Many of my ideas arrive on a walk, listening to podcasts through the headphones. Lately, I noticed several podcasts all circling the same tension: the unexpectedly high cost of running AI at scale. It’s the flip side of the revenue growth everyone’s celebrating at Anthropic and OpenAI. Every dollar of that revenue is someone else’s cost line, and that someone (more often a CFO) is starting to notice.

The pushback has started. Executives are now trying to rein in the initial excitement for AI and the spend. Look harder at whether you need the frontier model at all, when there’s a cheaper or open-source one that’s genuinely capable, not at the frontier but close, for a fraction of the price. Do nearly the same thing for a tenth of the cost, and pocket the difference.

I understand the instinct completely because I spent years of my life making exactly that mistake in a different industry. And I think it’s wrong. Fundamentally wrong. In a way the whole expensive-versus-cheap argument has inherited without noticing.

The cheap model doesn’t do nearly the same thing. It does 95% of the same thing. And the last 5% is where all the value is.

What follows is a mental model, not a forecast. My walk is where the thinking happens, where my investment ideas come from, and this is the lens that came out of it: how AI reshapes an industry from the inside, and, because I’m an angel investor, where the opportunities fall out. The reshaping I’m fairly sure about. The opportunities are the reason I’m watching.

The letter after “cost-per-“

In 2000 I was running a pay-per-click business at Dark Blue Sea. The online advertising industry at that point was in the middle of a decade-long argument about which letter to put after “cost-per-“, and I had a front-row seat.

It started with CPM, cost per mille, the price of showing your ad a thousand times. Simple to measure, simple to bill. It was also close to meaningless, because showing an ad is not the point of an ad. So, the industry moved to CPC, cost per click. Better. A click is at least a flicker of intent. But a click isn’t the point either. You don’t want clicks, you want customers, so we got CPA, cost per acquisition. Now you’re paying for an outcome.

It didn’t stop there. Not all acquisitions are equal: a customer who buys once and vanishes is worth a fraction of one who stays for five years, making lifetime value the number that actually matters, not the cost of a single transaction. There’s even a higher number that sits above that, the one nobody could put on an invoice. Shareholder value, or the quality of the customer base you were building and the durability of the franchise. That’s what you were really optimising, whether you measured it or not.

Every rung on that ladder measures something closer to what you care about. And every rung is a temptation to stop one step too early, because the lower rungs are easy to measure and the higher ones are not. The cost of stopping early is small and quick to show up at the bottom, and large and slow to show up at the top.

Token cost is CPM

You can apply this same approach to the AI debate. “Too expensive per token” is a CPM statement. It’s pricing the impression, not the outcome. However, the token price hides a far more layered approach to reach the end goal.

You give a task to a cheap model, and it gets you to 95%. Impressive, cheap, fast. But the job isn’t 95% done in any sense that matters, because the remaining 5% doesn’t sit there politely waiting to be ignored. It has to close, by a human or by a better model, and until it’s closed the outcome doesn’t exist. You can’t ship 95% of a financial model or 95% of a contract.

The honest comparison isn’t cheap-model tokens against expensive-model tokens. It’s the cost of the finished outcome one way against the cost of the finished outcome the other way. Once you frame it like that, the expensive model changes character. A frontier model that costs ten to a hundred times more per token is still cheap if it closes the 5% that would otherwise need a person. Its CPM is absurd. Its cost per outcome is a bargain, because the thing it’s replacing isn’t a cheaper model, it’s a salary.

You don’t use the expensive model for the bulk of the work. You use it at the margin, to replace the marginal high-value human on the part of the task the cheap model can’t finish. And it’s a moving line. Today the cheap model does the grunt work and the expensive one finishes it. Next year a better-and-cheaper model does what the expensive one did, and the frontier moves up to a harder 5%. The waterline keeps rising. The question is never “cheap or expensive.” It’s what’s the right mix, where does the line sit today, and the line moves every few months.

To see what that does to an organisation, it helps to look at one built entirely out of these layers.

Five layers, three years each

I spent part of my career running Citigroup’s equity capital markets business in Australia, so the investment bank is the structure I reach for. An IB is a hierarchy of roughly five layers: analyst, associate, vice president, director, managing director. Consulting, law, and accounting firms are built the same way, albeit with different labels. Each layer is about three years wide. You come in at the bottom, and if you’re good you climb a rung every few years.

Within each layer there’s a distinction worth keeping straight: the difference between a task, an outcome and a job. A task is a thing you do. An outcome is a thing that gets finished. A job is a person with a title, a seat, a career and a future. AI eats tasks easily, outcomes with more difficulty, and jobs not at all, because a job is not a pile of tasks, it’s a position in a structure that goes somewhere.

I’m as convinced as I am of anything that cheap models will eventually do 80 to 95% of the work at every one of these layers. The open question, the one nobody can answer yet, is how the remaining slice splits between a human and an expensive frontier model, and how fast that split moves up the hierarchy. To me, that’s less of a “small uncertainty” and more the whole design problem of the future firm.

Start at the bottom. The analyst layer is mostly tasks: building models, formatting decks, pulling comparables. A cheap model does the bulk of that to 95%, an expensive model closes the 5%, and you’ve replaced the analyst layer at a fraction of the cost. The cost per outcome collapses. On the spreadsheet it’s the cleanest decision a CFO ever made.

The analyst was never producing deals

Here’s what the spreadsheet doesn’t show. The analyst layer was never really producing models and decks. That was the task. The output, the thing it existed to manufacture, was future managing directors.

The MD was an analyst nine years ago. The three-year rungs look like a pay scale, when it’s more like a production line taking the raw material of junior bankers, shaping them as they climb into senior roles. Cut the analyst layer to zero because a model can do the modelling, and your cost per deal looks wonderful for about three years. Then the associate bench is thin, because you stopped feeding it. Three years after that the VP bench is thin. Ten years out you have no MDs of your own, because you optimised away the entry point they all came through.

That’s CPA-optimisation ignoring lifetime value. You banked the cheap outcome and quietly destroyed the thing that produces every expensive outcome you’ll need in a decade. And above even lifetime value sits shareholder value, the durable franchise, which is precisely the thing you’ve spent, to make one financial year look good. The reputation, the relationships and the succession never had a chance to grow.

Same ladder as the ad industry, and thus the same error. Maybe even worse because there can be higher stakes that take ten years to show up instead of one quarter. By then the person who made the decision has been promoted on the strength of the margin they delivered.

The analyst doesn’t vanish. The analyst’s job changes

The factory analogy shows why “cut to zero” is the wrong instinct. When weaving moved to the mill, the mill sack all of the employees. It employed different people doing a higher-order task, keeping the machines running rather than throwing the shuttle by hand, and the higher volume needed them. The new analyst is that person: rather than build the model, they can coordinate a swarm of AI agents to do it, and check what comes back. That’s a real job that still sits at the bottom of the hierarchy, and it still feeds the production line. You still get your future MDs, they just came up through agent-coordination instead of manual spreadsheet-building.

The succession problem can be easily resolved by reconceiving it. Cut it to zero and you starve the top of the house in a decade. Reconceive it with fewer analysts doing a coordination job that didn’t exist before and they will still feed the ladder. You can keep the franchise while capturing most of the cost gain. Now that the modelling is free, the firms that get this right will rethink what the role an analyst before they rush to cut.

The part you can’t automate at all

There’s a harder limit, and it sits above all of this. Always remember that a transaction is not a document. It’s a negotiation between people. The managing director who has known the target’s CEO for fifteen years, the read on whether the other side is bluffing, the timing of a call, the ego management, the trust. That’s the actual job at the top of the house, and it’s not a 5% you close with a better model. It stays more of less human regardless of how good the models get, because it was never an information-processing problem in the first place.

This is the people element, that same force I wrote about in The Gravity of the Good Exit. It’s the human overlay of exhaustion, relationships, personal numbers and board dynamics that pulls so many exits into the same band and makes deal outcomes so hard to forecast. Deals are done by people, to people, and that part doesn’t industrialise.

Which means every deal has a line drawn through it. On one side, the machine: the modelling, the comparables, the first draft of the memo. On the other, the person: the room, the relationship, the judgment call, the handshake. Knowing exactly where that line sits, in each task, at each layer, in each deal, is the real skill of running an AI-era firm. Draw it too high and you’re paying humans to do what a model does for pennies. Draw it too low and you’ve automated away the thing the client was paying for. And the line moves, because the machine side keeps expanding into the human side a little more each year.

You can’t buy the outcome you can’t measure

There’s a reason the ad industry took a decade to climb from CPM to CPA, and it wasn’t stupidity. It was measurement. You cannot pay for an outcome you cannot see. Paying per acquisition required the pixel, the attribution model, the tracking, the exchange; an entire technical apparatus that had to be built before outcome-pricing was even possible. CPA was simply an engineering problem that took years to solve.

AI is at exactly that point now. Routing work to the cheapest model that can finish the outcome sounds obvious, but it’s unbuildable until something can verify the outcome is actually finished. You can’t automatically send the 5% to the expensive model unless you can detect the 5%. The measurement layer comes first. Everything else waits on it.

The Lumascape comes for tokens

If you were in online advertising, you will remember the Lumascape, those famously dense diagrams showing the hundreds of companies that sat between an advertiser and a viewer. Demand-side platforms, supply-side platforms, exchanges, data brokers, verification firms, analytics shops. It looked like chaos yet worked like connective tissue that let the industry price on outcomes at all. Programmatic advertising didn’t work until that ecosystem existed to measure, route and verify.

The same tissue is going to grow around models and tokens, and for the same reason. Somebody has to measure whether an outcome was finished. Somebody has to route the 95% to the cheap model and the 5% to the expensive one. Somebody has to verify, orchestrate, log and price. None of that is the model, and the first box on the diagram is verification, because nothing downstream works without it. This is the AI services stack that has to exist before anyone can buy cost per outcome instead of cost per token.

For an angel, this is the centre of the whole thing, so hold onto it, because I’ll come back to it. The frontier labs are not an angel game. But the stack around them, the measurement and routing and orchestration layer that a small AI-native firm rents instead of builds, is exactly the kind of fragmented, early, picks-and-shovels opportunity set where angel returns come from. It’s the same shape I wrote about in The Multiplex Moment. In a world where companies vie for headlines, the real winners are the businesses that appear once the infrastructure settles.

From bespoke to off-the-rack

So, what’s the steady state? Think about clothes. For most of history they were made by hand, one garment at a time, expensive enough that owning many was a mark of wealth. Then the process industrialised, and now clothes are for everyone, so cheap and abundant that we don’t think about them at all. The craft didn’t survive as a mass activity, nor did the market shrink. It exploded, and the product became invisible, something everyone simply has.

That’s the endgame for investment-grade analysis, and for legal and consulting and audit work too. An investment bank has always been throughput-constrained. Its output is roughly deals per year times revenue per deal, and deals per year is capped by human hours. That cap is why banking has a floor on deal size: below some level of transaction value the fee doesn’t cover the hours, so the deal simply doesn’t get served. A large tail of smaller companies goes unbanked, not for lack of demand but because the cost to serve them is fixed and high. Bespoke, in other words, and priced like it.

AI moves the cost to serve. Lift the throughput cap and a firm can reach down-market and serve the tail, and in a competitive market it will be forced to, because this is the discipline of the Google and Facebook ad auctions: in an outcome-priced market you run best practice everywhere or you get undercut and lose the volume. The steady state I’d bet on is a higher quantity of deals with a lower revenue per deal, and a lower bar to access. It will mean a whole class of founders and smaller companies who could never afford a banker are suddenly able to.

There has to be some elasticity here; lower the price of anything and more of it gets consumed. Whether the industry is net bigger or smaller afterwards is the genuinely open question (more likely bigger) and the reason is history. Everything that has gone from bespoke to off-the-rack, be it clothes, printing, brewing, steel, has ended up a larger industry, not a smaller one. I’ll flag that as a bet rather than a forecast.

Who builds the factory

Here’s where the model gets specific, and where an angel should pay attention, because the bespoke-to-off-the-rack transition doesn’t treat all incumbents the same.

The giants adapt. I would not bet against Goldman Sachs or JP Morgan as businesses. They have the balance sheet to build the factory themselves. They have the brand that survives the transition, and the client relationships, the people element, that don’t industrialise. They will re-price their layers, reconceive their analysts, and keep the top of the house intact. A new batch of AI-native firms will be born alongside them, built factory-first to serve the newly-bankable tail, carrying none of the legacy cost structure. Those two groups split the enlarged market between them.

The losers are in the middle: the large and mid-tier firms that don’t lean in. They are too big to move like a startup yet too small to out-invest the giants, still pricing bespoke work in a world that’s gone off-the-rack. They are the cottage weavers who bought a second loom while the mill went up across the valley, optimising the wrong rung one step too early, right up to the moment the market they were defending stopped existing. The squeezed middle is where disruption usually hurts most, and this will be no exception.

The opportunity is the pattern

Let’s cycle back to my earlier point: The frontier labs are not an angel game. None of this is confined to banking. The five-layer stack, analyst to managing director, is the shape of consulting, law, accounting and audit too, and of a dozen other professional-services industries built on leverage and apprenticeship. The same play runs in all of them: cheap models take the bulk, the frontier and the humans fight over the margin, the access bar drops, the giants adapt, the AI-natives are born, the middle gets hollowed out. And every one of those industries grows its own Lumascape, its own AI services stack of measurement, routing, verification and orchestration businesses that have to exist before any of the rest can work.

What I’m looking at, as an angel, isn’t one opportunity. It’s the same opportunity repeating, fractally, across every professional-services industry there is, each with its own connective tissue to be built, most of it too small and too early to be on any frontier lab’s roadmap. Every one of those stacks is a map of where the deals are. That’s exactly the ground where angels do their best work. I think there will be more of these opportunities than any of us can count. Finding them is just a matter of running the model, layer by layer, industry by industry, and asking the same question each time: where’s the 5%, who closes it, and who builds the thing that routes the work.

That’s the whole point of having a mental model. It’s not a forecast you wait to see confirmed. It’s a lens you run across one industry after another, and the opportunities fall out the other side. The number that matters was never the cost of the token. It’s the cost of the finished outcome, and, one rung up from that, whether you’re still the kind of firm that can produce those outcomes in ten years. Most of the debate is still arguing about the price of the impression.

Richard Moore is co-founder of MooCoo Ventures, an angel syndicate that co-invests alongside Brisbane Angels, one of Australia’s most active angel groups. He has made over ninety personal angel investments since 2013.

moocoo.vc

DISCLAIMER: The information provided is for general informational and educational purposes only and should not be construed as financial, investment, or legal advice. Any decisions you make based on this information are your sole responsibility. You should consult with a qualified financial advisor or other professional before making any financial decisions. Past performance does not guarantee future results.