For fifteen years the argument about Palantir was whether it counted as a software company at all, or was really a consulting shop wearing software margins. Bank of America once ran a note saying it was misunderstood on Wall Street, where analysts had it filed as a government contractor. That argument is still live, and resurfaced in the trade press again last month.
On 3 August the company reported revenue of $1.94bn against $1.80bn expected, up 93% on the year, with US commercial revenue up 149%. The stock rose 12%.
Six of the largest companies in technology have spent the last eighteen months taking a page from its book.
What Palantir worked out
The model started at Palantir in the early 2010s, where it was called Delta before it had a name anyone recognised. They put engineers inside customer organisations because their intelligence customers could not always say what they needed. You had to go and look.
The detail I find most telling sits in Palantir’s own 2020 listing prospectus, which describes its forward deployed engineers as the first line in identifying research and development opportunities for its platforms. Watching customers work was the research method. The product came back from the field.
Their founding view was that the failure they existed to fix was never a shortage of data or algorithms. It was an inability to join information across systems that were never designed to talk to each other. Fifteen years on, that is still the first problem in almost every business I look at.
The last eighteen months
OpenAI announced Frontier Alliances in February, pairing its own forward deployed engineers with McKinsey and BCG on strategy and Accenture and Capgemini on implementation. In May it raised over $4bn for a majority-owned joint venture valued at $10bn before the money, and bought Tomoro, a London consultancy, to get delivery capacity faster than it could hire it. Anthropic launched Ode in July with Blackstone and Hellman & Friedman, built on the acquired Fractional AI team. Amazon has committed $1bn to a dedicated AWS forward deployed engineering organisation. Microsoft stood up a Frontier Company business unit. ServiceNow and Accenture launched a joint programme.
Constellation Research expects more than 85% of technology providers to be running an equivalent programme by the end of this year. The Financial Times reported in November 2025 that monthly job listings for the role rose more than 800% between January and September that year, using Indeed postings data.
The capital behind it is on a scale that is hard to picture. Amazon guided to $220bn of capital spending for 2026, Alphabet to between $195bn and $205bn and Meta to between $130bn and $145bn, all at their second quarter results in July 2026. With Microsoft added, the four are on course for roughly $750bn this year against about $410bn between them in 2025, and credible estimates of the total run from about $700bn to $900bn depending on who you count. People disagree, loudly, about whether that pace is sustainable. That argument belongs to shareholders.
What interests me is what all that money is buying. Alongside the data centres, it is buying a method, and that method is being written down in public as they go.
Why it went this way
For most of the last decade, implementation work was something software companies tried to engineer away. Every hour a person spent configuring your product was an hour that hurt the margin story, so the goal was always to make the thing self-serve.
AI broke that. A model is general and the value it produces is not. Getting anything useful out of one depends on a specific company’s data, its processes, its exceptions, and the several things everybody in the building knows but nobody wrote down. None of that ships in a box.
The clearest description of the problem I have read came from Colin Jarvis, who runs forward deployed engineering at OpenAI: what the customer describes in scoping does not match the data and system reality on the ground. Closing that gap is the entire job.
Worth noting what their own process does about it. There are three phases, and the middle one is building evals rather than running a demo. What comes out is a test set that tells you whether the thing works. I have been arguing for that as a contractual deliverable for a while, and it is quietly reassuring to see the people with the best view of the technology arrive at the same place.
We have seen a build-out like this before
In the late nineties, telecoms companies laid around 45 million miles of fibre optic cable on the conviction that internet demand would arrive faster than it did. By 2007, about two thirds of it was still dark. Unused.
At the time there was no shortage of people predicting that the whole thing would collapse, and on share prices a lot of them were right. What almost everyone missed was the second half of the story. The cable stayed in the ground. It was bought cheaply, consolidated into a handful of networks, and went on to carry the cloud, streaming and mobile era that arrived roughly a decade later than the prospectuses promised. Google, Microsoft and Meta eventually built on that surplus and bypassed the carriers who had paid for it.
The internet was never a fad. It did not happen on the schedule its backers sold. The names above the door changed. What had been built stayed.
The obvious objection, which is a good one
A GPU is not a length of fibre, and anyone drawing this comparison should deal with that honestly.
Fibre and the buildings it runs into have useful lives measured in decades, commonly put at 30 to 50 years. Nvidia’s chips have a shelf life of perhaps five, and there is real disagreement about whether the right number is three, five or seven, which matters enormously to anyone financing them. Operators generally hope to earn the cost back inside one or two years. That is a completely different asset.
So the narrow version of the point is the honest one. Land, power connections, substations, fibre and buildings hold their value across decades. The chips depreciate on a three-year cycle whether or not anyone wants them.
The part that actually compounds
Which brings me to what I think is being built here that nobody is putting on a balance sheet.
Thousands of engineers are currently learning how to make this technology work inside real organisations, against real data, with all the exceptions and undocumented rules that live in every business. A generation of companies is being taught what good looks like, what a proper evaluation set is, and how to tell a working system from a convincing demo.
That knowledge does not depreciate. It does not go bankrupt if a valuation corrects. It walks out of the building in people’s heads and turns up somewhere else, which is exactly what happened after 2000. The capital was destroyed and the capability was not. The people who learned to build for the web during the boom went on to build everything that came after it, on infrastructure somebody else had already paid for.
Why that should change what you do now
The sensible-sounding response to all this is to wait and see how it shakes out. I think that is the wrong call, and the reason is that the two questions have nothing to do with each other.
Whether Nvidia is correctly valued is a question about share prices. Whether a system that drafts your quotes correctly saves you a day a week is a question about your business. The second answer does not move when the first one does.
And the method being proven at the top of this market is not proprietary. Someone technical goes into a business. They find where the value actually is, which is almost never where the demo suggested. They build against real data and the awkward cases rather than the tidy ones. They measure against a baseline written down before they started. Then they stay long enough to fix what breaks on contact with reality.
None of those steps depends on scale. All of them are being validated right now at enormous expense, on somebody else’s budget, and written up in public as they go.
The businesses that came out of the last build-out ahead were not the ones that predicted the crash. They were the ones who learned how to use what got built.
Where I would start is smaller than most people expect. One repetitive task, one number written down before you begin. Our guide to finding your first real AI use case covers choosing it, and how to tell whether an AI system actually works covers the measuring. If your data sits in four systems that cannot reach each other, that is a plumbing problem and it comes first.
Frequently asked questions
What is a forward deployed engineer?
An engineer embedded inside a client’s organisation who builds working systems in that client’s own environment and owns whether the problem got solved. The distinction from a consultant is accountability for the outcome rather than for the recommendation.
Why are AI companies buying consultancies?
Delivery capacity is the constraint and hiring it takes too long. OpenAI acquired Tomoro to staff its Deployment Company, and Ode was built on the acquired Fractional AI team. Both bought people who already knew how to work inside somebody else’s business.
What happened to the infrastructure after the dot-com crash?
Telecoms firms laid around 45 million miles of fibre and roughly two thirds was still unused by 2007. Most of the companies that built it failed. The cable was bought cheaply, consolidated, and went on to carry the cloud and streaming era.
Can a smaller business use this approach?
Yes. The ventures named here target mid-size companies and above, but the method does not depend on scale. Mapping a task, building against real data, measuring against a baseline and staying to fix what breaks work at any size.
Jamie Francis is the founder of Flux Dynamics, a fractional CTO who builds. Tell us what you are weighing up.