Three of the world’s largest AI vendors have spent the past ten days admitting something enterprise buyers have suspected for a while: the model was never the hard part.
On 30 June 2026, AWS committed $1 billion to a new Forward Deployed Engineering unit, sending pods of engineers directly into customer organisations to build and deploy agentic AI systems on-site. Three days later, Microsoft answered with Microsoft Frontier Company, a $2.5 billion commitment embedding 6,000 industry and engineering experts inside client organisations to co-design, deploy, and run AI systems against measured business outcomes. Combined, that is 3.5 billion dollars committed by two vendors in under two weeks, and neither of them spent a cent of it on a better model.
They spent it on people, sent to sit inside your organisation and do the work your own team was supposed to already be doing.
This Is Not Two Companies. It Is a Pattern.
Treat this as an isolated Microsoft-versus-Amazon story and you miss what is actually happening. Both moves followed a pattern already set earlier in 2026 by the AI labs themselves. Anthropic and OpenAI both launched joint ventures for enterprise AI deployment on the same day, 4 May 2026. Anthropic’s is a $1.5 billion venture backed by Blackstone, Hellman & Friedman, and Goldman Sachs. OpenAI’s is The Deployment Company, a $10 billion vehicle anchored by TPG. The technique itself, forward-deployed engineering, sending a vendor’s own technical staff to embed inside a customer’s operations rather than selling software and walking away, was not invented in 2026 either. Palantir built its entire early growth on exactly this model more than a decade ago.
What changed in the space of two months is who is now doing it. Every major AI vendor, model builders and cloud hyperscalers alike, has independently reached the same conclusion at the same time: licensing the technology and leaving customers to figure out deployment is no longer a viable strategy for demonstrating that AI investment produces returns.
Why Now, and Why All at Once
The timing is not a coincidence, and the reason is uncomfortable for anyone who has spent the last two years running an internal AI programme on the assumption that the tooling was the hard part.
MIT’s Project NANDA research, based on 150 leadership interviews, a survey of 350 employees, and an analysis of 300 public AI deployments, found that 95% of organisations deploying generative AI saw zero measurable business return, despite an estimated 30 to 40 billion dollars in enterprise investment. The same research found that internal builds succeed at roughly a third of the rate of purchased tools paired with a genuine implementation partnership, and that the deployments which did work shared one trait: ownership sat with the domain leaders actually running the process, not with a centralised AI lab several layers removed from where the work happens.
That is the number every AI vendor is now responding to. A 95% pilot failure rate cannot be fixed by shipping a better model. It is an execution problem, and for the first time, the vendors are the ones saying so, with their own balance sheets rather than a slide in a sales deck.
What 3.5 Billion Dollars of Vendor Behaviour Actually Tells You
If AWS and Microsoft believed their own customers could close this gap with the tools already on the market, they would not be spending a combined 3.5 billion dollars putting engineers on the ground to do it for them. Vendors do not fund headcount at this scale to solve a problem their existing product already solves.
That is the signal worth sitting with if you are running, sponsoring, or governing an AI programme right now. The two organisations with the clearest commercial incentive to tell you that your existing licence is sufficient are instead telling you, with 3.5 billion dollars of capital allocation, that it is not.
What This Means for Your Own Programme
None of this means the answer is to wait for a vendor’s forward-deployed team to arrive and do the work instead of building the capability internally. Vendor-embedded engineers close the gap for as long as they are in the building, and then they leave, taking the capability with them unless the organisation has built something durable underneath it.
What it does mean is that the excuse most transformation programmes have been running on, that the tooling was not yet mature enough to deliver value, is no longer available. The vendors have just spent 3.5 billion dollars telling the market that the tooling works. The 95% failure rate MIT documented was never about the model. It was about exactly the things forward-deployed engineering exists to fix: ownership sitting in the wrong place, workflows that were never redesigned around the tool, and outcomes that were never defined before the build started.
Those are governance problems, sitting inside the organisation rather than inside the platform, and vendors were never going to be the ones to fix them permanently. They can staff their way around the symptoms for the length of an engagement. Your organisation has to solve the underlying problem itself, and the sooner that distinction is made explicit at the programme level, the less it will cost to fix later.
The model was never the hard part. The vendors just spent 3.5 billion dollars confirming it.