OpenAI and Anthropic stood up deployment companies a week apart in May, both on outside capital, some of it from the same firms already financing the chips.
An enterprise AI services venture backed by Anthropic, Blackstone, and Hellman & Friedman was unveiled on May 4, with each backer committing roughly $300 million and Goldman Sachs adding about $150 million as a founding investor, toward a consortium reported at $1.5 billion. Seven days later, OpenAI announced the Deployment Company, capitalized with more than $4 billion from 19 outside firms led by TPG. In a single week in May, the two leading US AI labs stood up dedicated companies to embed engineers inside enterprises and rebuild operations around their models. Neither built the thing on its own P&L.
Some of that outside capital is the same capital. Blackstone and Apollo, two of the firms behind a roughly $35 billion private-credit vehicle that finances Google TPUs leased to Anthropic, both appear in the investor consortium behind Anthropic's services firm, now branded Ode with Anthropic. So does Goldman Sachs. The same private-capital complex that financed Anthropic's compute off its balance sheet is now financing its labor layer off its P&L. That is the pattern this article traces: the labs keep the models, and structured outside capital carries the capital-heavy layers around them. It is an interpretation, not a claim any participant has made. The deal documents are what invite it.
The Anthropic venture surfaced first. Bloomberg-sourced reporting described the three-way split: roughly $300 million each from Anthropic, Blackstone, and Hellman & Friedman, plus about $150 million from Goldman Sachs, toward a target of $1.5 billion. Do the arithmetic and Anthropic's stake lands near 20 percent. It is a founding partner in a company it does not control.
OpenAI's version arrived on May 11 with the opposite control posture and the same capital logic. The Deployment Company is majority owned and controlled by OpenAI, in the company's own words, and will "operate as an extension of OpenAI." The money underneath it is someone else's: more than $4 billion of initial capital from the 19 firms, led by TPG, with Advent, Bain Capital, and Brookfield as co-lead founding partners, Goldman Sachs, SoftBank Corp., BBVA, and Warburg Pincus among the founding partners, and Bain & Company, Capgemini, and McKinsey signed on as integrator partners. The venture folded in the applied-AI firm Tomoro, bringing roughly 150 forward-deployed engineers in on day one. Reporting on the structure put the new company's valuation at $14 billion and described the private-equity backers as guaranteed a 17.5 percent annualized return over five years.
The Anthropic-side venture spent the summer assembling itself. On May 21 it acquired Fractional AI, a startup founded in 2024 by Chris Taylor, Eddie Siegel, and Travis May. On July 15 it launched publicly as Ode with Anthropic, led by Fractional's Chris Taylor as chief executive and Eddie Siegel as chief technology officer, with a consortium the launch release says has widened to include General Atlantic, Leonard Green & Partners, Apollo Global Management, GIC, and Sequoia Capital alongside Anthropic, Blackstone, Hellman & Friedman, and Goldman Sachs. TechCrunch reported a staff of roughly 100 engineers, more than half of them former founders. The launch release aims the firm at midsize enterprises.
One number needs care. TechCrunch describes the $1.5 billion as Ode's valuation; Bloomberg-sourced reporting describes it as committed capital. The component commitments are better attested than either label, so this article presents the split and leaves the label open.
The control difference between the two ventures is real, and it is a difference of degree rather than kind. OpenAI holds the majority; Anthropic holds a minority stake near 20 percent. Anthropic also keeps an internal team for strategic deployments while Ode handles broader commercial implementation, per TechCrunch, and Ode's mandate is Claude-first: it implements Anthropic technology wherever possible and can reach for competitors where necessary. Neither lab externalized deployment outright. Both externalized the capital.
The Ode consortium did not invent this structure for labor. It imported it from compute. In June, Apollo and Blackstone closed the private-credit vehicle of roughly $35 billion that finances Google TPU hardware leased to Anthropic, a structure that keeps the hardware off Anthropic's balance sheet ahead of an IPO, with Broadcom providing credit support on $30 billion of the tranches and Apollo's Atlas SP unit holding roughly $800 million of equity, per Axios's reporting on the deal. SCN has covered the scale of the commitment being financed: Anthropic's TPU arrangement with Google runs to 3.5 gigawatts of capacity through 2031.
Put the two structures side by side and the week in May reads less like a services story and more like a capital-formation story. The chips sit in an SPV with structured credit. The deployment engineers sit in ventures with outside investors, in OpenAI's case investors guaranteed a 17.5 percent annualized return, which is capital priced like debt. The honest counterread is that a guaranteed return is also what private equity demands when it doubts the downside, and a services company with no operating record gives it reasons. Either way, the model stays on the lab's books while the capital-intensive layers around it sit on someone else's.
The overlap has a second property worth stating plainly, because it follows from the documents already cited rather than from any new disclosure. Blackstone and Apollo hold credit claims on the SPV, whose debt is serviced by Anthropic's lease payments, which depend on Claude revenue. Both firms are also investors in Ode, a company whose mandate is to generate sustained enterprise demand for Claude. The same firms hold claims on the supply of one lab's compute and on a vehicle built to create the demand that makes the compute pay.
Two readings fit that structure, and this article holds them evenly. One is conviction: capital that believes the enterprise-demand thesis is capturing both layers of it. The other is reflexivity: demand partly funded by the financiers of supply is no longer a fully independent signal of what enterprises want. The concern has recent precedent. When Nvidia committed up to $100 billion to OpenAI in September 2025, analysts asked whether circular financing was inflating an AI bubble. Neither reading implies the firms are propping up their own credit book, and nothing in the public record suggests it. The structure is what it is; both readings of it are this article's interpretation, not anyone's stated intent.
Three things mute the circularity in practice. The claims sit in different vehicles and different funds, with separate limited partners and separate mandates, so "same firm" does not mean one pool of money feeding itself. The scale is mismatched: equity commitments to Ode in the hundreds of millions against credit exposure in the tens of billions, two orders of magnitude apart, which makes Ode a rounding error as a demand pump and coherent as an option on the demand thesis. And Broadcom's credit support on $30 billion of the tranches already loosens the credit side's dependence on Anthropic's demand. The loop exists on paper; the counterweights shrink it without closing it.
Why a lab would want it this way is visible in the margins. Sacra, a private-markets research firm, estimates Anthropic's revenue run rate at $47 billion annualized as of May 2026; that figure is Sacra's estimate, and Anthropic has not disclosed it. Investor materials reported last November project Anthropic's gross margin climbing from roughly 50 percent toward 77 percent by 2028. The Information has since reported that Anthropic lowered its gross-margin projections after inference costs ran 23 percent higher than anticipated. OpenAI's gross margin ran 33 percent in 2025, per Sacra's tracking, constrained by inference costs that reached $8.4 billion that year, and The Information has reported that both labs missed their own gross-margin forecasts.
The pressure comes from below as well. SCN has tracked the mounting bets against Nvidia's inference margin, the clearest sign that inference economics are being contested at every layer of the stack.
Now set the comparators. Accenture's gross margin runs around 32 percent, per its fiscal 2025 results. Bessemer's State of AI 2025 put LLM-native gross margins near 65 percent against the 80 to 90 percent ceiling of classic cloud software, and ICONIQ's data has AI product margins closer to 52 percent. A services company is a consulting-margin business. Bolting one onto a P&L that investors are asked to price as software works against the multiple, and both labs have valuation narratives in play, Anthropic's with an IPO in view. Holding the deployment business in an outside-funded structure lets a lab collect what it wants from the arrangement, token demand and enterprise lock-in, without carrying the headcount. The public materials do not prove the accounting treatment or the margin effect. The structures are consistent with that reading, and that is as far as the evidence goes.
There is a version of embedded services that escapes consulting margins. Palantir's forward-deployed-engineer model produced gross margins around 80 percent in 2023, per its 10-K. The practitioner reading of how, as The Pragmatic Engineer describes it, is that Palantir's field engineers function as product discovery: they sit upstream of the roadmap and turn what they learn inside customers into product, rather than billing hours downstream of a signed contract.
That reading offers one lens on the control difference. If deployment engineering is product discovery, the learning inside those engagements is worth owning, which would explain paying for majority control as OpenAI did. A minority stake with a Claude-first mandate, Anthropic's position, buys the demand pull-through and leaves more of the learning loop, and all of the headcount, outside. Whether either venture converts a thousand midsize deployments into signal that reshapes the model, or installs software and bills for the labor, is the variable that decides what these companies are worth, and there is no operating record yet to judge it by.
The labs' moves look sharper against the third posture available. At Cloud Next on April 22, Google Cloud committed $750 million to accelerate its partners' agentic AI development, funding Accenture, Capgemini, Cognizant, Deloitte, HCLTech, PwC, and TCS, and embedding Google's own forward-deployed engineers alongside them. Google is paying the incumbent channel to do the deployment work. The labs built companies that compete with that channel, though OpenAI hedged by signing Capgemini and McKinsey into its structure as integrator partners. Three postures, then: own the deployment company, hold it beside you at a minority stake, or fund the integrators who already have the relationships. The channel-conflict question belongs to the first two.
For readers who plan compute rather than trade equities, the structures change the shape of the demand they are modeling against, which is where this stops being a finance story and becomes an infrastructure one.
Start with the counterparties. The entities signing compute commitments are increasingly not the labs themselves but SPVs and consortium companies with structured returns, which changes whose balance sheet the buildout risk sits on and who absorbs the damage if utilization disappoints. A capacity forecast that treats "Anthropic demand" or "OpenAI demand" as a single credit is already out of date. And where the deployment vehicle and the compute financing share investors, the demand those deployments generate is not a fully independent signal; planners reading embedded-deployment growth as market evidence should know who funded the deployments.
Then the load shape. Model sales produce lumpy, training-heavy demand. Thousands of embedded midsize deployments would produce a different kind: sustained, inference-heavy pull that runs as long as the customer's operations run on it. The premise of putting engineers on site is to close the persistent gap between enterprise AI ambition and working deployment, the same gap SCN examined in IBM's dual-architecture enterprise push, where the go-to-market ran ahead of the benchmarks. If the embedded model works at scale, utilization profiles could flatten and lengthen, and forecasts built on training cycles could start to mismodel the load. The demand would not disappear so much as change texture.
It also sharpens a question this publication tracks under its Access and Science mandate: whether the AI-factory buildout is producing supercomputing capacity that science and research can reach, or whether commercial AI is capturing the resource. If the buildout's revenue case comes to rest on enterprise-services demand carried by structured vehicles, then what commercial AI captures is not just machine capacity but the value stack built on top of it, and the risk of the whole arrangement migrates to whoever funds the vehicles. The cap tables say that group is already international: the Singaporean sovereign wealth fund GIC sits in Ode's consortium, and SoftBank Corp. and the Spanish bank BBVA are founding partners in OpenAI's structure. The deployment layer is being globalized at the investor level before it is at the customer level. That is reported fact; what it means for where the capacity lands is the open question.
Two labs reached the same conclusion in the same week: the implementation layer is worth a dedicated company, and worth building with someone else's money. Part of the capital that agreed came from the same firms that had already agreed to finance the chips. Whether the labs are quarantining a margin-dilutive business, buying product discovery, or both at once is not settled by the deal documents. What the documents do establish is that the AI stack is being assembled around the labs by private capital with structured claims on the returns, and for anyone planning infrastructure against that stack, the signature on the compute commitment is becoming as informative as the commitment itself, and so is whether the same names appear on both sides of it. Early signal, not verdict. The next few quarters of Ode and Deployment Company engagements will start to settle it.
The SCN Staff is a small AI editorial squad working under human direction. Each agent owns one job.
Scout does the research. It runs down primary sources and checks what's already been published, on SCN and everywhere else, before a story gets written. If a claim can't be traced back to a real document, Scout flags it.
Forge writes. It takes what Scout found and turns it into a draft, argument and sentences and all. Every SCN piece starts here, then gets sharpened.
Cipher handles search: the titles, descriptions, and keyphrase work that decides whether a good article ever gets found. Least glamorous job on the squad. Also one that matters more than it looks.
Pixel makes the visuals. Images, charts, the occasional diagram, all built to SCN's brand instead of pulled from a stock library. When something's easier to see than to read, it goes to Pixel.
Editorial judgment and the final call stay with the humans. So does the fact-checking.
