Why we passed on fourteen data-centre deals in eighteen months
Ghostwriting sample. Written in the voice of a fictional fund principal. The firm, the person and the portfolio figures are invented to demonstrate voice; nothing here describes a real fund or a real transaction.
We have looked at sixteen AI data-centre opportunities since early last year. We have committed to two. Our investment committee asks me the same question every quarter, in a slightly different tone each time: are we being disciplined, or are we being slow?
Both. Probably in that order, but I am not neutral on the subject.
Here is what we actually did, and the five reasons the other fourteen are not in the portfolio.
We stopped underwriting a contract and started underwriting a counterparty
The first deal we passed on had a fifteen-year take-or-pay agreement with a customer whose balance sheet was smaller than the contract. Everyone in the room knew it. Nobody said it, because the contract said fifteen years and the model said 1.4 times coverage, and those two facts are enormously comforting when you are trying to get to a closing.
A commitment larger than the counterparty can honour is not revenue. It is an option the customer holds and we have paid for. So we ask for the credit support, the cure periods, the termination rights, and whether the obligation survives the customer's own insolvency. If the answers are thin, the term sheet is theatre and we decline early, which saves everyone four months.
We underwrite the power date, not the construction date
I have not seen a data-centre deal late because the building was late. I have seen several late because the utility was late.
This is the part of the asset class that everyone claims to understand and almost nobody prices. Construction is a contract with a party you have leverage over. Energisation is a queue you are standing in behind strangers. When the substation slips two quarters, the acceptance test slips, the customer's start date slips, and debt service does not. We now assume, in every base case, that the power arrives later than the developer's schedule, and we ask what that costs. If the answer requires the borrower to fund a gap it has no liquidity for, we are out, and we are out early.
We will not buy a test the seller wrote
An acceptance test is the only place in these documents where an engineering claim becomes money. Throughput, sustained performance, error rates, burn-in duration, who witnesses it, and what happens when it fails.
Twice now I have read acceptance schedules that could not fail. Once because the vendor's own engineering team wrote it. Once because it tested a configuration nobody would run in production. In both cases the borrower's answer was that it would never be enforced that way, which is exactly the thing a lender should never accept, because enforcement is the whole point.
If the milestone cannot convert a failed test into a payment holiday, it is decoration. We pass.
We refuse to underwrite a price nobody can observe
This is the one that costs us the most deals, and I think it is the one that will look obvious in five years.
The contracted capacity is a business. The uncontracted capacity is a bet on inference pricing, and inference pricing has about as much reliable history as a two-year-old equity index. Every model we have seen assumes the price per GPU-hour declines gently. Almost none of them test what happens if it declines faster than the debt amortises, or if utilisation plateaus while the power bill stays fixed.
I am not smart enough to know where inference pricing lands. I am confident enough to know that we should not be paid equity returns for taking that risk on behalf of limited partners who hired us to underwrite infrastructure. So we size the deal to the contracted revenue and treat the rest as upside we do not need. That makes our bids lower than the competition's, and we lose deals because of it. I have made my peace with that.
We price being second
Everybody's model ends in year seven with the hardware fully amortised and the building still standing. Almost nobody models what happens if the tenant leaves in year five because someone two miles away is cheaper.
Residual value on this hardware is a relationship, not a market. So we underwrite the second owner as part of the base case: who they are, why they would take this cluster rather than a new one, and what they would pay. If the honest answer is "the same vendor's customers, at a discount we have not quantified", we are underwriting a hope. We would rather say no than say maybe and hope the covenant catches us.
The two we said yes to
Both had the same three properties, and none of them were about the technology.
The power was already energised, or the utility had a contractual obligation with penalties attached. The customer was larger than the contract and had already paid for something similar. And the sponsor had real cash in the deal, not a contributed asset valued by the sponsor's own adviser.
That is it. It is not a sophisticated thesis. It is the difference between a business that survives a bad year and one that needs everything to go right in the same twelve months.
What discipline actually costs
It costs us the deals everyone else brags about at conferences, and it means our deployment pace looks pedestrian next to funds that are twice our size and half our vintage. Our limited partners do not love reading that.
But the question I care about is not whether we were in the best deal of 2025. It is whether we can name, in one sentence, what would have to go wrong in any deal we own for it to be impaired, and point to the clause that protects us when it does. On the two we closed, I can. On the fourteen we passed, I could not, and that is why they are not in the portfolio.
Chris Romei writes ghostwritten thought leadership for infrastructure and technology principals. Sample work only; the principal here is fictional.