Last week, Amazon closed its San Francisco AGI Lab, cut roles across its AGI organization, and retired roughly twenty AI products and features , while pouring a further billion dollars into its Forward Deployed Engineering (FDE) organization and raising its 2026 capital spending toward $220 billion. Read as a headline, it looks like retreat. Read as a strategy, it is a decision about which layers of the AI business it considers to be worth owning.

While many have reported on this news and analyzed it for what it means for AI, this piece focuses on a different question. What does this mean for the businesses that are buyers of Amazon (and its competitors’) AI services? What should they take away from this development for their own AI strategy?

For a detailed analysis of what Amazon did (and an interpretation I personally find very sound), I would direct you to a piece by Julien Simon of Hugging Face, “ Amazon Priced the Frontier and Declined It ”. A few observations:

  • Amazon appears to have made a conscious choice to retain and invest in the metering, customer-facing, and hardware-adjacent layers of the stack, where revenue can be made regardless of whose model is being run (theirs, Anthropic’s, OpenAI’s, Microsoft's, Google’s, etc.)
  • The layers of the stack close to the hardware were preserved and appear to be getting increasing attention.
  • They have maintained a small effort in their model research, and continue to provide Nova, but these appear to be hedge bets, and the organizational structure suggests they are intended to stay that way)
  • They continue to provide their own models, but this is likely to be to force price pressure rather than to attempt to get market share from other model providers.

Since this article was written, two notable announcements support this direction:

  • Amazon’s stock popped after their earnings announcement , suggesting the market approves.
  • Friday’s announcement of DeepSeek’s new bargain model further adds price pressure to the frontier model layer, the very layer Amazon appears to have chosen to exit.

What Does This Mean - For The Buyer?

Most companies do not compete with Amazon; they do, however, buy AI infrastructure from it or its competitors. While the move is strategically interesting for the AI industry, it is also important for buyers to note. We outline 5 implications.

1. Multi-model, yes, multi-infrastructure, maybe not. Amazon has consciously decided to offer other frontier models in their infrastructure. Given the widely shifting model cost and capability landscapes, not to mention soaring AI bills in businesses , staying switchable at the model layer is a wise move. Infrastructure lock-in, on the other hand, may be unavoidable. As many businesses learned with multi-cloud strategies, multi-vendor infrastructure can be impractical and expensive. However, both should be conscious choices, pricing the dependency rather than sliding into it. Amazon bets that the model layer is now a commodity; the buyer's takeaway is that the model is an ingredient, not the dish.

2. Would you notice if they left? For each layer of your stack, ask whether a single vendor exiting would cripple you. For an analogy, consider jam in a grocery store. If a jam vendor chose to exit the market, would you notice? Would you change anything except pick up another jam from the same aisle? On the other hand, if your grocery store went out of business, you would likely need to change where you drive to get groceries. Rent the layers where the answer is no, and plan your exposure and exit for the layers where the answer is yes.

3. Capability must live within. Amazon is not the only vendor investing heavily in FDE. Anthropic and others are doing the same, following a trend long established by Palantir. The thesis is that the vendor’s engineers will sit in your offices, with your own people, helping build a solution which you will care for once they exit. This can be a very useful thing, but as recent events with the OpenAI model attack on Hugging Face have shown, there is no equivalent for having capability in-house. The judgment that you will need to deploy of when to switch, what to run, how to secure across models, how to stand up a Kimi- or DeepSeek-class model when you need to, cannot be outsourced, because it fails you exactly when the vendor's interests diverge from yours. Help is procurable; capability is not.

4. Fit-to-task. Amazon has bet that the frontier model is now a commodity. Going back to our jam analogy, you may buy garlic paste from the same grocery store as you do jam. The same bread uses garlic paste when you want garlic bread, and jam when you want to serve breakfast. Last week’s arrival of the new DeepSeek model reinforces the need to maintain a deliberate portfolio (generalist + domain-specific + open-weight) of model resources (ideally within the same infrastructure and build the organizational judgment to match the model to the required task. Anyone can buy the jars of garlic paste and jam; knowing which to spread is the capability that the cook holds. This judgment should stay within your capabilities and be cultivated inside your organization.

5. Corporate Taste as the top layer. If the model layer commoditizes and the infrastructure layer is a landlord’s game you mostly rent, the only durable moat left for a normal enterprise is the layer above: how you use the models, which is judgment, which is Corporate Taste . Notice the division of labor across these five points. The first four are defensive. They keep you from overpaying, over-committing, or getting locked in; they manage cost and preserve optionality. Corporate Taste is the only one that plays offense: it is the layer where judgment about your domain, your customers, and what not to build turns rented tools into an advantage no competitor can buy. Everything else on this list protects your investment; this is the one that drives your return. What does your organization understand about your domain, your customers, their needs and wants? What have you learned about what not to do? For the only non-negotiable, ROI, you need both the defense and the offense.

Amazon appears to have decided to focus on the grocery store and not the jars within. As a customer, this is interesting, but the real strategy question is what menus you make from wide varieties of jars, not which grocery store you bought what from.