The Forward-Deployed Engineer Goes Private Equity

The interesting thing about Anthropic's $1.5bn services vehicle, announced on 4 May with Blackstone, Hellman & Friedman, and Goldman Sachs, is not that the model labs are moving into services. That was always going to happen. The interesting thing is who is funding it, and what that tells you about how the people closest to the cash flows of mid-market software actually expect the next five years to play out.

OpenAI's parallel venture with TPG and Bain Capital, announced the same week, is not a coincidence. It is two of the largest private equity complexes in the world deciding, in close succession, that the right way to extract returns from generative AI is not to back another model lab and not to back another SaaS rollup, but to fund a forward-deployed engineering practice that sits inside the customer and bills against outcomes. Palantir spent fifteen years proving the model could work. Anthropic and OpenAI are now rolling it out with private-equity-grade capital behind it, on a compressed timeline, against a customer base, mid-market American companies between $100m and $2bn in revenue, that has historically been the dominant grazing ground of the big consulting and systems-integration firms.

I think this is the most consequential structural move in enterprise AI deployment since the API itself. I also think it has been almost entirely misread, by the people I would expect to read it most carefully, as a story about Anthropic and OpenAI competing for enterprise share. It is not. It is a story about private equity deciding that the consulting layer is the part of the stack where the margin is going to land, and moving to own it before the incumbents can defend it.

The interesting question is not whether the labs are moving into services. It is why two of the largest PE complexes in the world decided, in the same week, that this was the trade.

Start with what an FDE actually is. Palantir's forward-deployed engineer is not a consultant in the McKinsey sense and not a systems integrator in the Accenture sense. The role is closer to an embedded product engineer who happens to work for the vendor rather than the customer. They sit inside the client, they have commit access to the platform, they ship code against the customer's actual workflows, and they are measured on whether the deployment produces the outcome the customer paid for. Palantir's gross margins on Foundry, around 80% at the platform level, with services running closer to 30%, work because the FDE creates the conditions under which the platform can be billed at software economics. The services exist to make the software work. Without the services, the software does not stick; with them, the software becomes load-bearing inside the customer.

This is exactly the problem Anthropic and OpenAI have. Their models are extraordinary, and their enterprise sell-through is, by any honest reading of the public commentary from their own customers, mediocre. The story of 2025 was that capability ran ahead of deployment. Mid-market companies bought seats, ran pilots, found the pilots interesting, and could not get from pilot to production without an enormous amount of bespoke work that neither the lab nor the customer was set up to do. The labs were not staffed for it; the customers did not have the engineering depth; the incumbent consultancies were quoting it but, in the candid assessments I have read from people running these programmes, were largely reselling junior bodies against problems that needed senior product judgement.

The FDE vehicle is the response. Hire the senior product judgement, embed it inside the customer at the lab's expense, take the deployment risk on the lab's balance sheet, and bill against whatever value framework the customer will accept, usually a mix of platform fees and outcome-linked components. Palantir's commercial business has been growing at roughly 50% year-on-year on this model. Anthropic and OpenAI looked at that number and concluded, correctly, that the binding constraint on their enterprise revenue is not the model and not the API, it is the absence of a deployment layer that can carry the customer from interest to production.

So why private equity, and why now? This is the part I think most of the commentary has missed. The labs could have built FDE practices on their own balance sheets. Anthropic raised at $61.5bn in March; OpenAI is sitting on cash at a scale that makes a $1.5bn services build-out look like a rounding error. They did not need outside capital to do this. They chose to take it.

The choice tells you what the labs think the services business actually is. If you believe the FDE practice is a margin-thin support function, a cost of goods sold for the platform, you keep it on your own balance sheet, run it as close to break-even as you can, and use it to make the platform stickier. That is, broadly, how Palantir runs its services line. If, on the other hand, you believe the FDE practice is itself a large standalone business, that the deployment work, valued at the prices the market will actually pay, is worth tens of billions of dollars over the next decade, then you put it in a separate vehicle, you take outside capital that is hungry for that specific kind of return, and you let the vehicle scale faster than the lab itself could politically justify.

The presence of Blackstone and TPG in these structures tells you the labs have taken the second view. Blackstone does not show up to fund a cost centre. Its private equity arm has been consistent, over the past three years, about where it sees mid-market software returns coming from: not from buying SaaS companies at fifteen times revenue and praying, but from buying or building services capacity that can be billed against transformation work the buyer has already committed to. The Blackstone thesis on AI services, which their portfolio team has articulated with unusual clarity in their last two LP letters, is that the consulting layer is going to consolidate hard over the next five years, that the consolidator will look more like Palantir than Accenture, and that the entry point is now.

Hellman & Friedman's involvement on the Anthropic side is the same trade from a different angle. H&F has been the most disciplined buyer of vertical software businesses in the mid-market for a decade. They know exactly what it costs to deploy enterprise software into a $500m revenue company, because they own dozens of them, and they know the deployment cost on AI-native systems is currently absurd. The bet is that an FDE practice running at scale, with the lab's models behind it, can compress that deployment cost by an order of magnitude, and that the savings show up on the FDE vehicle's books as margin, not on the customer's books as a discount.

Goldman is along for the underwriting. Bain on the OpenAI side is along for the operating expertise. None of this is unusual once you accept the premise that the FDE vehicle is being built as a billion-dollar-revenue services business in its own right, not as an accessory to the lab.

The labs did not need outside capital to build these vehicles. They chose to take it. That choice is the entire story.

What does this mean for the incumbents? This is where I think the consensus reading is most badly wrong. The dominant framing, in the few pieces I have read that engage seriously with the FDE vehicles, is that Accenture, Deloitte, and the systems integrators are about to be disrupted in a fairly conventional way: faster, cheaper, more technically competent competitor enters the market and takes share over a five-to-ten-year cycle. I do not think this is right, and I think it underestimates what is happening to the incumbents in two specific ways.

The first is that the incumbents are not losing share at the edges; they are losing the centre of their book. Accenture's AI-related bookings in calendar 2025 ran at roughly $4.6bn, which sounds large until you note that the firm's total bookings were over $80bn and that the AI line is growing at a rate slower than the underlying revenue base. The work Accenture is winning on AI is, by their own description in their last two earnings calls, predominantly advisory and integration work, the layer above the deployment, not the deployment itself. The deployment work is where the FDE vehicles are aimed, and the deployment work is where the customer's actual budget for AI transformation lives. Accenture is, in effect, being pushed into a thinner slice of the spend at exactly the moment the spend itself is migrating from advisory to deployment.

The second, more serious problem is the business model. Accenture and Deloitte bill by the hour. The hour, as a unit of professional services pricing, does not survive an FDE practice that is willing to bill against outcomes and willing to put the lab's own models behind the deployment at marginal cost. If you are a CFO at a $1bn revenue company looking at an AI transformation programme, and you have a quote from Accenture for $12m of advisory and integration work over eighteen months, billed against partner and manager rates, and a quote from the Anthropic FDE vehicle for a fixed platform fee plus an outcome-linked component, against a deployment they will run end-to-end with their own engineers, the rational choice is not even close. The Accenture quote does not survive contact with the FDE quote in any honest procurement process.

This is the part that I think the time-based-pricing question gets wrong when it is treated as a generic point about agents replacing knowledge work. The specific mechanism by which time-based pricing breaks is not that agents make consultants faster; it is that a competitor enters the market with a fundamentally different cost structure, billing against a fundamentally different unit, and the incumbent cannot match the price without abandoning the model that funds its partnership. Accenture cannot reprice its book to compete with the FDE vehicles without breaking the partner economics that hold the firm together. Deloitte is in the same position. The Big Four are worse off, because their consulting arms are structurally subsidised by audit and tax in ways that make any aggressive repricing politically impossible.

I want to be careful here, because I am close to the priors I am supposed to be testing. The strongest counter-case is that the FDE model has historically been hard to scale, that Palantir's growth, while real, has been concentrated in a relatively small number of very large accounts, and that the mid-market, which is where Anthropic and OpenAI are aiming, has historically resisted the embedded-engineer model in favour of cheaper, more transactional engagements. There is a serious version of the argument that says the FDE vehicles will work brilliantly for the top hundred mid-market accounts, run out of senior engineering talent at around three hundred, and stall out at a revenue level that is significant for the labs but not transformational for the consulting industry.

I take that seriously. I think it is probably wrong, but for a non-obvious reason: the talent constraint that has historically capped FDE-style practices is, specifically, the constraint the labs are best placed in the world to relax. An Anthropic FDE with full Claude access, working inside a $400m revenue manufacturing business, is dramatically more productive than a Palantir FDE was in 2018. The leverage that the model itself provides on the engineer's output is precisely what makes the practice scale. The talent ceiling that capped Palantir's mid-market growth is meaningfully higher when the engineer has a model behind them that can carry a substantial fraction of the deployment work.

Accenture cannot reprice its book to compete with the FDE vehicles without breaking the partner economics that hold the firm together.

The PE structure tells you the timeline. The vehicles are structured as separate entities with their own capital and their own equity, which means they are designed to scale at PE pace, not lab pace. PE pace, in services rollouts of this kind, is roughly: hire the first hundred FDEs in the first year, three hundred in the second, eight hundred in the third, and aim for a recapitalisation event somewhere between year four and year six. If both vehicles execute against that profile, and the capital is in place for them to do so, then by the end of 2028 there will be roughly two thousand lab-affiliated FDEs deployed across the American mid-market, billing somewhere between $4bn and $6bn of annualised services revenue between them, against a directly addressable market that the incumbents currently treat as core.

That is not a slow disruption. That is, on the timeline that matters for partnership economics, an extinction-level event for the parts of Accenture and Deloitte that depend on AI advisory and integration as the growth engine of the next cycle. The incumbents will not disappear. They will, I expect, retreat into the two parts of their book that the FDE model does not credibly attack: the very large enterprise accounts where the political complexity of the engagement is the actual product, and the regulated industries where the customer cannot accept lab-affiliated engineers for compliance reasons. Those are real, defensible businesses. They are not the businesses Accenture and Deloitte have been pricing themselves on.

There is a second-order question here that I do not yet have a confident view on. If the FDE vehicles work as the capital structure suggests they are designed to, then by 2029 there will be a layer of deeply embedded, lab-affiliated engineering capacity inside several thousand American mid-market companies. That layer will have commit access, deployment authority, and a commercial relationship with the lab that is, in important ways, deeper than the relationship the customer has with its own internal IT function. The question this raises, and I think it is the question that the next round of analysis on these vehicles needs to start engaging with seriously, is what happens to the customer's negotiating position when the FDE vehicle is the only entity that knows how the customer's AI estate actually works.

Palantir's history on this is instructive but not conclusive. The firm's largest customers have, over time, found themselves in a relationship where switching is technically possible but practically extremely difficult, and where pricing has drifted upward in ways that suggest the lock-in is real and material. Whether the same dynamic plays out in the mid-market, where the customer base is more fragmented, less individually valuable, and historically more willing to switch vendors, is genuinely unclear to me. The PE sponsors clearly believe it does; the recapitalisation thesis only works if the FDE practice can defend its renewal economics, and the renewal economics only work if the customer is sticky.

I am holding that question open. I think the deployment story over the next two years is the FDE vehicles building out and the incumbents losing the centre of their book; I think the question of what the resulting market structure looks like in 2030, whether it is a healthy services oligopoly, or something that looks more like the capture dynamics that Palantir's largest customers have ended up inside, is the question that the next round of writing on this needs to address.

What to watch. Three things. First, the FDE headcount numbers, both vehicles have committed, in their announcement materials, to publishing engineering headcount on a quarterly basis, and the trajectory of those numbers against the PE pace I described above will tell you very quickly whether the vehicles are scaling as designed. Second, the renewal cohort. The first wave of customers signed in 2026 will hit renewal in 2027 and 2028, and the renewal economics, both the retention rate and the price uplift, will tell you whether the lock-in thesis is real. Third, the response from Accenture, Deloitte, and the Big Four. The incumbents have, so far, responded to the FDE vehicles with the kind of public confidence that usually indicates private alarm. The first serious sign of repricing or restructuring inside one of the major consultancies will be the moment the consensus catches up to what the capital has already worked out.

The labs did not need PE capital to build these vehicles. They took it because the people closest to the cash flows of mid-market American software told them, in two parallel conversations in the same week, that this was the trade. I think the people closest to the cash flows are right. I think the consulting layer is moving, and I think the move is roughly four years ahead of where the consensus has it.

Footnotes

¶