Two labs, one playbook, $5.5 billion: OpenAI and Anthropic both decide the moat is human

Within seven days of each other, the two frontier labs that have spent four years insisting the model is the product launched services firms whose explicit pitch is that the model is not, in fact, the product. OpenAI announced its Deployment Company on 11 May, capitalised at $4 billion and built around the acqui-hire of Tomoro and its roughly 150 forward-deployed engineers.¹² Anthropic's enterprise-services venture launched seven days earlier, on 4 May, at $1.5 billion, backed by Blackstone, Goldman Sachs, and Hellman & Friedman.³ Both announcements name-check Palantir's Forward Deployed Engineer model as the template. Neither has yet disclosed a single reference client outcome.

That is the combined pre-outcome commitment. It is a useful number to sit with for a moment, because it is doing several things at once.

What the primary documents actually say. OpenAI's launch post is unusually direct for the genre: the new entity will "embed engineers directly inside enterprise clients to drive adoption."¹ This is not the language of an API company. It is the language of a systems integrator. The Tomoro acquisition is structured as OpenAI's first material acqui-hire whose purpose is go-to-market capacity rather than model or research capability — a structural first that is easy to miss in the noise around the dollar figure.² Anthropic's vehicle is more financially legible and less operationally specific: a new entity, three sponsors, $1.5 billion, services framing.³

The frame this is a case of is FDE market structure — the question of how AI capability actually gets deployed inside enterprise, and who owns the relationship when it does. The frame predicted, broadly, that as model capability commoditised the binding constraint would move to deployment: data plumbing, workflow custom-build, change management, the unglamorous work of making a model do something useful inside a company that runs on Workday and ServiceNow. Both labs have now confirmed the frame, on the same week, with capital. That is a strong signal. It is also worth noticing what the signal is not saying.

The investor split is the second story. OpenAI's vehicle is backed by 19 investors. Anthropic's by three. There is, as far as has been disclosed, zero overlap.³ This is not the same capital hedging across both bets, which is what one normally sees when an institutional thesis is in formation; it is two distinct capital stacks underwriting the same structural read independently. The PE consortium behind Anthropic, Blackstone, Goldman, Hellman & Friedman, is precisely the buyer-set you would assemble if you were planning to run a services business with predictable margin profiles and a future secondary exit. OpenAI's 19-investor cap table looks more like a strategic round than a services-PE deal. Same thesis, different exit paths, different operating cultures already implied by the financing.

The financial establishment has placed two separate bets on the same thesis. That is rarely accidental.

Where the Palantir comparison gets uncomfortable. Palantir spent roughly a decade refining the FDE model, and it did so in defence and intelligence — environments where switching costs are near-infinite, the procurement cycle selects for incumbents, and the client has, in many cases, no functional alternative.⁴ The 640% return figure that everyone is quoting is the return over the period during which FDE was the dominant motion, which is to say the return on a decade of compounding inside a specifically unusual customer base.⁴ Commercial enterprise looks different. Accenture, Deloitte, IBM, and KPMG have been hiring AI capability aggressively and arrive at the procurement meeting with existing master services agreements, security postures, and partner-level relationships with the CIO. Neither OpenAI nor Anthropic has disclosed how its pricing, contracting, or delivery terms will differ from the incumbents. That is not a small omission.

There is also the small matter of 150 forward-deployed engineers against a $4 billion capitalisation. Even on generous assumptions, most of that $4 billion is earmarked for scaling a delivery model that has not yet been validated at commercial enterprise scale by the entity now scaling it. Palantir built its FDE bench over years; the OpenAI Deployment Company has to build one in quarters, in a labour market where the incumbent SIs are bidding for the same engineers.

Where the frame fits, and where it strains. The FDE market-structure read fits cleanly on the strategic intent: both labs are vertically integrating into the services layer because the services layer is where the durable enterprise relationship lives. It also fits on the financing: the Anthropic consortium in particular is the obvious buyer-set for a services rollup. Where the frame strains is on the execution timeline. Palantir's moat was time and accumulated process; replicating that moat with capital and acqui-hire in a competitive labour market against entrenched incumbents is not the same trade.

The adjacent frame worth holding alongside is the SaaS apocalypse read. If per-seat pricing is under structural pressure from agents that replace seats, then services revenue, project-based, outcome-based, not seat-based, becomes the logical hedge for a lab whose own pricing trajectory is unfavourable. Viewed this way, the FDE pivot is not only about capturing enterprise value; it is about capturing enterprise value in a revenue form that the labs' own products may be in the process of compressing. That is a slightly awkward strategic position to be in, and it is one explanation for why both labs moved at once.

The performativity layer. $5.5 billion committed before a single disclosed client outcome will reshape the market regardless of whether the delivery model works as advertised. It creates a labour-market signal that draws engineers away from incumbents; it creates a client expectation that the labs will arrive with a services arm; it creates pressure on Accenture and Deloitte to respond, which they will, which will further validate the category. The spend is doing market-structure work in advance of the outcomes. Whether the outcomes follow is a separate question, and one the announcements politely decline to address.

What this is a case of. Vertical integration into services by a platform vendor that has discovered the platform alone does not retain the customer. The precedents are not flattering on speed: Salesforce took years and several acquisitions to build its services capability; IBM Global Services was decades in the making; even Palantir, the explicit template here, took ten years. The labs are attempting the same move in roughly twelve months, with the capital to force the pace but not, yet, the operational evidence that the pace can be forced.

What to watch. Three things. First, the first published enterprise client outcome from either entity, and specifically how the contract is priced — project, outcome, or something disguised as either. Second, the engineer-attrition numbers at the major SIs over the next two quarters; the labour market will tell us whether the FDE bench can be assembled at the speed the capitalisation implies. Third, whether the two consortia stay distinct or whether secondary trades start showing overlap, which would suggest the institutional thesis is consolidating rather than splitting. The split, for now, is the more interesting structural fact.

Footnotes

¶