Amazon’s $1B FDE Push Says AI Software Is Still a Services Business
Amazon is putting $1 billion behind engineers who sit inside customers’ businesses because selling AI software alone clearly isn’t cutting it.
Amazon is putting $1 billion behind engineers who sit inside customers’ businesses because selling AI software alone clearly isn’t cutting it.
That is the bit everyone pretending AI is a clean software-margin miracle should pay attention to.
On June 30, Amazon Web Services launched a new organisation of AI-focused forward-deployed engineers, or FDEs. These are not salespeople with a slide deck and a LinkedIn addiction. They are technical operators embedded with customers to build and deploy AI agents in the messy reality of actual companies.
AWS says it is committing $1 billion of internal resources to the effort. That is not a venture round or a flashy joint venture. It is Amazon spending real money to put capable people into the field because customers need help turning model demos into useful systems.
That is the story.
Not another chatbot. Not another benchmark. Not another bloke on X announcing that a model can now solve a puzzle nobody in business has ever been paid to solve.
The important admission is that AI’s bottleneck is no longer access to intelligence. It is implementation.
Amazon is paying for the unglamorous bit
For years, the AI pitch has sounded wonderfully simple: plug in a model, point it at your data, cut costs, grow revenue and let the robots do the boring work.
Lovely. Except businesses are not plug-and-play.
Their data is scattered across ancient systems. Their approval processes are built around humans. Their security teams say no for good reason. Their customer-service operation has twenty years of exceptions. Their best workers have knowledge stuck in their heads, their inboxes and a spreadsheet called FINAL_v18_USE_THIS_ONE.xlsx.
An AI model does not fix that by itself.
AWS’s new FDE group is built around a more honest proposition: send engineers into the customer, build purpose-made agent systems, move quickly, then leave the customer with systems and skills it can operate itself.
The forward-deployed model was popularised by Palantir, which figured out something plenty of software companies still resist: buyers do not pay for software because it is clever. They pay because it produces an outcome in their particular, inconvenient business.
That often requires someone from the vendor to get close to the work. Close enough to understand the data, the workflow, the politics, the failure points and the exact moment a human needs to take control.
Amazon is not alone in spotting this. OpenAI and Anthropic have each created forward-deployment ventures with private-equity partners, reportedly valued at $4 billion and $1.5 billion respectively. The labs are learning the same lesson: enterprise AI revenue is not won merely by having the best model. It is won by getting the thing into production before the customer loses patience.
That is much harder. It is also much more valuable.
The $1 billion is not generosity. It is defence.
Amazon’s cloud business is enormous. In the first quarter of 2026, AWS revenue rose 28% year on year to $37.6 billion, its quickest growth in 15 quarters. AI demand is a major reason.
But cloud infrastructure has a nasty feature: it can become interchangeable if the customer believes another provider can deliver the same compute, models and tools.
Amazon therefore needs more than servers. It needs to become part of the operating machinery of customers building AI products.
An FDE embedded with a bank, retailer, manufacturer or insurer is not merely helping that business ship an agent. That engineer is helping design the customer’s architecture, data flows, guardrails, cloud usage and habits around AWS.
Once that work is done well, switching clouds is no longer a procurement exercise. It becomes a painful operational rewrite.
That is why the FDE model matters. It turns cloud infrastructure from a commodity into a working relationship.
Amazon has form here. It has been building a broader AI stack: custom Trainium chips, Bedrock model services, a deep partnership with Anthropic and major work with OpenAI. AWS has said it agreed to supply OpenAI with 2 gigawatts of Trainium computing capacity as part of a $50 billion investment arrangement.
That is a lot of metal and electricity. But hardware capacity only earns its keep when customers build workloads that stay on it.
The FDE push is how Amazon tries to make that happen.
Here is the overlooked angle: AI may create a better kind of services company
Most investors hear “services” and reach for the smelling salts.
Fair enough. Traditional consulting can be a wonderful business until it becomes a people-heavy machine that grows revenue one expensive head at a time. Margins get squeezed. Delivery gets inconsistent. The client becomes dependent. Everyone agrees to “circle back” and somehow nothing improves.
But AI-forward deployment can be different if it is run properly.
The first implementation may be intensely hands-on. Yet the useful pieces can be repeated: connectors, security patterns, evaluation systems, workflow templates, monitoring, permissions models and playbooks for human escalation.
That is the real asset being built. Not the first custom agent. The implementation factory behind the next hundred.
AWS is explicitly framing these engagements around customer self-sufficiency. That matters. If Amazon simply creates permanent dependence on a platoon of expensive engineers, it has created a consulting business wearing an AI hat.
If it leaves behind reusable systems that make each next deployment faster, safer and cheaper, it has created something much better: a services-led wedge into software-like economics.
There is still a risk. Plenty of companies will confuse bespoke work with product development. They will build one-off monstrosities for big customers, call them “platforms,” and quietly discover they are running a body shop.
The winners will be ruthless about what gets standardised. They will identify the 20% of implementation work that repeats across customers and turn it into tooling. They will say no to custom requests that do not create reusable capability. And they will measure time-to-value like their business depends on it.
Because it does.
Why this is bad news for lazy AI startups
If Amazon, OpenAI and Anthropic are all putting serious money into deployment teams, the middle of the market has a problem.
A startup cannot win enterprise deals now by arriving with a polished interface, a few model-provider logos and a claim that “we integrate with everything.” Customers have heard that one before.
The new standard is: can you make this work inside my business, with my systems, under my compliance rules, and show me a measurable result before the budget owner gets bored?
That does not mean every founder should hire an army of consultants. Quite the opposite. It means founders must be fanatical about reducing implementation pain.
Build for ugly data. Build for permissioning. Build for audit trails. Build for humans to override the machine. Build a narrow wedge that can prove value in weeks, not a grand transformation that needs twelve workshops and an executive steering committee.
Snowflake’s recent $6 billion, five-year AWS agreement offers a clue about where this goes. As AI shifts from model training into day-to-day use and agentic workflows, demand spreads beyond GPUs into the plumbing around them, including CPU capacity. The value is not just in training a model once. It is in running useful systems repeatedly, reliably and at tolerable cost.
That favours operators who can make AI boring enough to trust.
Boring is underrated. Boring pays.
What this means for you
If you run a business, stop asking your team, “How are we using AI?” It is a rubbish question. It invites theatre.
Ask these instead:
1. Which workflow costs us real money every week? Pick one with volume, repetition and a measurable failure rate. Customer onboarding, proposal drafting, claims triage, invoice chasing and internal support are better starting points than “an AI strategy.”
2. What must be true for an agent to act safely? List the data it needs, the systems it must touch, the decisions it may make and the exact points where a human must approve. If you cannot answer that, you are not ready to deploy anything meaningful.
3. Can we prove value in 30 days? Define a number before you build: hours saved, conversion lifted, errors reduced, revenue recovered or response time cut. If nobody owns the metric, the project will become a very expensive demo.
4. What can we reuse after this first win? Every AI project should leave behind an asset: a clean data connection, an evaluation set, a permissions model, a workflow template or a documented operating rule. Otherwise you are paying tuition repeatedly.
5. Do we need outside operators, not just outside software? This is the uncomfortable one. Sometimes the fastest path is bringing in people who have shipped the sort of system you need. Just make sure the agreement includes knowledge transfer and a clean exit plan.
Amazon’s $1 billion bet is not proof that AI has failed. It is proof that the easy part is over.
The next fortunes in AI will not go only to the companies that make the smartest models. They will go to the ones that can drag those models through the mud of real operations and make them earn their keep.
That is where the money has always been.