The honest number

OpenAI is reportedly guaranteeing private equity partners at least 17.5% a year to back its new enterprise AI venture. The deal, structured as a $10bn joint venture with TPG, Bain, Brookfield, and Advent, is expected to close in early May. Anthropic is in parallel talks with Blackstone, Hellman & Friedman, and Permira on a smaller but structurally identical vehicle. The shape, in both cases, is the same: lab capital plus PE capital, engineers embedded inside portfolio companies, workflows restructured at the seams. A Palantir-style forward-deployed engineering arm, financed by buyout money.

The headlines have focused on the size of the cheque. The interesting figure is the 17.5%.

That number is the most honest line in the AI market right now. It is honest because it tells you, in the language allocators actually use, where the lab thinks the operating margin sits.

What a guaranteed return is really saying

A guaranteed minimum return is not a generous offer. It is a price ceiling on the participation of the counterparty. If the lab can credibly cap PE at 17.5% and still expect to win on the trade, then the operating return on the work that sits above that cap — the embedded engineering, the workflow compilation, the agentic infrastructure — is several multiples higher. The lab is keeping the part that compounds. The PE firm gets carry on the leftovers, plus a brand association and a distribution channel into its own portfolio.

This is not an unusual structure in capital markets. It is what happens when one party owns the productive asset and the other party brings the customer relationships. The party with the asset takes the upside. The party with the customers takes a coupon.

The unusual thing is that the customers, in this case, are PE's own portfolio companies. Funds are being asked to be the distribution layer for someone else's operating system, on terms that look more like a bond than an equity stake.

The thirty-year pitch is being rewritten

For three decades the PE pitch had three legs: capital, governance, and operating expertise. The first two are intact. The third is the one that is changing under everyone's feet.

The operating partner bench was the moat. It was the reason a fund could pay a high multiple for an asset and still earn its returns: send in the right operator, find the twenty highest-impact decisions, change them, and capture the spread. That model worked when the operator's edge was experience, network, and judgement honed over a thirty-year career.

The new structure asks a different question. If the operating expertise is going to live in the forward-deployed engineer who flies in from the lab, with frontier-model fluency, a clean Claude Code session, and the entire platform's deployment history in their head — what is the operating partner's job?

There is an answer. But it requires the fund to do something it has not historically had to do: build its own operator capability at the same hiring bar as the lab. Not subcontract it.

A worked example

Consider a mid-market industrial software platform that a fund acquires for $400m. The thesis is the usual one — there are eighteen SOPs in the back office that, once compiled into agent workflows, would lift gross margin by nine to twelve points and free up two years of revenue capacity per analyst.

Path A is the JV path. The fund signs the DeployCo agreement. Forward-deployed engineers from the lab arrive on day thirty. By month nine they have the workflows running. The lab has gross margin visibility into the platform, knows the quality of the data, knows the seams in the product, and is now in a structural position to recommend the same playbook to the next four buyers in the category. The fund has shipped the deal and locked in 17.5%. It has also handed over the most strategically valuable map in its portfolio.

Path B is the in-house path. The fund spends the first six months hiring three operators of its own at frontier grade — people who can read a codebase, write production prompts, and make the call on what to ship and what to ignore. By month twelve they have the same workflows running. The fund owns the compiler, the prompt library, the eval harness, and the strategic map. The lab is a vendor. If a better lab appears in 2027, the fund can swap.

Path A is faster and looks cheaper. Path B compounds.

The steel-manned case for the JV

There is a real argument for the JV that I want to take seriously, because if you do not steel-man it you will not see why intelligent funds are signing.

The case is this. The hiring bar that Palantir's FDE programme actually requires is not one most PE firms can clear. The talent is scarce, the compensation expectations are non-standard, the LP base does not understand why an operating bench needs to be paid like a software company. By outsourcing the engineering layer to the lab, the fund gets access to a hiring outcome it cannot reproduce internally. The 17.5% is the price of admission to that talent pool. It is, on this view, a perfectly reasonable trade.

That argument is not wrong. It is incomplete. It assumes the talent pool will remain external indefinitely. It will not. The compounding return on building the bench in-house is structural, and the funds that figure out how to do it — the funds that change their compensation philosophy and their hiring philosophy at the same time — will end the decade with an unfair advantage that does not require them to share margin with the labs at all.

There is a window. It is open. It will close inside eighteen months, because by then the talent will be more expensive, the JVs will be locked in, and the funds that did not move will discover they have rented the operating layer of their own portfolio companies on a five-year contract.

The second-order read

The most interesting consequence of this trade is not the dollars. It is the data.

When forward-deployed engineers from a frontier lab work inside a portfolio company for nine months, they learn things about that business that the buyer often does not learn in three years. They see the seams in the product, the limits of the data, the places where the SOPs break under load. That information has commercial value. In the JV structure, that value flows upstream to the lab.

Across a fund of forty companies, that is forty data points about how operations actually work in a sector. Across the four major PE-lab JVs being negotiated this quarter, that is several hundred. The labs are not buying revenue. They are buying a map of where the operational margin sits in the real economy. The 17.5% is what they are paying for the right to draw that map.

This is the part that should change how funds think.

What this means for the fund-level AI role

I have spent the last three years building SOP++ Factory at A.P. Moller – Maersk — a compiler that turns standard operating procedures into production GenAI workflows, now running across six countries, intercepting just under 52% of around a million emails per month, with documented savings in the tens of millions. The lesson from that work is unambiguous. The compiler is the asset. The prompts are the IP. The judgement about what to encode and what to leave as human is the operator's edge. Every one of those layers should sit on the buyer's side of the wall.

If you are running a PE fund and you are reading the DeployCo terms this week, the question is not "should we sign". The question is "what does our internal operator bench need to look like in twelve months, and are we hiring that bench now". The funds that ask the second question will not need the JV. The ones that do not will be paying 17.5% for a long time.

The honest number is 17.5%. What happens to the rest decides which lab gets your pension money.