AWS will embed its engineers inside your company. The model was never the moat — this is.

Amazon announced on the last day of June that it will spend a billion dollars to put its own engineers inside other people's companies. The framing was partnership, as it always is: a new AWS organization of "forward-deployed engineers" who embed with customers for weeks at a time, sit beside their business, engineering, and security teams, and build production AI systems alongside them. The named early customers are a roll-call of institutions that are not short of money or talent — the NFL, the NBA, Cox Automotive, Southwest Airlines, Ricoh, the Allen Institute. The stated goal is to turn those customers into self-sufficient operators of their own AI. It is a generous-sounding program, and it is one of the sharpest power moves the cloud industry has made in years. Both of those things are true, and only the second one is the story.

Here is the thesis, stated plainly so you can argue with it. For three years the entire contest in artificial intelligence has been narrated as a contest over models — who has the biggest, the smartest, the safest. That contest is real, but it has quietly stopped being the one that determines power, because the model layer is commoditizing in front of us. Frontier capability leaks, gets cloned, gets matched by open weights within months, and gets cheaper every quarter. What does not commoditize is the thing that turns a model into a system a company actually runs: the deployment, the integration, the human relationship wired into the customer's core workflow. AWS just bet a billion dollars that the durable power in AI is not the model. It's the engineer standing in your building, building the model into the load-bearing walls.

The Palantir playbook, finally industry-wide

None of this is new as a tactic. Palantir has run forward-deployed engineers for well over a decade, and the reason it became one of the most quietly entrenched companies in enterprise software is precisely that its engineers don't sell you a tool and leave — they move in, learn your operation better than your own staff document it, and build something you cannot easily run without them. What's new is everyone adopting it at once. Anthropic does it. Google Cloud does it. Salesforce does it. And now AWS is industrializing it with a billion-dollar budget and, by its own description, thousands of engineers organized into pods that deploy for roughly six-week stretches. When every major AI vendor independently arrives at the same model — put our humans inside the customer — that convergence is telling you where the moat moved. It moved off the slide deck about benchmarks and into the org chart of the customer.

Read the official language and the power structure is right there, undisguised. AWS describes a progression in which the customer's teams go from "observers to co-builders to autonomous operators," finishing the engagement with deployed systems, documentation, and "trained internal champions." Every word of that is true and every word of it is the mechanism. The systems are built on AWS's stack. The documentation describes AWS's patterns. The "champions" are employees of the customer who are now fluent, invested, and professionally identified with the AWS way of doing AI. The program graduates customers into self-sufficiency the way a company town graduates workers into home ownership: the house is real, and it's still the company's town.

The program turns customers into self-sufficient operators the way a company town turns workers into homeowners. The house is real. It's still the company's town.

Why a billion dollars, and why now

Follow the money's logic and the timing explains itself. A billion dollars of senior engineering labor, given away inside customer accounts, is not a services business that needs to make its own margin — the economics are far too thin for that. It is a customer-acquisition and customer-retention cost, paid in the most expensive currency AWS has, deployed against the one problem that actually threatens the cloud's AI revenue: the adoption gap. The models are ready; enterprises mostly are not. There is, in AWS's own words, "a ton of demand" from customers asking for help to drive agentic AI patterns into their workflows, and that sentence is the whole strategy. The bottleneck on AI spending is no longer capability. It's the customer's inability to turn capability into production. Whoever closes that gap captures the spending behind it — and captures it on their own platform, in their own services, with their own people holding the institutional memory of how it was built.

This is why the program targets customers who could plainly afford to hire their own engineers. The NFL is not strapped for cash. The point was never to subsidize the poor; it was to get inside the rich, because that's where the durable, high-volume AI workloads — and the multi-year compute commitments under them — are going to live. An AWS engineer who builds the NFL's production AI systems has done something no marketing budget can buy: made AWS's architecture the path of least resistance for every subsequent decision that institution makes about AI. The next workload doesn't go to tender. It goes to the stack the champions already know.

The strongest counter-argument, which nearly persuades me

Let me give the rebuttal its full weight, because it is good. Enterprises genuinely cannot hire enough AI engineers; the talent does not exist in the volume the moment demands. Forward-deployed engineers deliver real, measurable value — the NFL's own CIO says his teams launched into production in weeks rather than the quarters it would otherwise take. The knowledge transfer is real; some customers really do come out the other side more capable. And there is nothing sinister about a vendor being good at helping customers succeed on its platform; that is, in a sense, the entire history of enterprise computing. If the work is excellent and the customer wins, who exactly is harmed? This is close to the argument that, a few years ago, made me too sanguine about open-source models as a check on concentration. I assumed that if the capability diffused widely enough, the power would diffuse with it. I was wrong, and this is the shape of why.

Here is the answer the rebuttal doesn't survive. Capability diffusing and power diffusing are not the same event, and they have just decoupled in plain sight. The models did democratize — they are cheaper, more open, and more abundant than the gloomiest forecasts predicted. And in exactly the same window, the ability to deploy them at scale inside a serious institution re-concentrated, because that ability turns out to require something the open weights never touched: thousands of senior engineers you can afford to give away. Open-source answered "who can have a model." It did nothing about "who can put a thousand engineers in the field for free," and that second question is now the one that allocates the power. The barrier didn't fall. It moved to a place open-source can't reach.

Self-sufficiency in whose house

The tell in the whole program is the word "self-sufficient," because it concedes the thing it's meant to obscure. Self-sufficiency is offered as the customer's win — and it is a real one, at the operational level. But self-sufficiency on a platform is not independence from the platform; it is fluency in it. A customer who emerges from an AWS engagement able to run its own agentic systems is a customer who can now run more of them, faster, all of them on AWS, staffed by people whose hard-won expertise is non-portable to a competitor's stack. The more capable the customer becomes, the more deeply specified to AWS its capability is. That is not a bug the program tolerates. It is the product. The deepest lock-in has never been the contract you can't leave; it's the competence you can only exercise in one place.

And this is what makes the forward-deployed model a more complete form of power than the one I usually write about. The compute was always the hard floor — own the chips, the data centers, the power contracts, and you own the terms underneath everyone else's ambitions. But compute is a landlord's power: necessary, extractive, and at some distance from what the tenant actually does day to day. Putting your engineers inside the customer closes that distance. It pairs the landlord's leverage over the building with a foreman's presence in every room, shaping not just where the AI runs but how the customer thinks about building it at all. The chips were the floor. The engineers are the walls and the wiring, and they were installed by the landlord, to the landlord's spec, by people the tenant now calls their own.

What to watch, and what it means

So watch the convergence, not the press release. When AWS, Anthropic, Google Cloud, Salesforce, and the original in Palantir all decide that the way to win AI is to occupy the customer with human beings, the era of the model as the trophy is ending and the era of the deployment as the moat has begun. A few things follow from that, and none of them are in the announcement:

• The advantage accrues to scale, not cleverness. Only a handful of companies can afford to field thousands of senior engineers as a loss leader, which means the deployment moat is narrower and better-capitalized than the model race it's replacing.
• Cheaper and more open models do not loosen this grip; they tighten it. The more commoditized the model, the more the differentiation — and the margin — shifts to the party that can install it, and installation is where the concentration now lives.
• "Self-sufficiency" is the metric to distrust. Ask self-sufficient on whose stack, portable to whom, staffed by people whose skills transfer where. The answers describe the lock-in.
• The customers least worried about this are the ones already deepest in it. The institutions taking the free engineers first are buying a real head start and a real dependency in the same transaction, and only one of those shows up this quarter.

I'll end where the announcement won't. A billion dollars given away is never charity; it is a price paid for something worth more than a billion dollars, and what it buys here is position — the position of the company whose people built the thing, whose patterns the customer now thinks in, whose platform the next decision defaults to. The model was always going to commoditize. The smartest players in AI have already accepted that and moved a layer up, to the place where capability becomes dependence: the engineer in your building, billed to no one, who will be very hard to ever fully replace. That's not a partnership. It's where the power went while everyone was still arguing about whose model is smartest.