Let me start with the number that ought to sink this essay.
In April this year two economists, Shuide Wen at Tsinghua and Beier Ku at Oxford, published a paper measuring what it actually costs to run the kind of AI-maintained knowledge system I have spent the last two essays arguing for. They built one, ran it against the two standard alternatives, and counted tokens. The result: the compounding system used roughly three and a half times more tokens than the cheapest stateless approach on the queries they ran, and in their thirty-day projection it never caught up. At the most favourable setting they modelled, it still cost about 3.8 times as much. There is no crossover on any horizon they could see.
If you have been sold an organisational AI memory on the promise that it saves you money per query, you have been sold something false. The people who built the most careful version of the experiment I know of say so plainly, and they called it their “central honest finding.”
So why am I leading with the number that should end the argument? Because Wen and Ku’s next move is the most useful thing anyone has written about corporate AI this year, and it only makes sense once you have accepted the bad news. Their claim is that the extra spend is not a cost at all. It is being booked on the wrong ledger.
Two kinds of token
Here is the distinction, and once you see it you will not be able to unsee it.
Most of what any organisation spends on a large language model is a consumable. You ask, the model generates, you read, the answer is gone. Tomorrow’s question pays the full price again and the system has learnt nothing. That is what a chatbot is; it is also what most “AI over our documents” deployments are, however sophisticated the retrieval. The spend behaves exactly like electricity or paper: consumed in the period, expensed in the period, nothing left at year-end.
But there is a second kind of spend. When a system takes what it just worked out and writes it back into a persistent, structured record (a synthesis page, an updated entity, a cross-reference between two things nobody had connected before) that spend produces something that is still there tomorrow. Wen and Ku’s experiment makes this concrete. Their third query, run after a deliberate restart, cost 28,000 tokens, of which about 18,000 went on searching for what the system did not yet know and writing the result back into its own pages. Their fourth query, on a brand-new angle of the same subject, then cost 4,000 tokens, because the page already held the answer. The stateless system would have paid its full price a fourth time and carried nothing forward.
That 18,000 tokens was not consumed. It was invested. The economists’ word for what it produced is the right one, and it is the word this whole essay turns on: a capital good.
They show it meets the four defining tests. It is a persistent product: the page exists after the answer has been read. It generates compound returns: every addition lowers the cost of the next in-domain piece of work, which is why the cost ratio in their model narrows from 6.3 times on day one to 3.8 times on day thirty. It is heritable across model generations: plain files carry across to whatever model you run next year, which is the “own the context, rent the intelligence” principle from Part 2 stated as an asset property. And, remarkably, it exhibits negative discounting. Physical capital depreciates. A well-kept knowledge base appreciates, because each refinement makes the whole more useful. There are not many assets you can say that about.
The accountants have been here before
If reclassifying a stream of spend from expense to asset sounds like a rhetorical trick, it is worth knowing it has a precedent, and that the precedent was not controversial for long.
Before 1985 every dollar a US company spent developing software was expensed as it was spent, indistinguishable on the income statement from the coffee. Then the Financial Accounting Standards Board issued Statement 86, which said something that now sounds obvious: once a piece of software has passed the point of technological feasibility, the money spent finishing it is building a durable asset and should be capitalised as one. Same engineers, same salaries, same code. A different and more truthful account of what the spend produced.
Wen and Ku’s proposal is structurally identical, and I think it is the correct frame for every organisation now deciding what to do about AI. The question stops being “what does each query cost?” and becomes the one a board actually cares about: what does the company own at year-end that it did not own in January?
I want to be careful here, because the paper is careful, and because any finance reader will already be objecting. No auditor will book this asset for you today. Under IAS 38, the standard that governs intangibles in Singapore, the UK and most of the world, internally generated know-how is expensed as incurred, maintenance is expensed, and nothing internally generated may be written up in value, however much it appreciates. SFAS 86 is a precedent in the economic logic, not a treatment you can apply next quarter. Where the asset does get recognised is at the moment it changes hands: a buyer’s purchase price allocation routinely records proprietary databases, documented processes and know-how as separately identifiable intangibles. Invisible on the seller’s books; visible in the buyer’s price. Hold that thought, because it is where the private equity section is heading.
Nor do Wen and Ku put a monetary value on the asset. They say, correctly, that pricing it needs assumptions about future query patterns, model migration and discount rates that a small experiment cannot support. Their claim is qualitative: an honest set of books has to record this asset somewhere, and standard token-cost reporting records it nowhere. I am going to hold the same line. I am not going to tell you what your organisation’s knowledge is worth. The one number you can defend is its cost of formation: the spend that went on writing back, plus the hours spent curating. That is a floor, not a value. But it is an auditable floor, which is more than the knowledge in your people’s heads can offer, and it lets you say with a straight face that the asset exists, that it is currently invisible, and that invisibility is not the same as worthlessness. It is usually the opposite.
The two readers who need no translation
The capital-goods frame applies to any organisation that does repeated, domain-bounded knowledge work. I will come back to that, because it is the real point. But it lands hardest with two audiences who already think in this vocabulary, and I want to speak to them directly before generalising, because they show what the argument looks like when the reader already has the concepts.
Private equity. A PE fund is in the business of buying an asset, improving it, and selling it for a higher multiple. Ask any operating partner where a portfolio company’s operating knowledge lives today and you will get an uncomfortable answer: in the heads of a dozen people, in a shared drive nobody can navigate, in decisions whose reasoning evaporated the week after they were made. When those people leave, and in a five-year hold some of them will, the knowledge leaves with them, and the buyer at exit is paying for a business that has partly forgotten how it works.
The market already knows this, and it already pays for it, just not as an asset. Look at how acquisitions are actually structured: earn-outs, retention bonuses, rollover equity, key-person clauses, transitional services agreements. Every one of those exists because the buyer knows that a material part of what it is paying for lives in a few heads, and it designs the deal to keep those heads in the building until enough has been transferred to survive their leaving. That is the market pricing operating knowledge. It prices it as a risk to be hedged through people: deferred consideration, tied to the goodwill of individuals who may walk the day the earn-out vests, buying a transfer that happens in conversations nobody writes down. It is an expensive, leaky and partial hedge, and everyone at the table knows it.
Now describe the alternative in the fund’s own language. An appreciating intangible asset, owned outright, held in a substrate the fund controls rather than inside a vendor’s tenancy, that captures how the business actually operates, survives management turnover, and can be read at diligence rather than interviewed. The buyer discounts less, because less of the value is contingent on who stays. Retention terms get shorter and cheaper, because they are protecting judgement rather than memory. And the asset itself appears in the buyer’s purchase price allocation, where the seller’s own books could never carry it. That is not a software subscription. That is a thesis. It changes the exit multiple of every company it runs in, and the fund can carry the discipline from one portfolio company to the next.
The buyer’s side of the table is already moving this way. Buyers no longer take claims of AI-driven value on trust; they want it quantified, governed and shown to survive a change of owner. EY’s 2026 study of exit readiness found that the share of managers who see AI as a challenge in preparing an asset for sale rose from 7 per cent to 17 per cent in a single year, and that evidencing value creation in exit EBITDA is now both the top challenge sellers report and the biggest single determinant of how the sale goes. Read those two findings together and the gap they describe is the one this essay is about: the value was created, and the evidence was not. Unsupported value is not discounted at signing; it is struck out. Which means the record of how a business is run is not a nice-to-have for the data room. It is the evidence room of the exit, and it is either built continuously through the hold, as an asset, or assembled in a hurry, from memory, in the three months before a sale. Credibility cannot be retrofitted. I do not know a PE professional who needs the argument explained twice.
Managers of physical assets. This is the world Janus Digital works in, and it is the audience for whom this essay writes itself. People who run buildings, infrastructure and portfolios of physical stock steward capital assets for a living. They know viscerally that an asset compounds in value if maintained and depreciates if neglected; that book value and operating cost are different columns; that there is a difference between renting and owning.
Say to that reader: your operating knowledge is a capital asset too. Every commissioning decision, every fault diagnosed and fixed, every tenant negotiation, every “we tried that in 2023 and here is why it failed.” Right now that asset is undocumented, uninsured, and depreciating at the rate your people move on. The system I am describing is the digital sibling of the assets you already manage, and it wants the same thing they do: to be owned, maintained, and allowed to compound. Nobody in this audience has ever asked me why an appreciating asset is worth investing in. They have only asked why it took this long for anyone to frame it that way.
Now generalise
Those two readers are not the boundary of the argument. They are the ones who already have the vocabulary. The argument itself is universal, and here is why.
Wen and Ku identify the condition under which the asset compounds fastest: high topic concentration. The system needs to be asked about the same domain, repeatedly, so that each answer has a good chance of being useful to the next question. Under low concentration (a consumer assistant asked about everything and nothing) coverage grows slowly and little accretes.
Look at that condition and then look at any real business. A law firm asks about its clients and its matters. A hospital asks about its equipment, its wards and its protocols. A manufacturer asks about its lines, its suppliers and its defects. A marketing team asks about its markets, its messages and what worked last quarter. Deep, repeated, domain-bounded operating work is not a special case. It is the definition of an operating business. The paper’s “favourable condition” is simply what most companies look like from the inside.
Which means every organisation is already running this experiment, whether or not it knows it. It is either accumulating an owned, appreciating record of how it operates, or it is leaking that knowledge through resignations, retirements and forgotten Slack threads. There is no third option where the knowledge is safely “somewhere.” The only question a CEO or CFO needs to answer is which of the two their organisation is doing, and whether they are content with the answer.
You cannot buy it. You can only grow it.
There is one property of this asset that changes the strategy around it, and it is worth understanding why it holds before reaching for the metaphor.
Two people can live through the same year and come out of it with different memories, because memory is not a recording; it is a selection. What each of us chooses to notice, keep and connect is a large part of what makes us who we are. Organisations are the same. Two firms can receive identical inputs, the same market, the same regulations, the same vendors, and end up with entirely different knowledge assets, because each one selected differently: which incidents it treated as lessons, which decisions it bothered to explain, which patterns it decided were worth naming. A colleague of mine who has spent his career around buildings puts it this way: no two buildings are alike, so you start with general knowledge and it gets selective fast. Put a model on top of that selection and you have a force multiplier. Put a model on top of general knowledge alone and you have a search engine.
This is why the asset cannot be copied even in principle. It is not merely that it took time to accumulate. It is that it is the record of what this particular organisation chose to remember, and nobody else can make those choices for it. Steve Yegge put the consequence better than I can. Writing in August about the fifty-agent system he has been running, he observed that intelligence grows around your domain “like ivy,” wrapping itself around your databases, your workflows, your org chart, your conventions. “I don’t think it’s transplantable, either. You can’t rip ivy off someone’s wall and stick it on someone else’s. You have to seed it, then grow it. There’s no shortcut.”
The ivy grows to the shape of the wall. For an ordinary IT purchase, non-transplantable would be a defect. For a capital asset it is the whole point. The reason a competitor cannot copy your knowledge asset is the same reason it is worth having: it is made of your decisions, your incidents, your context, accumulated over your time. A rival who wants one has to start now and lose the years you have already banked. This is the closest thing to a real moat I have seen in enterprise AI, and it is available to any company willing to plant the ivy.
It also disposes of the objection I hear most: “the model vendors are shipping memory features, so this will be commoditised.” A vendor’s memory feature is a consumable dressed as an asset; it lives in their tenancy, on their terms, and as I wrote in Part 2, we watched one of those tenancies go dark for eighteen days in June. The asset I am describing is yours precisely because it is not theirs. The memory-feature comparison is not a threat to the frame. It is the frame’s best illustration.
What we did about it
At Janus Digital we have been operating this way since the spring, long enough that the economics paper read to me less like a revelation and more like someone finally doing the sums. The Prime Radiant is our company’s memory held as structured, version-controlled text: decisions with their reasons, evaluations with their evidence, briefs that build on the last nine briefs, maintained by the AI and owned by us, governed by a rulebook we wrote and keep amending. It began in the AI Office and is now rolling out one function at a time.
What the capital frame changed for us is the question we ask about it. We stopped asking whether it was cheaper than the alternatives. It is not, and we now know roughly by how much. We started asking what the company owned at the end of each month that it had not owned at the start: how many pages, how many cross-references, how many decisions whose reasoning we could still reconstruct. Those numbers only ever move in one direction, and they are the ones I would put in front of a board.
The test
In Part 2 I borrowed a test from Satya Nadella: if you switched off your model provider tomorrow, would your company’s accumulated expertise stay, or walk out with the vendor?
Here is the Part 3 version, and it is a question for the balance sheet rather than the server room.
Ask what your organisation spent on AI last year. Then ask what it owns as a result. If the honest answer is “answers we have already used,” you bought consumables, and you will buy them again next year at the same price. If the answer is “a record of how we operate that is larger, denser and more useful than it was twelve months ago,” you have been forming capital, whether or not anyone booked it that way.
Wen and Ku’s line for the difference is the one I keep returning to. Two of the three systems they tested “leave the user with answers but no system. The third leaves the user with answers and a system that becomes more capable each day.”
Most organisations are still paying for the first kind and wondering why nothing accumulates. The ones who understand that they are building an asset, rather than buying a service, will have something on the balance sheet in five years that their competitors cannot buy at any price. Not because their AI was smarter. Because they started planting earlier.
Unabashedly written with the help of Claude. Ideas are my own.