Priced for perfection: A retired fund manager's warning on the AI boom

Every so often a piece of private research does the rounds that's too good to keep private. James Gubb's note on AI valuations, published in full below, is one of those.Gubb is a retired Chartered Financial Analyst, which means two things worth knowing upfront: he's spent a career being paid to separate genuine value from a good story, and he no longer has a book to talk. So, what follows isn't a stock tip. It's a skeptic's audit of the biggest capital deployment in corporate history. And it lands at an uncomfortable moment, with the CAPE ratio sitting near 41 - the second-highest reading in 140 years - and equity markets still grinding out fresh records on the back of AI euphoria.His central image is a cake. Picture the AI industry stacked in four layers: chips and hardware at the base (Nvidia, TSMC, Broadcom — worth $8.5 trillion, with genuine pricing power), then data centres and neoclouds, then the frontier labs building the large language models, then applications on top, worth roughly $13 trillion. The catch, Gubb argues, is that the two tiers actually spending the money - data centres and the labs - are worth a fraction of what's flowing through them. It's an unstable structure, and the narrow middle is where it's most likely to buckle.From there the note builds methodically. Capability, he shows, has become a commodity: the cost of matching last year's frontier model has collapsed roughly two hundredfold in three years, driven by open-weight competitors out of China, the US, Europe and Israel. Every new flagship model is, in Gubb’s words, an orange discarded onto an ever-growing free pile. Except unlike fruit, software doesn't spoil. To borrow from Harvard Prof Clayton Christensen's Innovator's Dilemma: the incumbents keep climbing the high-margin staircase even as cheaper alternatives quietly swallow the market beneath them.Then there's trust. Gubb walks through a string of 2026 episodes: Microsoft's restructuring of its OpenAI stake; Anthropic and OpenAI both launching consulting arms that compete with their own enterprise customers; and Figma's public falling-out with Anthropic over its Claude Design launch. Follows on to argue that big, security-conscious clients have reason to want their data off vendor servers entirely, accelerating the shift to models run in-house.The sharpest section, though, reframes the whole boom as a credit story rather than a technology one. Roughly $2.1 trillion in contracted compute backlog now sits across the major hyperscalers, much of it effectively owed by two loss-making labs. Treat those take-or-pay contracts as loans, Gubb argues, and the resemblance to pre-GFC 2008 — not the dot-com bust of 2000 — becomes hard to ignore: slow, orderly corrections belong to technology cycles; credit cycles seize.None of this is presented as certainty. Gubb is careful to flag where he's relying on estimates rather than disclosure, and he gives real weight to the Jevons-paradox counterargument that cheaper AI could simply unleash so much new demand that volumes swamp falling prices. But his own money is already positioned around the thesis: long Nvidia, Apple, Amazon, Snowflake, and Cloudflare; short a basket of Nasdaq exposure and the more-leveraged neocloud names..What makes the note worth your time isn't a prediction of when the boom breaks. It's the watchlist: OpenAI's next funding mark, the first quarter hyperscaler capex growth flattens, and the first capex cut that the market rewards rather than punishes. Those are the tripwires. Read the full note below and judge the arithmetic for yourself..Priced for a future that Token Economics forbidA sceptical view of current AI valuations.By James Gubb, July 17th 2026.This is a private discussion note. It’s my analysis, not investment advice. Prices and figures are as of mid July 2026 and sourced where stated; several material inputs are estimates the companies do not disclose. Technical and market terms are explained in footnotes at the bottom of the page on which they first appear, so a generalist reader outside the AI industry can follow the argument without too much prior background. And yes, Fable 5 helped me reduce the time taken to produce this note by around a factor of 8.The question is narrow and hard to dismiss. At current multiples the market prices a high and near-certain return on the largest capital programme in corporate history, while the potential range of outcomes is far wider than current pricing suggests. The CAPE¹ ratio around 41 is the second-highest reading in more than 140 years. Traditional fundamental stock market valuation ratios all flash deep red. Leverage in stock markets and sovereign debt are at historic highs. And equity markets in developed markets, despite all these warnings, continue to scale new all-time highs, bewitched by the biggest technical leap that mankind has ever produced, and now being made manifest with gusto. The four largest hyperscalers² are guiding to roughly $725bn of capex³ in 2026, up about 77% year on year, with consensus near $920bn for 2027 and parts of the sell side arguing consensus is too low. You do not need a crash thesis. You need only the observation that the probability of market expectations matching reality as I describe below, is priced close to zero. The argument is sequential; every step is observable today. I believe there is an opportunity to profit from this mismatch, timed and triggered by certain clearly defined events.1. The AI industry is a 4 Layered CakeExhibit 1a lays the industry out as a cake, and the shape of the cake is the first finding: the width of each tier is scaled to the market value of the companies in it. The base is chips and hardware: Nvidia designs the GPUs⁴, TSMC fabricates them, Broadcom builds custom silicon and three memory makers supply the high-bandwidth memory; roughly $8.5tn of market value, gross margins above 70%, and the only tier with a durable chokepoint. Above it sit the data centres and neoclouds⁵, the borrowers: the tier where the $725bn a year of capital is actually spent and where the debt sits, yet whose listed pure-plays are worth only some $0.4tn. Above them, the models: frontier labs⁶ spend billions of dollars of compute⁷ training the LLMs⁸ behind ChatGPT and Claude, carry private values near $2tn, and between them run at roughly $72bn of annualised revenue as of mid-2026, one of the two (OpenAI) still burning cash heavily. The top tier is applications and distribution, some $13tn of market value, where AI is finally sold to a paying customer by the token⁹ through an API¹⁰ or as a subscription, and where AI-attributable revenue is not yet material to anyone's earnings. The cash flows down the spine of the cake and the product flows up it. The candles on top are the products, and the moths circling them are the consumers. The cake is unstable by construction: the two profitable tiers are enormous, while the two tiers between them, the ones consuming the capital, are worth a fraction of the money passing through them. It is those narrow tiers that are the most vulnerable part of the cake..Two concepts complete the anatomy. The first is the bypass, drawn in Exhibit 1b and built from open weights¹¹ and harnesses¹². Some labs publish their trained models outright, so the capability is free to download, and a growing body of evidence says a cheap open model wrapped in a good harness, the scaffolding of prompts, tools and company data around it, matches an expensive closed model for most real work. An enterprise that takes the bypass routes around the paid model layer entirely: it runs free weights on rented compute and sells straight into applications. The model layer collects nothing. The very recent launch by Chinese president Xi of the latest open weight model Kimi-K3 raises the competitive stakes even further, with claims of surpassing the current coding king, Fable, on Code Arena, a coding industry benchmark. Open weight models are catching up very fast, with increased pressure on closed models to earn their keep. The second is the loop: to a significant degree, the demand in this industry is internal, with the machine feeding itself. This is dealt later in section 7..Exhibit 1c gives the scale of this capex build. The dotcom and telecom build-out of 1998 to 2003 spent roughly $500bn in total. The AI build of 2026 to 2030 is on course to spend roughly $5.9tn on the sell side's own path: nearly twelve times the dotcom cycle in nominal dollars, and roughly four times relative to the size of the economy of its day. Those are the bull's numbers, not the bear's. Everything that follows asks a single question of them: what needs to be true for the deployment of capital on that scale to earn a decent return?.2. Capability has commoditised, and the price sheet now says soThe AI industry is deeply competitive, particularly in the LLM space. This is no longer a two-horse American race. Frontier-adjacent models ship from China (DeepSeek, Moonshot's Kimi, Zhipu's GLM, Alibaba's Qwen, MiniMax), from US open-weight programmes (Meta's Llama, Nvidia's Nemotron), from Europe (Mistral) and from Israel (AI21), most under licences that let anyone download, run and modify them.The left panel of Exhibit 2a shows what four years of that competition has done, and it contains two stories that are easy to confuse. The three solid lines are the flagship prices of OpenAI, Anthropic and Google at each launch. They are almost flat. A vendor looking only at those lines would conclude that frontier pricing power is intact. The dashed gold line tells the other story. It holds capability constant, tracking the cheapest model that can do what the 2023 flagship could do, and its cost has fallen by roughly two hundred times in three years, from sixty dollars per million tokens to twenty-eight cents. Both lines are true at once. The frontier holds its price; what the frontier could do last year becomes almost free. The question for anyone underwriting this capital cycle is which of those two lines the mass of commercial demand sits on, and Section 3 argues it sits on the gold one..A necessary correction before drawing conclusions from any price sheet: a token is not a unit of capability, and per-token prices are not comparable across models. A frontier model packs more into each token and needs fewer of them; a cheap reasoning model may burn several times as many, thinking out loud and retrying, to reach the same answer; and vendors tokenise the same text differently. The honest unit is not the token but the finished job: dollars per problem solved, at a matched quality bar, counting every token consumed along the way.The right panel of Exhibit 2a does exactly that. Costed properly, on problems all six models actually solve, and scored on the same benchmark¹³, GPT-5.5 runs at roughly eleven times the cost of the cheapest adequate option, Claude Opus 4.8 at roughly ten times, Gemini 3.1 Pro at roughly five times. Put the other way: an enterprise doing work that an open-weight model can already handle is paying somewhere between 400 and 1000 % more than it needs to. Verbosity has been counted against the cheap models, and despite being handicapped for that, they still win by that margin. Kimi K3, the 2.8 trillion parameter Chinese heavyweight, recently became the first open-weight model to top the Frontend Code Arena, and the ProgramBench benchmarks, while running substantially cheaper than its US closed model competitors for comparable workloads, an advantage engineered through efficient architectures, cheap power and thin target margins, not dumping alone. When identical output costs ten times more from one supplier than another, the difference is not quality. It is margin, or it is subsidy, and fierce competition rapidly removes both.Where does the gold line stop falling? At the cost of the compute that performs the work: electricity plus amortised silicon¹⁴. For work an open-weight model can do, the capability itself is free to license; what remains billable is compute, and compute is a commodity that competition prices at cost, with cache-hit pricing¹⁵ already taking the repeated instructions of agent workloads to a few cents per million tokens. The price of getting a task of given difficulty done, quality held constant, therefore converges on the marginal cost of running it, which itself decays toward the energy floor. Tokens are merely the meter on the pipe; the task is the economic unit.3. The overshoot: capability now improves faster than our need for it growsPicture a monkey being fed oranges. You have most probably seen this on social media many times. It takes a bite, is offered a fresh one at no cost to itself, and discards the first on the ground, only to repeat the action immediately when a new orange is offered. Extravagantly wasteful, except that the monkey bears none of the cost of its waste. That is the AI model market today: a new flagship arrives roughly every quarter, priced at or below the old one, and the installed base discards a product of entirely decent utility because a better one is free to switch to. The subsidised entry pricing of Section 2 built the habit; the release cadence of new improved models feeds it.Now correct the metaphor, because the correction is the argument. Fruit spoils. Software does not. When the monkey drops the orange, it is spoilt; when the market drops last quarter's model, the model remains exactly as capable as the day it shipped, indefinitely, and where the weights are open it sits in a public orchard that anyone may pick from at near-zero price. The discarded oranges never rot. They accumulate, six a year, into a permanently fresh, permanently free inventory of capability below the frontier¹⁶. Every vendor of new oranges is now selling against that pile, and the pile is enormous.Exhibit 3a draws the consequence. The black staircase is flagship capability, a step every quarter or so since 2023, flattening as the data wall¹⁷ presses down: the stock of high-quality training text is exhausted, so each new step costs more to build and adds less. The green band describes what roughly 95% of all enterprise work comprises: the basic managerial functions of planning organising and controlling, in turn divided into activities like analysing, communicating, researching, reporting, forecasting, optimising and evaluating, that constitute most white-collar activity, and are highly suited for current AI capabilities. The band is drawn rising because needs of enterprises grow too. Anthropic's Fable 5 cleared that band around the middle of 2026. Every step since is sold to a shrinking audience of frontier buyers, while the price tags under the discarded steps show what each level of capability costs today: sixty dollars became twenty-eight cents. And you don't need to know quantum physics well to run Disney..This dynamic has a name: Christensen's The Innovator's Dilemma (1997) opens with an uncomfortable finding: great companies fail not through bad management but through good management, because listening to your best customers and chasing the highest margins is exactly what blinds you to the bottom of the market. He split progress into two kinds. Sustained innovation makes good products better for demanding customers, and the then dominant incumbents almost always win those battles: the frontier race in Exhibit 3a is a sustained battle, and the closed labs of OpenAI and Anthropic are winning it. Disruptive innovation enters at the low end, inferior on every traditional metric but cheaper, simpler and more accessible, and improves until it meets the mainstream's needs, which the incumbents have meanwhile overshot; performance oversupply is his name for that overshoot. Open-weight models running on a company's own hardware are the textbook disruptor at the lower end of the market: dismissed at first for hallucinating more and knowing less, but free once downloaded, available to anyone, and, decisively after Section 5, private, because the data never leaves the building. The cruellest part is the trap he called the value network, the web of customers, cost structure and investor expectations a business model is built around. The incumbents cannot fight down-market: their burn rates and their near-trillion-dollar private marks are written against premium margins, so matching open-weight prices means cannibalising the very revenue their valuations require. The rational move for every player is therefore to keep climbing the staircase even as it climbs away from the customers. Venture investors have already named the resulting price series: capability-adjusted model prices falling roughly ten times a year, the fastest sustained deflation ever recorded for a technology input.So ask the question directly: Is today's frontier model, a Claude Fable 5 or a GPT-5.5, good enough for most, say 95% of corporate tasks? If the answer is yes, the argument is essentially over: on the observed trajectory, that exact capability will list at roughly 3 to 5% of today's token cost within eighteen months, by around the end of 2027, and at 1 to 2% within two years. The buyer does not even need the cheap substitutes of Section 2 to exist; the frontier itself becomes the cheap model, on a conveyor, on a schedule you can read off four years of price data. And if the answer is no, the bull owes an answer to a harder question: which corporate tasks, specifically, does the current frontier fail at, how large is the spend attached to them, and why will the next six steps upward crack them when the last six did not?There is a precedent for how this all ends, and it is not the death of the technology. Between 1996 and 2002 roughly a trillion dollars of telecom capital laid fibre optic cable. After the crash, the great majority of it sat dark. Global Crossing and WorldCom died on the debt, the fibre was bought for cents on the dollar, and that discarded capacity went on to carry and serve YouTube and Netflix. The technology triumphed, the capital that built it was destroyed, and the stranded surplus became the plumbing of the next economy at a fraction of its build cost. That is the trajectory this exhibit describes: the discarded models become the everyday operating layer of enterprise, and the price at which they do so has scant chance of paying the training bill.So financing matters. An overshoot funded with equity destroys shareholders slowly; 2000 was survivable. This overshoot is increasingly funded with debt, take-or-pay contracts and GPU-collateralised loans, all sized to yesterday's price of capability. It has turned almost all (with the notable exception of Apple, I venture to say) the hyperscalers from free cash flow producing behemoths with decent moats and suitably high ratings to enterprises that have entered a record high entrance-cost race of uncertain duration and outcome, that has decimated both cash and competitive advantage. Debt funding is the next requirement to finishing the race. Equity can absorb a write-down. Debt written against a price that falls ten times a year cannot.4. The training bill cannot be recouped, and training itself has hit the data wallAgainst that falling price line stands a rising fixed-cost line. Each frontier model costs more to train than the last, the data wall of Exhibit 3a means each new one buys less improvement for the money, and the whole bill must be recovered from tokens whose price is collapsing toward cost, inside a selling window that the quarterly leapfrog cycle has squeezed to three months or less. Put plainly: a frontier model now becomes obsolete faster than it can pay for itself. The odds of recouping training spend at market token prices are slim, and every new entrant in Section 2 makes them slimmer.Worse still, LLM's can be copied. The practice is called distillation: training a cheaper model on the answers of a better one, so the student inherits much of the teacher's capability without its cost. Proposed as model compression in 2006 and formalised by Hinton in 2015, it went mainstream in 2023 when Stanford's Alpaca cloned GPT-3.5-class behaviour for a few hundred dollars of API calls, and in 2025 OpenAI accused DeepSeek of building its breakthrough partly on distilled OpenAI outputs. The Chinese open weight models have been widely accused of using distillation to rapidly catch up with their American competitors, at about 80% of their capability, for around 5% of their price. The implication for the closed labs is brutal: every API answer leaks capability out the front door, terms-of-service bans are unenforceable against foreign labs, and each billion-dollar flagship helps train its own cheap replacement..5. The moat is dissolving twice: by competition, and because the vendors became the vampiresThe bull case has already retreated once. "We own the best model" quietly became "we own the contracts, the distribution and the data" as Section 2's competition arrived. But a backlog priced on today's compute economics can be renegotiated the moment a competitor commoditises the compute underneath it; a contract is only a moat if the customer cannot get the same capability cheaper next year, and Sections 2 and 3 established that they can. That is the first dissolution, and it is about price.The second is about trust, and it is newer, faster, and harder to reverse. An enterprise that routes its work through a closed model sends its prompts, its documents and its workflows, the raw material of its competitive edge, through a vendor's systems. That was always an act of faith. Through 2026, a sequence of documented events, laid out in Exhibit 5a, gave enterprises reason to become very, very worried. Take them in order:Firstly, in 2020 the Wall Street Journal documented Amazon employees using individual third-party sellers' sales data, against Amazon's own stated policy, to build competing private-label products; the reported example was a seller's car-boot organiser that Amazon studied and then cloned. The EU charged Amazon that year over its use of non-public seller data, and Amazon settled in 2022 with binding commitments. The relevance is not that AI labs have done the same. It is that the platform playbook exists, is documented, and was executed by one of the most trusted companies in the world against the very customers who made its marketplace valuable.Secondly, April 2026: Microsoft restructured its OpenAI relationship, keeping its 27% equity while shedding exclusivity and its obligation to supply compute. On its face a corporate reorganisation; in context, the best-informed counterparty in the industry repositioning to hold the upside without the entanglement. A smart and telling move, imho.Thirdly, June 2026: Anthropic launched a services firm with Blackstone, Hellman & Friedman and Goldman Sachs, placing itself in direct competition with the consulting firms that had spent two years building businesses on its models. The vendor's own arm now bids against its channel.Fourthly, June into July: OpenAI was reported to be assembling the identical structure with TPG and Bain. Two data points make a pattern: the labs' path to revenue runs through their own customers' markets.Fifthly, July 2026, the flashpoint that moved the issue from trade-press grumbling to boardroom agenda: Figma pulled out of Anthropic's Claude Design launch and publicly accused it of deceptive communication, after the tool it had been asked to endorse emerged as a direct competitor to Figma itself. In the same weeks, Palantir's chief executive Alex Karp called the predatory dynamic out by name, citing clients' "rational concerns around proprietary insights migrating into the weights of closed models," and offered the cure in the same breath: Nvidia's open Nemotron models, run inside the client's own walls, where nothing leaves the building..I want to be clear about what this note does and does not show: there is no documented case of a lab extracting a customer's alpha¹⁸ from API traffic, and this note does not claim one. There does not need to be. The risk is unverifiable from the outside, the vendors are now visibly competing with their customers downstream, and the Amazon precedent shows how the platform-versus-customer story ends when the incentive is left in place. Faced with an unverifiable risk from an LLM that has total access to one's proprietary data, its alpha, with a demonstrated conflict of interest, the rational enterprise response is not to demand promises. It is to own the weights. That response has now been productised, and, in the most delicious double irony of the whole affair. The competing product is not only being sold by Nvidia itself, which profits identically whether the model running on its chips is closed or free, but is also being copied or "distilled" in the same way the LLM's copied the vast amounts of "free" data on the internet! Thieves stealing from thieves...The ramification for the closed labs is structural, not reputational. Section 2 showed the 95% of demand leaving on price. This section shows a slice of the remaining 5%, the largest, most security-conscious enterprises and governments who were precisely the customers able to pay frontier prices, leaving on trust. The customer base willing to pay a premium for a closed frontier model is being shrunk from both ends at once, and the trust end, unlike the price end, does not come back when the vendor cuts its rates.6. So, the question is ROIC, and the arithmetic does not closeDamodaran¹⁹ is right that the depreciation-schedule debate²⁰ is a sideshow; the main event is ROIC, or return on invested capital²¹. Roughly $725bn of 2026 capex, several trillion cumulative by the end of the decade on the sell side's own build-out arithmetic, is being deployed to sell a product whose price converges on cost. Aggregate hyperscaler cash capex is on track to exceed aggregate operating cash flow²² around Q3 2026, funded increasingly with debt at interest rates that are staying stubbornly high. The frontier labs the spend is partly built for reached a combined run-rate of roughly $72bn by mid-2026 (OpenAI ~$25bn and flat since February; Anthropic ~$47bn and compounding violently on Claude Code), still only about a tenth of a single year's infrastructure bill, and nearer 7-8% on recognised rather than annualised revenue. To earn even a utility-grade return on the cumulative programme requires hundreds of billions per year of durable revenue priced above cost, and "priced above cost" is exactly the property the competition is removing. The bet is not that AI fails; it is that a commodity earns commodity returns on record-priced capital.The cross-subsidy defence, that Search or Windows profits can carry the AI division indefinitely, fails on the conglomerate discount: a high-return compounder that becomes a low-return, capital-intensive operator does not keep its growth multiple, it trades down toward a utility's. And depreciation policy deserves one paragraph, as a clock rather than a thesis. GPUs booked on five-to-six-year lives against a two-to-three-year competitive cycle do not change the economics; they set the date the economics reach reported earnings. The understatement is self-concealing while capex grows and reverses when capex growth flattens. That is a signal to watch, and not the main argument..7. The financing is a credit cycle wearing a technology costumeEverything to this point is an argument about returns. The final layer, set out in a widely circulated July 2026 essay by the research blog Groundbreaker ("The Second Derivative"), is an argument about financing, and it explains why the unwind, when it comes, is unlikely to be gentle. The essay's frame deserves restating because it is the piece most analysis misses.Any growing quantity can be described by three numbers: its level, its growth rate, and the change in that growth rate, the second derivative²³. Markets watch the first two obsessively and the third almost never. Yet when a financing structure embeds a growth assumption, it is the third that kills it. The 2008 precedent is exact: US subprime mortgage delinquencies inflected upward in 2006 while house prices were still rising, because the loans were designed to be refinanced rather than repaid, and refinancing required prices to keep rising faster. The deceleration broke the machine a year before the level ever fell.The frame matters because it reclassifies the asset. The market prices AI as a technology cycle, whose risks are obsolescence and competition and whose corrections arrive as slow multiple compression: 2000 was that kind of event, painful but gradual, because equity is patient capital. The actual anatomy of this build-out is a credit-financed property cycle: data centres built with borrowed money and leased to tenants on long take-or-pay terms. Property cycles do not decline gradually; they seize, because credit covenants are step functions and the exposures are negatively convex²⁴. 2008, not 2000, is the relevant template, and 2008 broke on a deceleration in the rate of demand growth against the fixed supply the boom had just built.The AI build-out has quietly adopted the same architecture. A frontier lab signs a multi-year take-or-pay²⁵ contract for compute it has not yet consumed; the counterparty books the promise as backlog, or RPO²⁶, and borrows against it to pour concrete and rack chips. Strip the arrangement to its skeleton and it is a loan: the data centre is the collateral, the lab's contracted payments are the debt service, and record capex figures are better read as loan origination volume than as demand. On Groundbreaker's reconstruction, roughly $2.1 trillion of contracted backlog now sits across the big four platforms, and about half of it is owed by two loss-making labs, OpenAI and Anthropic: about 49% of Microsoft's book, 54% of Oracle's (roughly $300bn owed by OpenAI alone), 43% of Google's and 51% of Amazon's. Those composition figures are the author's estimates rather than company disclosure, but the direction is not seriously disputed: Wall Street is valuing as forward revenue what is, in economic substance, a concentrated credit exposure to borrowers with no operating income.The two borrowers are not the same credit. Anthropic is a risky credit, but an insulated one: roughly 80% enterprise revenue, about $1.70 of revenue per dollar of compute, guiding to profitability as soon as Q2 2026 on a $10.9bn quarterly revenue target, and its largest compute liabilities carry what amounts to a monoline wrap²⁷: in the reported ~$35bn Apollo and Blackstone financing for its Google TPU capacity, Google guarantees lease shortfalls and Broadcom the residual value of the silicon, so lenders are underwriting Google and Broadcom, not the lab. OpenAI is the naked borrower: roughly 60% consumer revenue, burn near 57% of revenue, cumulative cash consumption heading toward $115bn by 2029 on the essay's arithmetic, and no co-signer since Microsoft restructured the relationship in April 2026, keeping its 27% equity while ending the revenue share, exclusivity and its right of first refusal on supplying compute. Read that restructuring plainly: the counterparty with the best information in the industry kept the upside and walked away from the bills and is now reported to be renting neocloud capacity on five-year terms rather than owning through the trough. I repeat, smart move Microsoft! Behind it, SoftBank, the largest remaining backer, is reported to be seeking a $10bn margin loan²⁸ against its OpenAI stake secured by a personal guarantee; the lender of last resort is itself borrowing against the position. A company in OpenAI's position services its commitments not from earnings but from financing, and financing is available on one condition: that each round prices above the last. In Groundbreaker's five words, "the markup is the cash flow." The step-ups²⁹ tell the story: roughly 1.83x, 1.91x, 1.67x and 1.70x across 2024-25, compressing to about 1.23x implied by the reported target for a public offering that has now been delayed. The revenue line tells the same story: OpenAI's run rate sat flat near $25bn from February through the spring, with missed internal monthly targets reported, while Anthropic's ran from $19bn to $47bn over the same months. And the two forces compressing the step-up are precisely the forces of Sections 2 to 4: token-efficiency gains eroding revenue per task, and Chinese open-weight models repricing the commodity middle of inference, reportedly carrying over 60% of tokens on the largest independent routing marketplace at a fraction of closed-model prices..Meanwhile the lenders are crossing a line of their own. Aggregate hyperscaler capex has climbed from about 30% of operating cash flow in 2022 to roughly 100% on 2026 consensus; above that line every marginal dollar of capacity is funded from the balance sheet, and the group's debt load has already doubled in a year. And the second derivative has, on the essay's numbers, already turned: capex growth of +51%, +81%, +77% and +52% across 2024 to 2027E means the acceleration peaked into 2025 and is now falling (Exhibit 7a). The level is at records, the velocity is high, and the one number the structure is written on has rolled over. One recursion completes the picture, and it is easiest to see by following a single dollar. Amazon invests in a lab; the lab spends that money buying Amazon's cloud compute; Amazon books the spending as AI revenue and cites it to the market as proof of demand; the market rewards Amazon for building more capacity, and the cycle repeats. The dollar made a round trip and was counted as a customer on the way through. In any other industry this is called vendor financing. Nvidia's revenue is the hyperscalers' capex; the neoclouds' revenue is that capex with leverage added; and a meaningful slice of the labs' revenue began life as hyperscaler investment. Strip out the money that never left the machine, and the genuinely external demand, real customers paying from their own P&Ls, is a fraction of the headline the valuations rest on..There is a name for the moment this becomes visible in equity prices: the reflexive flip. The capex race holds only while the market rewards the next dollar of spending; the first time a hyperscaler cuts an announced capex number and its multiple expands on the news, every CFO's payoff table is destroyed at once, and discipline cascades as fast as the spending it replaces. The plausible first mover is Meta (founder control, the weakest direct monetisation of its AI capex, and a rewarded precedent for cutting); Google is least likely to blink, Oracle cannot afford to, and Amazon may have its hand forced by already-negative free cash flow.Why this matters for the blast radius rather than just the timing: OpenAI is, directly or one step removed, the largest customer of nearly every name in the build, more than half of Oracle's backlog and roughly two-thirds of CoreWeave's book traced through to the underlying tenant. The exposures are negatively convex: the same take-or-pay contract is an asset to everyone on the way up and a liability to all of them simultaneously the day belief reprices. And the paper has started to move: CoreWeave's $3.1bn GPU-backed facility syndicated in May 2026 was the first such loan built to trade in the secondary market, the originate-to-distribute³⁰ step that in 2008 spread a concentrated risk through the whole credit complex. The terminal refinance is the public market itself: an IPO³¹ whose S-1 is a trap by construction, because the document that unlocks the capital must disclose the audited losses, the customer concentration and the $600bn-plus of take-or-pay obligations that made the refinance necessary. The reported step, from a ~$852bn private valuation to above $1tn, must simultaneously clear a price the most unforgiving investor pool will pay and retire a cumulative burn heading toward $115bn. The essay's judgment, which I share, is that it cannot do both, and the delay is itself the signal..If it does break, the essay orders the sequence, and the order is worth internalising because it says where to look and where not to stand. First, the terminal refinance package is priced below the required step-up. Second the borrower pulls back on compute commitments, and a take-or-pay pull-back is not a lost sale but a covenant event for whoever collateralised the contract. Third the breach lands on the neoclouds (Coreweave and friends), spread businesses with no balance-sheet cushion that are insolvent by construction when the spread inverts, then on Oracle, which bleeds rather than seizes. Fourth credit freezes: backlog reprices from a demand signal to a counterparty exposure and GPU-backed notes cannot roll. Fifth, equity follows with capex repriced from a growth option to a fixed cost. Last, the strongest balance sheets buy the stranded capacity for a fraction of its build cost. Nothing physical breaks anywhere in that chain; belief reprices, and the single variable the whole of it keys off is whether OpenAI clears its next mark. That, not any valuation level, is why its recent IPO delay moved chip stocks..This section strengthens rather than replaces the thesis. Sections 2 to 6 say the returns will not be there; this says the financing breaks before the returns are even measured, on deceleration alone. My standing trigger was the first quarter in which capex growth flattens. The credit lens moves the tripwire one derivative earlier, and on the essay's arithmetic that wire has already been crossed.8. The exit door: what passive flows do to the way downOne final input, filed deliberately at arm's length because it is a short-seller's narrative rather than established fact: Carson Block of Muddy Waters has argued a demand-side scenario that bears on how any repricing propagates. The mechanics run as follows. The steadiest buyer of the US index is the retirement contribution flowing into passive funds³², and the marginal contributor to those flows is the well-paid knowledge worker. If AI displaces a meaningful share of those workers over the next few years, the flow flips sign: contributions stop and withdrawals begin. Passive redemptions sell every stock mechanically at index weight, and on the flow-multiplier research³³ the price impact concentrates in the largest constituents, the same mega-caps whose weightings the inflows built. The irony is exact: the AI winners would be hit hardest by the labour displacement their own technology causes, through the pipe their own index weight created.Handle it with care. Block is talking his book; the displacement timing is asserted rather than demonstrated; it dismisses the same Jevons rebuttal this note takes seriously; the multiplier's magnitude is contested; and retirement demography is a slow-moving beast that his scenario requires to move fast. What survives those discounts is narrower and still useful: it is not a trigger, it is an amplifier. If the financing mechanics of Section 7 start the repricing, mechanical passive flows are a credible reason the move is violent and concentrated in the biggest names rather than orderly and diffuse. For anyone hedged, the practical consequence is that protection must be structured so that it survives gaps: owned put options, insurance paid for in advance, rather than standing stop-loss orders, which assume someone will still be buying on the way down, in the size that you want. Fat chance of that happening. The scenario in which you need the floor is precisely the one in which liquidity evaporates, and prices gap down dramatically.9. What the current price requires you to believeFor today's multiples to be right, all the following must hold simultaneously. Each is possible. The current price requires almost all of them occurring simultaneously. At a cyclically adjusted multiple bettered only by December 1999, the market is charging near-certainty for a chain of individually unlikely events..The honest counterweight, and the triggerBears have been early before, and I have been very guilty of that over the recent past. Frontier-lab revenue is still compounding at extraordinary rates, hyperscalers report being supply-constrained rather than demand-constrained, a slice of capex is long-lived power and buildings rather than short-lived silicon, and the strongest rebuttal is Jevons³⁴: collapsing token prices could induce so much new demand that volume outruns price. Note what that rescue requires: volume growth steep enough to offset a price line converging on cost, sustained for years, accruing to the incumbents rather than to whoever serves open weights cheapest, and arriving fast enough to keep the refinancing treadmill of Section 7 turning. Three stretches could extend the borrowed time: a larger-than-expected private round; a sovereign or strategic backstop, and a proposal for a 5% US federal stake in OpenAI has already been reported, so this stretch may be in play; and the hyperscalers' remaining capacity to lever up, which seems unlikely, with the group's debt load already doubled in a year. The thesis does not need total acceptance of the timing; it needs only that the market is pricing the whole failure mode near zero, and I strongly believe that at these levels, it most certainly is.The watch list then, in order: First, whether OpenAI's next mark clears, privately or at the IPO, and at what step-up; this is the single variable half the backlog is wired to. Second, the first quarter in which aggregate hyperscaler capex growth flattens outright, the moment the arms-race logic loses its forcing function and the depreciation understatement stops being self-concealing. Third, the first capex cut that is rewarded with a higher multiple, the reflexive flip that ends the race; (the plausible first mover remains Meta.) Fourth, GPU rental rates, any refinancing stress at the neocloud layer (I think that is happening now, the weak share price of Coreweave is a giveaway), and the secondary-market price of the newly syndicated GPU-backed paper, the first domino by construction and now the first visible signal. Fifth, Oracle's credit, the most visible single-name fuse.Bottom line. AI capability is a commodity being priced like a franchise. Global competition, open weights and a trust break between vendors and customers are driving the realisable price of intelligence toward the cost of energy, while the cost of staying at the frontier rises into a data wall. The spend racing to build that capacity is financed like a property boom, on contracts that are loans in substance, extended to a central borrower who can pay only by refinancing at ever-higher marks, and the acceleration those marks require has already rolled over. Record capital is being deployed against that spread at the second-highest starting valuation in market history. The asymmetry does not favour the buyer at these prices.Position Declaration:Long: Nvidia, Apple, Amazon, Snowflake, Cloudflare.Short: Various duration and strike Nasdaq 100 put options, Coreweave, Nebius, Iren, SoftbankSubject to constant change..*Prepared as a private discussion note. Not investment advice and not a recommendation to buy or sell any security. The author is not a licensed investment adviser, just a retired Chartered Financial Analyst. Exhibit sources: provider list prices and benchmark trackers as of late June 2026; lab revenue run-rates from company disclosures and press reporting, May-July 2026; sell-side, Epoch AI and Groundbreaker ("The Second Derivative: Why No One Understands the AI Boom", July 2026) capex and backlog work as cited in the text. Backlog-composition and OpenAI cash-flow figures in Section 7 are that essay's estimates, not company disclosure. Section 8 summarises a scenario argued by Carson Block (Muddy Waters) and flow research associated with Michael Green, presented as a scenario rather than a claim of fact. The capability index in Exhibit 3a is illustrative and Exhibit 4a is a model, not a forecast; the prices, dates and capex figures shown in both are sourced as above..Sources and further readingInline claims are sourced below. Where a figure is an estimate or a reconstruction rather than company disclosure, the text says so at the point of use.Primary documents and disclosuresAnthropic, Series H announcement, 28 May 2026 ($65bn raised at a $965bn post-money valuation; run-rate revenue disclosed above $47bn, stated gross of hyperscaler reseller share). anthropic.com/newsAnthropic, "Anthropic expands partnership with Google and Broadcom", 6 April 2026 (run-rate revenue above $30bn; the TPU capacity programme wrapped by Google and Broadcom).Palantir Technologies and NVIDIA, "Palantir Launches Engine for Deploying NVIDIA Nemotron Open Models in Sovereign Environments", Business Wire, 29 June 2026 (source of the Karp quotation on proprietary insights and closed weights, and of the "domain-optimized harnesses… frontier capabilities" claim).Analysis and commentaryGroundbreaker, "The Second Derivative: Why No One Understands the AI Boom", Substack, July 2026 (groundbreakerre.substack.com). The framework of Section 7: the loan-book reading of backlog, the borrower split, the step-up series, capex against operating cash flow, and the ordered break sequence. Composition and cash-flow figures are that essay's reconstruction, not disclosure.Clayton M. Christensen, The Innovator's Dilemma, Harvard Business School Press, 1997 (performance oversupply; sustaining versus disruptive innovation; the value network).Guido Appenzeller, "Welcome to LLMflation", Andreessen Horowitz, November 2024 (capability-adjusted model prices falling roughly ten times per year).Aswath Damodaran, Musings on Markets (the valuation framing adopted in Section 6: depreciation schedules as a sideshow to return on invested capital).Epoch AI: Villalobos et al. on the exhaustion of high-quality public training text (the data wall), and 2026 analysis of aggregate hyperscaler capex against operating cash flow.Michael Green, research on passive flows and index-fund market impact (the flow multiplier of Section 8); Carson Block (Muddy Waters), public commentary on AI-driven labour displacement and passive flows, 2026. Both are presented in Section 8 as a scenario, at arm's length.Press reportingDana Mattioli, "Amazon Scooped Up Data From Its Own Sellers to Launch Competing Products", The Wall Street Journal, 23 April 2020; European Commission antitrust proceedings against Amazon (Statement of Objections, 10 November 2020; binding commitments accepted, 20 December 2022).The Information, reporting on Anthropic's application launches and enterprise-partner conflict, including Figma's withdrawal from the Claude Design launch, June-July 2026.Press reporting, March-July 2026 (multiple outlets): the April 2026 Microsoft-OpenAI restructuring; OpenAI's March 2026 financing at a reported ~$852bn valuation and the reported terms of its delayed public offering; the reported SoftBank margin loan; the reported proposal for a US federal stake in OpenAI; and CoreWeave's $3.1bn GPU-backed syndicated credit facility, May 2026.Prices and benchmarks: provider published price lists and public benchmark trackers (SWE-bench Verified, SWE-Bench Pro), as of late June 2026. Capex figures: company guidance and sell-side consensus for the big-four hyperscalers, as cited in the text..ReferencesConsolidated from footnotes appearing at the bottom of each page in the original document. Numbers correspond to the superscript markers in the text above.CAPE: the cyclically adjusted price-to-earnings ratio, or Shiller PE. The price of the S&P 500 divided by its average inflation-adjusted earnings over the previous ten years, which smooths out the business cycle. The long-run average is about 17. A reading near 41 has been exceeded only once, in December 1999.Hyperscalers: the handful of companies operating cloud computing at global scale, principally Microsoft (Azure), Amazon (AWS), Google (Google Cloud) and Meta, with Oracle increasingly included. They own most of the data centres on which AI runs.Capex: capital expenditure, cash spent on long-lived assets such as buildings, power equipment and chips. Accounting treats it as an asset that is written off (depreciated) over years, rather than as an immediate expense.GPU: graphics processing unit, a chip that performs thousands of simple calculations in parallel, which happens to be exactly what AI needs; Nvidia dominates the market. A CPU is the general-purpose processor in every computer. RAM is a computer's working memory; VRAM is the memory attached to the GPU itself, and the amount of VRAM is the main constraint on how large a model a machine can run.Neoclouds: specialist GPU-rental businesses such as CoreWeave, Lambda, Nebius and Crusoe. They borrow money, buy chips and lease out the computing, which makes them economically closer to leveraged equipment-leasing companies than to software firms.Frontier labs: the few firms training the most capable models, principally OpenAI, Anthropic, Google DeepMind, xAI and Meta. "Frontier" means at the limit of what is currently possible; a "frontier-adjacent" model is one within touching distance of that limit at a fraction of the price.Compute: industry shorthand for raw processing capacity: chips, the servers around them and the power that feeds them, bought outright or rented by the hour or consumed per token.LLM: large language model, software trained on vast amounts of text to predict the next fragment of text. ChatGPT, Claude and Gemini are products built on LLMs. Training (building the model, done once, enormously expensive) is distinct from inference (running the trained model to answer queries, a recurring cost per use).Tokens: the unit in which model work is measured and billed, roughly three-quarters of a word. Input tokens are the customer's prompt; output tokens are the model's reply, which cost several times more to produce. Prices are quoted per million tokens.API: application programming interface, the tap through which businesses buy model access by the token rather than through a chat window. The API price series is the published per-token price over time; API traffic is the customer's own data, prompts and replies, flowing through the vendor's systems.Open weights: a model whose trained parameters (its "weights", the billions of numbers that constitute the model) are published for anyone to download, run and modify, often under permissive MIT or Apache software licences. The capability itself is then free; only the hardware to run it costs money. Not the same as open source: the training data and code are generally not published.Harness: the software scaffolding built around a model: the prompts, tools, company data and checking logic that turn a raw model into a useful product. Much of an AI application's value lives in the harness, which is why a cheap model in a good harness often matches an expensive model without one.Benchmarks are standardised exams for models. SWE-bench Verified scores whether a model can fix real bugs in real open-source software, and is currently the most watched test of practical coding ability. SWE-Bench Pro is a harder variant.Serving or inference cost: the recurring cost of answering queries with an already-trained model, chiefly electricity and the amortised cost of the chips.Cache-hit pricing: providers charge far less for input tokens they have recently seen and stored ("cached"). Automated agent workloads resend the same instructions thousands of times, so their effective input cost collapses toward the cached rate.The knowledge or efficiency frontier: the line of best currently achievable capability. Pushing the line upward is invention; filling in gaps along it is the diffusion of what already exists. The distinction matters because diffusion is quickly copied and does not command premium prices for long.The data wall: frontier models have already been trained on essentially the entire stock of high-quality public text. Researchers at Epoch AI projected exhaustion in roughly this period. Once absorbed, further capability must be bought with synthetic (machine-generated) data and heavier post-training computation, at rising cost per increment of improvement.A customer's alpha: a firm's proprietary edge, its strategies, data, methods or trade secrets. The fear is that queries sent through a vendor's model could, directly or indirectly, inform products the vendor later builds in competition with that customer.Aswath Damodaran: professor of finance at NYU Stern, widely regarded as the leading academic authority on company valuation. His argument, adopted here, is that the depreciation controversy is secondary and the return on invested capital is the question that matters.Depreciation and the schedule debate: depreciation spreads an asset's cost over its assumed useful life as an annual charge against profit. The controversy: AI chips are booked over five to six years while the competitive replacement cycle runs two to three, and a longer assumed life directly flatters reported profits today.ROIC: return on invested capital, operating profit after tax divided by the capital tied up in the business. It is the test of whether a spending programme creates or destroys value: a company earning less than its cost of capital destroys value faster the more it grows.Operating cash flow is the cash a business generates from its operations. Free cash flow is what remains after capex, the cash genuinely available to shareholders. Once capex exceeds 100% of operating cash flow, every additional dollar of building must be funded with debt or new equity.The second derivative: for any quantity, the level is how big it is, the first derivative (velocity) is how fast it is growing, and the second derivative (acceleration) is whether that growth is speeding up or slowing down. Financing structures that depend on continual refinancing break when the acceleration turns down, long before the level ever falls. Subprime mortgage delinquencies turned up in 2006 while US house prices were still rising, because prices had stopped rising faster.Negative convexity: a payoff that is capped on the upside and unbounded on the downside, such as a lender earning a fixed coupon while bearing the full loss in default. Equity stories deflate gradually; negatively convex credit structures snap, because their covenants are step functions.Take-or-pay: a contract obliging the customer to pay for reserved capacity whether or not it is used. It converts what looks like usage-based revenue into a fixed, debt-like obligation on the customer.RPO or backlog: remaining performance obligations, the contracted future revenue a company has signed but not yet delivered. Wall Street reads a growing backlog as locked-in demand; it is in fact a promise, worth exactly the creditworthiness of the counterparty that signed it.Monoline wrap: borrowed from bond insurance, where a specialised insurer guarantees a security so that investors underwrite the guarantor's credit rather than the borrower's. Used here by analogy: guarantees from investment-grade companies sitting behind a speculative borrower's obligations.Margin loan: borrowing secured against a portfolio of shares. If the shares fall, the lender demands more collateral or sells the position. A margin loan against an illiquid private stake is the same structure with no easy exit.Up-round and step-up: a private funding round priced above the previous one; the ratio between the two valuations is the step-up. For a loss-making company funded round to round, the step-up is the lifeline: each markup is what pays the bills the last round could not.Originate-to-distribute and securitisation: making loans in order to package and sell them to other investors. Once the paper trades in the wider credit market, losses no longer stay with the originator. This was the plumbing of 2008.IPO and S-1: an initial public offering is the sale of shares to the public market; the S-1 is the disclosure document that must precede it, laying out audited losses, customer concentration and contractual obligations for any reader.Passive or index funds: funds that buy every stock in an index such as the S&P 500 mechanically, in proportion to size, whenever money flows in, and sell the same way when money flows out. They now hold the majority of US equity fund assets, and retirement contributions are their steadiest inflow.The flow multiplier: research associated with Michael Green argues that because passive funds transact regardless of price, a dollar of net flow can move the market value of the largest index constituents by many times that dollar. The size of the multiplier is contested; the direction, that mechanical flows amplify moves in the biggest names both ways, is not seriously disputed.Jevons paradox: the observation, from 19th-century coal economics, that making a resource cheaper to use can increase total spending on it, because demand expands by more than the price falls. It is the strongest single rebuttal to this essay.

Priced for perfection: A retired fund manager's warning on the AI boom

Related Stories

What fruit flies can teach us about the human brain

Ivo Vegter: Activists cast SA's data centre boom as a monster to slay

Nvidia's trillion-dollar reset: Cheapest since 2019 despite record profits

BN Power Lunch: Joburg's Budget crisis, the $3tn IPO wave, and what it means for SA markets