seller services at $46.8B in a quarter is the network as revenue. A household does not hold two Prime badges. A seller does not split FBA across two fulfillment graphs that both confer the badge. GPU rental has no such loop: CoreWeave's and Nebius's analysis both mark the network cell as not applicable, because compute is fungible across landlords.
• Proprietary data — the exhaust the ads line bills. Purchase intent and last-mile telemetry are what advertising at $19.8B, +26%, is charging for. Muse that never walks the store never feeds that exhaust. That is the cash reason for the block. It is also why the block, if it holds, protects the data pillar rather than proving it already died.

The cloud half of the same list is shorter because the neoclouds do not claim the other cells. They claim price and speed on a GPU generation. That is a real product. It is the product the build piece already sorted as borrowed capacity with no cash engine to retreat to. Microsoft and Alphabet sit in this cohort as the control: they are the other places an enterprise already dual-runs a record. Dual-running Azure and AWS is common. Dual-running AWS and CoreWeave as the identity plane is not a sentence anyone in coverage has a print for.

The other side of the store half is not a straw man, and it is not CoreWeave. It is Shopify, and a reader who thinks Muse is how Amazon's ads line dies is reading a real architecture, not a headline. Meta's own September 8 post names Shop Pay as coming, next to Link by Stripe. That is an invitation. Shopify's transaction-embedding and network pillars are the ones the framework rates as strong because the merchant already made Shopify the checkout, the wallet, and the app store — and the company has spent a year telling anyone who will listen that 'Commerce for Agents' is the point of the platform. An agent that can see the catalog, compare, and pay on Shop Pay never needs amazon.com. If that path takes the reorders that today hit 1-Click, the ads line is not a protected exhaust. It is a residual of a storefront people stopped opening.

Amazon has already conceded the mechanism in the other direction. Buy for Me, as The Register notes, sends Amazon's own agent onto other retailers' sites, identifies itself, and lets those retailers opt out. That is the same opt-in rule Amazon is now enforcing against Muse, pointed outward. A reader who calls that hypocrisy is not confused. They are noticing that Amazon wants to be the agent that skips other people's storefronts and the storefront that other people's agents cannot skip. If the category standard becomes the Shopify invitation rather than the Amazon wall — if a federal appeals court that already lifted a Perplexity bar in August keeps lifting them — then the wall is a delay, not a moat, and the learned-interface cell is not the only one that moves. Discovery leaving is how you would see it: advertising growth rolling over while third-party seller services hold, because the seller still needs FBA and the buyer no longer needs the search box.

Walmart is the physical version of the same warning. The framework already writes the app down and the store grid up. An agent that shops around can still have to clear 10,900 stores for groceries that have to arrive in hours. That is why Walmart belongs on this page as a control, not as a victim. It is also why the Amazon store half is not automatically safe: the items that do not need a same-day node — the SKU a Shop Pay merchant can ship from their own warehouse — are the ones an agent can take with it. The MAG7 piece treated the warehouse as the increment that Microsoft does not have. The increment only applies to the SKUs that use it.

The cloud half of the other side is Microsoft, and it is the same honest counter-case the MAG7 piece already ran. Active Directory is a worse thing to leave than an S3 bucket. Enterprises already multi-cloud the workload. If the next wave of training and inference is what Azure and the neoclouds split between them, AWS can keep the record and still lose the growth that is paying for the $220B capex year. CoreWeave's own chief executive called Q2 an inflection toward enterprise, not only frontier labs. Take that sentence at full strength. A specialist that starts as overflow and ends as the default training cloud for the Fortune 500 is exactly the path AWS itself ran from 2006. The print that would make that true is not a $2.6B quarter. It is an enterprise identity story CoreWeave and Nebius do not yet have, and that Microsoft already does. Anyone who wants to invert this page has to argue that GPU overflow becomes the record. That argument is coherent. It is also the argument the neoclouds have to prove, not the one the Q2 scale table currently supports.

This page is not a claim that Amazon is cheap, and it is not a claim that Muse will fail. The Muse note already said the agent is two weeks into a habit test and has no named revenue line; nothing in a weekend block changes that. The MAG7 piece already said durability is the argument and the price is a different page. The job here is narrower. The market took a locked door and a pair of GPU landlords and sorted Amazon as a company whose moat is slowly becoming a service anyone can offer. The five-pillar grid sorts the same facts as a weak interface being policed and a cloud record still compounding, with overflow sitting in the cells that have no record to lose.

The lesson generalises past this weekend. When a company locks a door, check whether the door is on the fortress or on the room it already knew was a hallway. Amazon locked the storefront. Muse still has to call the warehouse if the item is an Amazon item. CoreWeave still has to call a hyperscaler account when the job is more than a rack. The MAG7 factor hears a threat in any new product that says it can do the same job. The skip test asks whether the job it can do is the one that was ever the moat.