From View From The Top, Issue #41
The Fed confirmed the weather. Apollo sized the wave. The question nobody is asking is what it costs to deploy it — and how the industry will change.
Last Wednesday the Fed raised rates by a quarter point, taking the target range to 3.75–4 per cent. 12 to 0. The first hike since 2023. The following day the Bank of England held at 3.75, but three of its nine voted to go, with CPI at 3.1 per cent in August and the Bank’s own forecast now pointing above 4 per cent by early next year.
What struck me the most was that nobody was surprised. No financial market tantrum. No repricing. The market had already absorbed what I was writing about last summer: this is the climate now. Money costs something, and it is going to keep costing something.
Which puts us back in the game I started out in years ago — before a decade of liquidity made everyone look very clever. Nobody will be riding a falling discount rate any more. Returns will come from what you buy, what you build, how you operate and by trying to find alpha in the operational arbitrage — the part this series is about — what it costs you to run the machine that does both.
The same week, at the Barclays conference in New York, one of the largest credit managers, Apollo’s Scott Kleinman put numbers on where the capital is going. More than two hundred structured transactions. He described the capital needed over the next decade as once-in-a-generation in scale. And on private credit replacing traditional fixed income: he said, “we are just pulling out of New York Harbor.” Early days.
Put those two events side-by-side. Rates that are not going back to zero. And the most sophisticated credit house in the world telling you a generational wall of capital is heading into structured, investment-grade private lending.
But every pound of it has to be originated, underwritten, documented, monitored and reported — by someone, at some cost. The winners I think of the next decade will not be decided by who can raise that capital. One of the largest credit managers in the world has just told us the capital is coming. They will be decided by who can deploy it fastest and at the lowest cost per deal without losing the judgement that keeps it safe.
A BIG CHEQUE, AND WHAT I READ INTO THIS
And what I find interesting, is a good deal of that institutional money is going into the infrastructure this new technology runs on. Their president was on Bloomberg the same week on the scale of that capital expenditure, and on rates staying higher for a while yet.
Nobody at Apollo is telling you what agentic AI does to a lending book. But you do not put that kind of money behind a technology you think is a fad. I read it as a straightforward vote of confidence in the thing itself — and if it is real enough to underwrite at that scale, it is real enough to expect how it will change the business of underwriting.
I should declare an interest here, and a failure. Three or four years ago at Hilltop we tried to build precisely this: an automated credit platform, the underwriting done by learning. I knew exactly what I wanted it to do but the language models could not do it. It was expensive and we went through a few developers who kept re-starting. We were too early, and too early is the same as wrong, only more expensive. And when I raised it with people in the industry at the time I mostly got blank looks. Even my own team indulged me politely and rolled their eyes as we tried to build it out.
Now it appears everyone is talking about it, everyone is doing it or trying, Apollo included. I don’t mention it to claim foresight. I mention it because I have already been through the exercise of working out which parts of a lending business are information work, which are extremely manual and outdated and which parts are pure judgement and experience. The proportions are not what most people assume.
Which brings me back to my story of the trading floor.
WHAT HAPPENED TO WALL STREET IS THE TEMPLATE
Years ago, I worked on those floors. Chicago first, then the Amex (the “Curb”), then the New York Stock Exchange and futures exchanges. Thousands of people. All running around in distinctive coloured jackets, clerks matching tickets. Runners carrying paper. Specialists making markets by hand. Independent clearing firms reconciling every trade overnight.
Every one of those jobs existed because information moved slowly and processing was manual. The moment technology made information instant and processing automatic, the economics collapsed. Not gradually — abruptly. Fixed dollar commissions went to cents, then to nothing. Hundreds of clearing firms became a handful. An entire ecosystem gone inside fifteen years. Poof.
But here’s the part people forget. Trading did not shrink. It exploded!
Volumes went up a hundredfold and more. The work did not disappear, it got bigger. But the cost of the work disappeared, and the volume the same capital could support went vertical.
Today real estate credit in 2026 is the trading floor in 1985. A business where information moves slowly, processing is manual, and the entire cost structure assumes both. And the technology that changes it has just arrived.
THE COST ANATOMY OF A LENDING BOOK
But think about this — ask most lenders what it costs them to do a deal and you will not get an answer. Not because they are hiding it. Because they have never had a reason to work it out. It’s not like selling widgets where you know the cost basis; the cost is embedded across an entire direct and indirect process.
Cost sits in salary lines. Underwriting is a department with a headcount and a budget, and so is operations, and so is portfolio management. Nobody allocates that cost across the activities those people actually perform, so nobody knows what any single activity costs. The P&L tells you what the business spends. It does not tell you what anything costs to produce.
The method is not complicated. Take total platform cost — every salary, every system, every outsourced line, all the overhead. Allocate it across the activities rather than the departments: origination, underwriting, execution, monitoring, reporting, fund administration. Then divide by output.
What comes out is a short list of numbers that almost nobody in this industry measures and I think most would be surprised:
Cost per deal, including the ones that died. This is the one people resist and the one that matters most. If you screen forty and close ten, the thirty that failed are part of what the ten cost you.
Cost per pound lent. Which tells you immediately why the small loan is unattractive — and the reason is not credit.
Monitoring cost per loan-year. The cost that arrives after the fee has been taken, every month, until redemption.
Hours from pack to committee. How much human time a single decision consumes from end to end.
Loans per head. The crudest of the five and the most revealing.
I should be straight about where I am with this. I do not have my own numbers yet. We are running the exercise now — the method above, applied properly to our own book — and I am comparing notes with a handful of people who run lending platforms and are asking the same question. I will publish what comes back.
If you run one and you have done this work, I would like to hear what you found. Even a rough ratio. Nobody publishes these, which is exactly why nobody knows whether their own is any good.
WHAT BECOMES SOFTWARE
Take that activity list and mark every line against a single question. Is this work collecting, verifying, reconciling, modelling, documenting or reporting?
If it is any of those, it is information work. It is moving data from one place to another and checking that it arrived intact.
Do that honestly across a lending platform and the proportion lands somewhere between sixty and seventy per cent. Not the decisions — the work that surrounds the decisions. Assembling the pack. Chasing the valuation that has not arrived. Rekeying the appraisal into the model. Reconciling the drawdown schedule. Producing the monthly report that three people read.
That is the part that becomes software. Not eventually. Now.
WHAT STAYS HUMAN — THE IMPORTANT BIT
Probably less than you would like to hear.
But the relationships that produce the deal in the first place can only be done by building those relationships, pressing the flesh and building trust. A broker calls you rather than somebody else because of fifteen years, not because of your interface with a platform. The most expensive part, and revenue generating part, stays analogue.
The judgement on the sponsor. Whether this particular developer, on this particular site, finishes the job when the ground comes back worse than the survey said. No model has seen enough of those to be useful, and the ones that matter are never in the data.
And the call that comes on a loan which has gone past its term. You wrote it for twelve months. It is now month fourteen, the scheme is ninety per cent finished, and the borrower needs more time and more money to get the roof on. Somebody has to decide — based on information — whether that is a delay worth funding or a failure worth enforcing, and the two look identical on the file. That is not a data problem. That is a person who has been wrong before.
More on this next week.
WHERE THE SAVING GOES
Here is where people get it wrong. They assume a falling cost structure means falling returns, and that lenders will simply pocket the difference in the meantime.
Neither, for long.
The risk still sets required returns. The Fed has just reminded everybody of that — money costs something again, and no amount of automation changes what a development loan in a difficult market has to yield. The saving does not go into margin. I think it goes into two other places.
It goes into spread. In a market where everyone is chasing the same prime assets, the platform with the lower cost base prices tighter and still makes its number. That is how competitive markets work and there is no reason this one is exempt.
And it goes into the minimum ticket, which is the more interesting one. Today the £3m loan does not get done by big platforms. It simply takes too many resources. Not because the credit is bad, but because it costs nearly as much to underwrite as the £30m loan. Take the cost of underwriting down far enough and the small loan becomes worth doing. That’s a game changer for UK housing and could make a huge impact.
Which is exactly where the volume is. The developer building fifteen units is not underserved because lenders dislike him. He is underserved because he is uneconomic to serve.
And remember what happened on the trading floors. The cost fell and the volume exploded. Kleinman’s generational wall of capital has to go somewhere, and it will go to whoever can deploy it at the lowest cost per deal.
HARD LANDING
I watched an entire industry’s cost structure disappear from the inside of it.
The people who survived were not the ones with the biggest floors, or the most clerks, or the deepest relationships with the specialists. They were the ones who worked out earliest that the floor itself was the cost.
Every lending platform’s organisation chart is a fossil. It is a record of what the work used to require — how many people it took to move information around back when information moved slowly. Most of those boxes exist because of a constraint that is now being removed.
The next decade’s winners get to draw theirs from scratch.
