That is October in one company. The demand is real and accelerating. The bill for serving it has outgrown what even the richest builders earn, so they are bringing in lenders and new shareholders, and outside money arrives with a claim on the reward. My reading is that the market is becoming less willing to treat rapid AI growth as a complete investment case. The customers can be real, the business can grow, and the return to an existing shareholder can still disappoint. October sharpened two questions: what must the builders give up to finance their next expansion, and which businesses can earn without waiting for it?

When Intelligence Is Free researches how AI reshapes the economy: where profits disappear, where new opportunities emerge, and how to invest through the transition. These monthly reviews follow its capital cycle—from the rush to build, through the test of returns, to AI’s spread through the wider economy. October showed the strain of the first phase reaching its strongest builders.

The buyer of its own shares became a seller

Meta supplied the clearest reversal. In the first nine months of 2025 it spent $26.2bn buying back its own shares; in the same period of 2026 it bought back none. In the third quarter its capital spending of $34.9bn, including finance-lease payments, exceeded the $32.8bn its operations generated, its first negative free cash flow in more than a decade. It ended September with $82.5bn of cash and marketable securities against $80.9bn of long-term debt (Meta’s results). Then it started selling shares. On 27 October Meta priced $25bn of common stock at $625 a share through forward sales that settle by November 2027, and $12.5bn of preferred stock that converts into common shares in 2029. It has also created a non-voting share class, which its board says will give the company the ability to finance its growth, and plans a further programme of up to $60bn of sales once that class trades. The resulting $97.5bn headline adds together financing at different stages: the forward proceeds arrive on settlement, and the later programme is still a plan.

Amazon shows the scale behind such sums. Its cloud business grew 39%, its fastest in nearly five years, yet over the twelve months to September the company spent $42.4bn more on property and equipment than its operations produced, a gap Amazon attributes primarily to investment in AI. It still expects about $220bn of cash capital spending this year, and its long-term debt rose from $65.6bn at the end of 2025 to $173.1bn at the end of September. Even Microsoft, which expects to keep producing free cash flow this fiscal year, chose not to rely on it alone: on 29 October it sold $25bn of bonds, its first US investment-grade sale in more than nine years, into more than $100bn of orders.

None of this is distress. Every deal was oversubscribed, and Alphabet’s free cash flow over the past twelve months was still positive. These companies can draw down cash, borrow, sell shares, bring in partners or slow down. What has changed is that the pace they have chosen now requires an explicit decision about who finances it, at the top of the industry rather than only among more stretched borrowers such as CoreWeave and the lenders behind Oracle-leased data centres, as the Q3 review described. For existing shareholders that decision has a price. Lenders are paid first; new shareholders take a slice of every future dollar. Either can still enrich existing owners if the new capital earns more than it costs. Cloud growth proves that customers want the service. It does not prove that the next dollar invested will clear that hurdle, and a bigger AI business can be a worse investment per share.

The customers are paying, and the clouds carry the machines

The strongest evidence against a story of empty speculation arrived in the same week. Anthropic’s registration statement for its stock-market listing, filed on 26 October, showed revenue of $16.3bn in the first half of 2026, against $1.2bn a year earlier, and well above the $9bn it spent on computing and infrastructure. Operating cash flow turned positive, with $3.9bn coming in against $1.5bn going out a year before, and the second quarter was its first with an operating profit. The six-month figures are unaudited, the company still recorded a small operating loss for the half, and it holds $131.7bn of cash raised from investors. Even so, the filing gives the boom something more substantial than another forecast: a fast-growing business collecting cash from its customers.

Set that beside the platforms and the roles look reversed. A quarter ago, the model companies looked like the borrowers in this story and the cloud platforms like the bank behind them. In October, the first big lab to open its books reported cash coming in, while the platforms went to the capital markets. That is not a coincidence. A model company mostly rents its computing; the cloud that rents it out has to buy the chips, buildings and power. Cash can improve at the tenant while the investment burden grows at the landlord. The lab’s progress is real, but it moves the construction bill rather than removing it.

Nor can every company’s revenue be added up as independent demand. A business pays a model company, which pays a cloud, which buys equipment. Each transaction is genuine revenue, but the same money passes through several sets of accounts, and some of the customers at the start of the chain are spending investors’ money rather than cash their own businesses generate. What matters is how much sustained customer spending reaches the system, what it costs to serve, and what remains to reward the capital committed. October strengthened the evidence that customers will pay. It also showed how much more capital the builders are committing before anyone knows whether that payment will be enough.

Paid to build, or paid to run?

The most useful signal of the month appeared inside the infrastructure trade itself. From the end of September to the 30 October close, chip stocks fell 14.2%, measured by the VanEck Semiconductor ETF. Makers of grid equipment fell as hard or harder: Hyosung Heavy Industries, a Korean transformer maker, lost about a third of its value, and Powell Industries, which makes switchgear for utilities, industrial plants and data centres, fell 15.3%. Owners of infrastructure already running went the other way. Talen Energy, which sells power from operating plants, including to an Amazon data-centre campus beside its Susquehanna nuclear station, rose 9.7%, and Kinder Morgan’s gas pipelines rose 12.3%. Renting an existing Nvidia H100 chip got more expensive through the selloff: Silicon Data’s neo-cloud rental index rose from $2.78 per GPU-hour on 3 October to $2.92 on 31 October.

These businesses are usually lumped together as AI infrastructure. October pulled them apart along a single line. An equipment maker depends on customers placing orders for the next project. A power plant, a pipeline or a rented chip collects revenue from capacity that already exists. Rising rental prices alongside falling chip shares expose the same distinction from another angle: demand for today’s installed machines can stay strong while investors mark down the returns they expect from supplying tomorrow’s.

A month of share prices cannot prove that financing caused each move, and the distinction is about dependencies, not safety. An equipment maker may have a protected backlog, service revenue and customers outside AI. A pipeline’s steady fees make it trade partly like a bond, so some of its gain had nothing to do with AI, and an operating plant may have debt to refinance, contracts to renew or prices that regulators can rewrite. “Already built” does not mean “safe.” What the month offers is a question I want to carry into every piece of research: how much of this company’s value requires another expansion to be financed, and how much is supported by customers using what already exists? If financing gets harder while usage keeps growing, those two claims can move in opposite directions. That creates opportunities on either side: an operating asset with durable pricing power, or an equipment supplier whose share price has fallen much further than its likely earnings.

Still building, but the test has started

My judgment at the October close remains late Buildout: companies are still committing capital to expand capacity faster than its returns can be proven. No large builder cut its plans. Alphabet said spending would rise significantly again in 2027, Meta raised the bottom of its 2026 range to $135bn, and Amazon kept its $220bn. The physical bottleneck has not eased either, since chips already installed are renting for more, not less. Digestion, the phase in which the industry has to show that what it built earns its cost, has not arrived.

What October changed is the route there. Shareholders do not have to wait for empty data centres to impose that test. If the next expansion needs more outside money, takes longer to pay back or leaves less profit for each existing share, they can simply pay less for the shares today. That is how revenue can accelerate while valuations fall, and why the test of returns can begin before the spending stops. Monetary policy is not about to solve the problem on the builders’ timetable: the Federal Reserve held rates at 3.75–4.00% on 28 October, and two of its officials voted to raise them.

The case for buying October’s fear

The strongest objection is that October offered exactly the prices investors should want. On this view the builders are doing what great companies should do: raising money while investors are eager, preserving flexibility and financing demand that is accelerating. Microsoft’s bonds drew four times the orders it needed, and Anthropic’s filing showed paying usage at scale. If the new capacity fills at attractive margins, the extra earnings will more than compensate for the new shares and the interest. Selling because a company raises capital would then mean selling just as it gains the means to grow. Pessimism about AI financing has become popular enough that the market sold the fastest cloud growth on record, which is usually a better sign for buyers than for sellers.

I take that case seriously, and the disagreement is narrower than it looks: it is about the return on the next dollar invested. I would become more constructive if operating cash increasingly covered investment at Alphabet, one of the portfolio’s largest holdings, and at Meta when they report the fourth quarter in late January and February, without any slowdown in customer growth. Serving each additional dollar of revenue with progressively less new capital, and rising profit and cash per share, would show that existing owners are benefiting too. The opposite would move the evidence toward Digestion: share or bond sales absorbed only at falling prices, or pulled altogether, and spending cuts explicitly blamed on disappointing returns.

What October did to the portfolio

For the month ended 31 October 2026, the portfolio returned −2.28% net, against −6.65% for the Nasdaq-100 price index, both in US dollars. It lost money while losing less than the benchmark, and the attribution makes the result less comfortable than a simple claim that the cycle view worked.

Official month ended 31 October 2026; USD; unaudited. Contributions are rounded independently.

The largest loss had nothing to do with AI demand. It came from Macro, chiefly Strategy Inc., the bitcoin-holding company, and silver, and it cost 5.08 points in a month when the Fed leaned toward tighter money. Expecting financial pressure eventually to provoke a policy response does not make the positions built for that response protective while the pressure is rising. The Reserve, which holds cash, short-term bonds and put options for protection, paid off as chip stocks fell and contributed 2.99 points, with AI beneficiaries adding 1.05. Physical infrastructure, which holds businesses on both sides of the line described above, cost only 0.40 points in a month when chips fell 14%.

At 31 October 2026, Reserve was 33.86% of the portfolio, Macro 29.27% and Physical infrastructure 13.10%. Macro has come down from 36.8% at 30 September, but the mismatch remains visible: more of the portfolio sits in monetary bets than in the physical bottlenecks central to the research, and one favourable month does not resolve that. The reserve preserves room to act if financing pressure forces good assets to change hands at better prices. It also has a cost if the boom continues and those prices never arrive. That is the choice the portfolio should be judged on, alongside the assets it owns.

Who gets paid for the wait

The next reports need to answer more than whether demand grew again. Nvidia reports in late November, and its filing will show how much more it is lending to customers, or investing in them, to keep orders flowing. Anthropic’s listing, expected during the month, will show what public investors will pay for a model company’s growth. Meta’s new share class becomes effective on 19 January, opening the door to its larger programme of share sales. And the fourth-quarter results in late January and February will show whether the builders’ cash after investment improves, on what terms they raise more, and whether suppliers’ contracts can survive a slower pace of new construction.

Go back to 29 October. Selling the fastest-growing cloud business in the world looks irrational if the question is whether AI pays. It looks rational once the question becomes who gets paid. The richest builders can still raise all the money they want, but every dollar they raise gives someone new a claim on the outcome. The most useful opportunity may be a business whose customers keep paying even when they stop expanding. To find it, follow the payment: who owes it, what earns it, and does another financing round have to close first?

When Intelligence Is Free is independent research on where economic value moves as intelligence becomes abundant. Explore the framework or see how the thesis is expressed in the live portfolio.

Concentrated, high-volatility strategy; capital may be lost. Past performance does not guarantee future results. Nothing here is investment advice or an offer.

Subscribe now

That changes how this boom might break. We have been watching for spending to outrun demand. The quarter revealed another route: customers keep arriving, but some owners cannot afford the interval between winning an order and delivering it. Useful assets can change hands without the technology failing. For the next owner, the previous owner’s financing problem may become an opportunity.

We remain in late Buildout, the phase in which capital races to install capacity. What changed this quarter is the route into Digestion, when expansion slows and prices, debts and ownership adjust. Delivery and financing stress could bring that adjustment even while demand grows. Beyond it lies Diffusion, when cheaper, widely available AI spreads through ordinary businesses. These phases can overlap; a loss for today’s builder need not be a loss for tomorrow’s user.

To see why, follow the customer’s payment down the chain and the obligations back up it.

When Intelligence Is Free researches how AI reshapes the economy: where profits disappear, where new opportunities emerge, and how to invest through the transition. These quarterly reviews follow its capital cycle—from the rush to build, through the test of returns, to AI’s spread through the wider economy. Q3 showed how stronger demand can coexist with a more fragile financing structure.

The order looked different from each desk

From the customer’s side, the story was becoming easier to believe. A business buying an AI tool cares whether it saves time, improves a product or does work worth paying for. It does not need to settle the investment case for the entire infrastructure supporting that tool. The disclosures arriving during Q3 made those individual decisions harder to dismiss.

Google’s July earnings commentary reported cloud revenue growth of 82% and model-API throughput rising from roughly 16bn to 22bn tokens per minute in a quarter. Private-company figures pointed the same way, with less visibility: Reuters, citing a person familiar with the matter, put Anthropic’s annualised revenue above $65bn at July’s end, versus $47bn in May. Its 29 September report on OpenAI put the figure near $70bn, also citing an unnamed source. Those run-rates are reported estimates, not audited annual revenue or profit. They nevertheless reinforced the evidence that customers were paying.

The same order looks different at the model company. It brings revenue, but fulfilling it requires computing. At the cloud provider, it becomes a reason to secure more chips, buildings and power. At the chip supplier, it helps justify manufacturing commitments arranged well before the final customer has paid for years of service. Follow one payment through the chain and several companies can report revenue from it; follow the obligations in the opposite direction and someone must finance the distance between today’s construction and tomorrow’s receipts.

Nvidia’s August filing made that distance tangible. Supply and capacity commitments rose from $119bn to $279bn in one quarter, and some investment-grade customers received payment terms of 90 days to a year. Nvidia also disclosed August guarantees capped at $105bn for specified lease and power obligations at an SB Energy campus serving an OpenAI affiliate. They generally become effective as successive leases begin, with the first phase expected in Nvidia’s fiscal 2029; payment depends on certain tenant defaults. They cover only defined obligations and can terminate, including if OpenAI attains a satisfactory credit rating. The cap is neither cash spent nor an estimate of loss. It shows how selling the equipment can also involve supporting the buyer’s ability to house and power it.

There is no need to invent a villain in that chain. The customer wants useful software, the lab wants computing, the cloud wants to keep the customer, and the supplier wants the next order. Each can make a reasonable decision while the system becomes more dependent on everyone delivering on time. Strong demand keeps the expansion going; it does not remove the financing risk.

The lender had a different calendar

September made that distinction harder to ignore. The Federal Reserve raised its target range to 3.75–4.00%, increasing the pressure on new borrowing and refinancing. A project owner may expect excellent demand over the next decade; a lender still needs to know who pays before the project opens.

Project Jupiter illustrated the uncertainty. On 18 September, Reuters relayed a Financial Times report that approximately $18bn of loans behind the Oracle-leased New Mexico development were quoted at 89–91 cents on the dollar. Six days later, Reuters reported, citing a person familiar with the matter, that Oracle had issued a force-majeure notice concerning possible power-delivery delays. On 30 September, Oracle and STACK said the project remained on schedule. The discounted loan quotes preceded the reported notice; neither establishes a default or a missed opening.

This is evidence of financing and delivery risk, not proof that Jupiter has failed. The mechanism matters beyond this project: a full order book cannot switch on an unfinished facility. If delivery slips, the interval without operating income lengthens while the financing bill continues. Demand can survive that interval more easily than the owner’s balance sheet.

There is also substantial evidence against an imminent industry-wide break. Oracle reported 850 megawatts of capacity delivered in its fiscal first quarter, cloud-infrastructure revenue up 121%, and a completed $20bn equity sale. Capacity is reaching customers and capital is still available. That is why the cycle remains in Buildout: financing has become more demanding, but expansion has not broadly stopped. Whether operating cash arrives before obligations become unbearable remains an open contest.

How the paths changed

The quarter did not simply make every adverse outcome more likely. Faster monetisation improved the chance that the industry could grow through its commitments, even as dearer money reduced the prospect of an early rescue. Ranked by our assessment at September’s close, the near-term paths now look like this:

The biggest change was away from expecting rescue and towards having to endure expensive money. Treasury buybacks do not overturn that reading: the announced programme supports market liquidity, rather than promising to cap yields. Making bonds easier to trade is not the same as making financing cheap.

The reset path weakened across Q3 even though it strengthened again in September. Keeping that distinction matters in a quarterly review. Rising demand was real counterevidence, not an inconvenient detail to remove from a bearish story. What became harder to defend was reliance on policy easing before financing strain had done damage.

Over the next decade, our leading path strengthened modestly: useful AI spreads, while physical and institutional constraints retain value. The possibility of a prolonged capability plateau was little changed, as was that of a sudden capability leap accompanied by political and monetary disruption. An abundance shock that rapidly dissolves today’s scarcities became less persuasive.

That is compatible with intelligence getting cheaper. Epoch’s September research estimated that the cost of a given performance level across five benchmarks had fallen about 47% per quarter since 2023, while cautioning against simple translation into business outcomes. Cheaper intelligence encourages use; it does not automatically make power or dependable deployment abundant. The quarter strengthened the case for adoption without showing that all the constraints on serving it had disappeared.

The profit hid the exposure

The public book provides an uncomfortable test of that argument. It gained 8.74% in Q3, against 0.44% for the Nasdaq-100, but much of the advantage came from a strong August in bitcoin-linked and silver exposure. September returned −0.79%, against the benchmark’s +3.23%. A profitable quarter did not mean the public book was well aligned with the businesses whose prospects were improving.

At the September close, Strategy Inc. (MSTR) represented 22.54% of the public book and silver 14.26%. Both can have a long-term monetary rationale. Neither is interchangeable with protection against the immediate consequences of expensive money.

MSTR combines bitcoin exposure with corporate financing and a share price that can command a premium to the underlying assets. In a squeeze, bitcoin can fall, that premium can contract and access to attractive new financing can weaken together, while claims senior to common shareholders remain. The company’s own disclosures distinguish those senior claims from common-equity exposure. An eventual monetary rescue does not protect the shareholder from the route taken to reach it.

Silver presents a different mismatch. Its monetary appeal sits alongside substantial industrial demand. In a credit squeeze, that industrial dependence can hurt just when protection is needed. Physical infrastructure is no automatic refuge either: a useful asset can still be overvalued, heavily financed or exposed to falling demand. Tangibility does not make a share price defensive.

The implication for Q4 is to reduce dependence on a monetary rescue and preserve the capacity to buy through a squeeze. The lesson of Q3 is that a correct long-term thesis offers little comfort if the claims chosen to express it cannot survive the journey.

Where the next owner could find value

Return to the customer’s order. It still needs computing and electricity, regardless of who owns the building. If financing forces a sale, the asset may carry on serving the same demand under an owner with less debt and a lower purchase price. The economic usefulness survives while the distribution of returns changes.

That is where the WIF proposition that what AI consumes can gain value becomes concrete. As intelligence gets cheaper and use expands, dependable electricity, connected sites and capacity already in service can become more valuable relative to promises of future delivery. But scarcity reaches the shareholder only if the business can capture it. A generator able to reprice output has different economics from one locked into a long, fixed-price contract; an operating facility has different risks from a site still waiting for power.

The next advantage may therefore belong to businesses that combine delivered capacity, pricing power and the ability to fund their own growth. Owners dependent on repeated refinancing face the opposite pressure. A useful asset is not necessarily a good investment at today’s price, but a forced sale can change that price without destroying the use. The opportunity is to distinguish a broken financing structure from a broken business.

Further into Diffusion, value can also move beyond the builders. Cheaper AI can let a company serve customers or perform work that was previously uneconomic. Keeping the gain will require something competitors cannot obtain merely by buying the same model: customer access, trust, specialised data or operations that are difficult to replicate. Falling input costs enlarge the opportunity; competition decides who retains the profit.

The two possibilities belong to the same story. A reset can lower the cost of the infrastructure, helping the technology spread. The industry’s next phase may be financed partly by losses borne by its previous owners.

The ending we must be willing to miss

The strongest objection is that no disruptive transfer of ownership may be necessary. Revenue could keep catching up, delivery could hold, and receptive capital markets could carry the builders through. A more liquid public book would then miss some of the gains. A genuine monetary easing could also reward the very exposures whose timing risk now looks uncomfortable. Preparing for strain has an opportunity cost.

The next earnings rounds can challenge the argument. Revenue converting into cash, promised capacity entering service and continued access to funding would strengthen the case that the industry can grow through its obligations. Broader project-debt stress, weaker financing access or capital-spending cuts would strengthen the case for Digestion. A genuine shift towards monetary easing would weaken the policy-trap reading; routine liquidity operations alone would not. These are tests of the interpretation, not reasons to explain away whichever outcome arrives.

Q3 brought the customer, the builder and the lender into the same story. The customer could be delighted while the builder’s shareholder lost money; a supplier could book record business while taking on more of the risk required to produce it. The important question is increasingly who can afford to stay in the chain until the receipts arrive.

We finish the quarter more convinced by the uses of AI, but more attentive to who will own the assets that serve them. If attractive assets eventually change hands, we want to have the money to buy them.

When Intelligence Is Free is independent research on where economic value moves as intelligence becomes abundant. Explore the framework, see how the thesis is expressed in the public investment book, or apply the same reasoning to a company or market decision through advisory.

Concentrated, high-volatility strategy; capital may be lost. Past performance does not guarantee future results. Nothing here is investment advice or an offer.

A century and a half later, telecom repeated the trick with glass. The internet was obviously growing, carriers borrowed heavily to lay fibre ahead of traffic, and when some customers ran out of their own financing capacity, equipment makers helped finance the customers buying their equipment. A Canadian telecom operator, for example, disclosed a $315 million Lucent vendor-financing facility. Nortel described providing not only equipment finance but money for installation, working capital and even equity, and disclosed more than $5 billion of customer-financing exposure at the end of 2000. Its subsequent SEC filing shows what happened when capital markets closed: third-party placement became difficult, customers failed financing conditions, purchase commitments fell and provisions against the loans rose sharply.

The revenues were real while the financing lasted. Then refinancing stopped, telecom equity was destroyed, carriers went through bankruptcy and huge amounts of fibre changed hands for a fraction of what they had cost to lay. The internet did not fail. The capital structure did. And the wreckage became an input into everything that followed: broadband, cloud computing, streaming and eventually today’s AI industry inherited communications capacity that somebody else had financed at the wrong price.

That is the distinction this essay is about. The founding essay asked where value goes when intelligence gets cheap. There is a second question underneath it: when does that value move? The answer runs through the capital cycle.

Subscribe now

1. Every revolution has a time path

Infrastructure-heavy technologies have an awkward property: they have to be built before anybody can know exactly what they will earn. You cannot wait for railway traffic before building the railway, for internet traffic before laying fibre, or for an AI-native economy before building the power plants, substations, fibre, cooling systems, data centres and compute required to run it.

So capital finances the future backwards. An expectation supports a valuation; the valuation makes equity available; equity builds the first assets. Debt follows because it is cheaper and allows the buildout to move faster. When corporate balance sheets become constrained, project finance and private credit take more of the load. Customers prepay. Suppliers extend terms or help assemble pools of capital around their buyers. This is not a defect in technological capitalism. It is one of the reasons a new infrastructure system can be built in years rather than decades.

But it creates a predictable sequence. Buildout is when the scarce thing is infrastructure and capital earns by supplying what everybody suddenly needs. Digestion starts when the market stops asking how fast capacity can be built and starts asking what the capacity already built actually earns. Projects are cancelled, refinanced or transferred; assets once valued on scarcity get valued on cash flow. Diffusion is what the whole exercise was for: capacity exists, the cost of using it has fallen and ordinary companies can finally redesign themselves around the technology. The gain moves from building abundance to using abundance.

The most expensive mistake in a transition like this is therefore not being wrong about the technology. It is being completely right about the destination and standing in the wrong phase.

AI is still in Buildout. Hyperscaler spending has not rolled over, physical bottlenecks remain binding and financing markets remain open. But the thing that should weaken before spending itself does is already moving: the funding channel is becoming more expensive and more complicated. That does not mean Digestion has begun. It means the bridge toward it is becoming visible.

2. Follow the marginal dollar

Two years ago, the answer to “who pays for AI?” was almost boring. Microsoft, Amazon, Alphabet and Meta had enormous cash flows from existing businesses and strategic reasons to spend them. Internal cash is unusually patient money. Nobody calls it back because an enterprise AI programme took two years longer than expected.

That answer stopped being sufficient. By early September, hyperscalers had issued more than $200 billion of debt in 2026, already more than twice their 2025 issuance. Amazon’s latest bond sale is one visible piece of that migration. Investor appetite remains substantial, but it has become less automatic: Apollo found hyperscaler bond cover ratios falling from almost five times in February to below two by July. Its data show investors demanding more compensation as supply grows.

Oracle shows what comes next. In fiscal 2026 it generated a record $32 billion of operating cash flow and spent enough on infrastructure to leave free cash flow negative $23.7 billion. It raised $43 billion of debt and $5 billion of equity. At the same time, its backlog reached $638 billion, with $75 billion of large AI contracts already structured around customers either prepaying Oracle for GPUs or supplying the hardware themselves. Oracle’s own year-end release lays out all three sides of that equation: extraordinary demand, extraordinary capital intensity and customers increasingly helping finance the assets required to serve them.

And then Oracle supplied some useful counterevidence. On September 10, its next quarter showed backlog rising again to $664 billion, while $11.36 billion of $28.5 billion of quarterly capex was offset by customer prepayments. Management also said much of its newest contracted demand would not require proportionate new Oracle capital. The September quarter therefore cuts both ways. It shows that the financing problem is real enough for contract structure to change, but also that strong customers can move part of the burden away from Oracle’s own balance sheet. A capital cycle is not a one-way ratchet.

Then the supplier entered the financing story. On August 10, Nvidia announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR intended to mobilise more than $500 billion of third-party capital for AI infrastructure. Nvidia’s announcement explicitly describes the goal as expanding the pool of capital available for compute infrastructure. Nvidia is not lending half a trillion dollars itself, which matters. But another fact matters too: the company selling the scarce equipment now considers the financing capacity of its buyers important enough to organise some of the world’s largest pools of capital around it.

Its own balance sheet points in the same direction. Nvidia ended the July quarter with $63.1 billion of receivables and 60 days of sales outstanding, up from 45 days, as it extended longer payment terms on large multi-quarter purchases by investment-grade customers. Its SEC filing says those terms can range from 90 days to one year and explicitly notes that financing arrangements and credit support will continue to affect operating cash flow. The disclosure is unusually direct.

Put the pieces together and the marginal dollar has been moving outward:

1. Operating cash flow from highly profitable technology companies.

2. Equity and corporate bonds issued against their future earnings.

3. Project and private finance issued against individual infrastructure assets and contracted cash flows.

4. Customer prepayments that fund the hardware required to fulfil the contract.

5. Supplier-supported finance, credit support and longer payment terms that make it easier for customers to keep buying.

These are not five mandatory stages that every company passes through in order. The important observation is simpler: more of the buildout now depends on capital with a price, a maturity date or both.

CoreWeave makes the issue unusually easy to see. Its August facility runs roughly five years while the customer contracts supporting it average about three. CoreWeave presents that, reasonably, as evidence that lenders are willing to underwrite renewal risk. The company’s own announcement says exactly that. It is also exactly what it sounds like: the currently contracted revenue expires before the loan does.

At June 30, CoreWeave’s quarterly revenue had more than doubled to $2.575 billion, while net interest expense reached $640 million. The Q2 filing shows the scale of both the growth and the financing burden. The growth is real. The financing is real. They run on different calendars.

That mismatch is completely ordinary for a toll road. A road can support thirty-year debt because the useful life of the road is the least controversial assumption in the model. A rack of accelerators is not a road. The machine may still switch on perfectly in 2031; what nobody can tell you today is what an hour of its output will be worth.

3. Revenue is real. That is not the same as saying the economic loop has closed.

This is where the AI debate becomes unnecessarily binary. One side sees huge AI revenues and concludes that monetisation is proven. The other notices capital circulating among the same companies and calls the revenue fake. Both shortcuts miss the interesting question: where did the customer’s dollar come from?

Suppose a manufacturer earns $100 selling machines and spends $5 of the resulting cash on AI because AI reduced its costs. That dollar entered the AI economy from outside. It is external economic demand. Now suppose an AI company raises $10 billion and spends several billion of it on cloud compute. The cloud provider has earned perfectly real revenue, but the payer is funding the purchase with new capital rather than cash generated by its own customers. That is financed demand. The AI company pays the cloud, the cloud buys accelerators, and the accelerator supplier may invest back into the lab or infrastructure provider. Those are internal stack payments. Sometimes disclosure simply does not tell us where the payer’s money ultimately came from, leaving the provenance unresolved.

All reported revenue remains revenue. But aggregate economic demand should be counted only once, where money enters the AI stack from outside it. Otherwise the same dollar can appear economically larger every time it moves between a lab, cloud provider and supplier.

Microsoft gives us an unusually useful glimpse. Its 2026 annual report says Microsoft owns roughly 25% of OpenAI on an as-converted basis, had funded $11.9 billion of a $13 billion investment commitment, and simultaneously recorded $24.1 billion of revenue from commercial arrangements with OpenAI, with $6 billion still receivable at year-end. Microsoft discloses the related-party economics directly in its 10-K. Azure delivered real services and Microsoft earned real revenue. But the payer is also a company partly financed by Microsoft and other participants in the same capital cycle.

Amazon produces an even larger version of the same pattern. By June it carried $28.7 billion of OpenAI preferred stock; AWS had expanded OpenAI’s cloud commitment from $38 billion by another $100 billion over eight years. Amazon had also invested another $10 billion in Anthropic during the quarter while Anthropic expanded its AWS commitment to more than $100 billion over ten years. Those figures come from Amazon’s own filing. Alphabet has its own version: Anthropic reportedly committed about $200 billion to Google Cloud and related chips while Alphabet was simultaneously investing up to $40 billion in Anthropic. Reuters reported the relationship in May.

None of this is evidence of accounting fraud. Strategic investment, customer finance and supplier finance have existed for centuries. The telecom analogy matters for a much more mundane reason: revenue funded by a customer’s operating profits and revenue funded by a customer’s latest capital raise look much more similar on the supplier’s income statement than they do when the financing window closes.

The balancing evidence is equally important. OpenAI’s annualised revenue run rate reportedly passed $40 billion in August, driven by paying consumers, enterprises and coding products. Bloomberg’s reporting shows how quickly genuine outside demand is scaling. Microsoft says its cloud business exceeded $214 billion of annual revenue and nearly 90% came from customers outside frontier-model companies; all sequential RPO growth in its latest quarter came from customers outside the frontier labs. Microsoft’s earnings call is unusually helpful on this point.

So AI has real end demand. The unresolved question is whether outermost economic demand is growing as fast as the infrastructure obligations being accumulated upstream. The industry publishes exquisite statistics on tokens, model scores, GPU shipments and contracted backlog. The number I most want barely exists: what percentage of the marginal AI infrastructure dollar is ultimately being paid from cash generated outside the AI capital cycle?

That is the number Digestion will force the market to find.

4. The technology can move too fast while the customer moves too slowly

The monetisation problem is not one problem. Three mechanisms push on the same future cash flow from different directions.

Enterprises move slowly

Enterprise adoption is real and already broad. Stanford’s 2026 AI Index found 88% of surveyed organisations using AI in at least one function, while agent deployment remained in the single digits across almost every business function. The distance between those numbers is more important than either number alone.

An employee asking a model to draft a presentation is adoption. A bank rebuilding credit underwriting around agents is diffusion. The first takes an afternoon; the second runs into old databases, fragmented permissions, security controls, regulatory rules, procurement, works councils, budget cycles and managers with perfectly rational reasons not to redesign a process that currently works. This is the electrification problem all over again: the large productivity gains did not arrive when factories merely swapped steam for an electric motor, but when the factories themselves were redesigned around what electricity made possible.

AI distributes much faster than electricity ever could, so the analogy has limits. Organisations are still organisations. One calendar is measured in model releases, another in transformation programmes, and infrastructure debt sits between them with its own maturity date.

The product improves too quickly

Waiting for enterprise diffusion would be easier if intelligence itself were economically stable. It is not. Stanford’s technical-performance data show frontier systems gaining around 30 percentage points in a year on Humanity’s Last Exam, while OSWorld accuracy rose from roughly 12% to 66.3%. Benchmarks expected to remain difficult for years are being consumed in months.

For society, this is excellent news. For infrastructure underwriting, it creates a nasty question: how much compute will the economy need in 2029 to produce one unit of useful intelligence? One possibility is Jevons: intelligence gets dramatically cheaper, workloads multiply even faster and total compute demand explodes. The opposite is that efficiency outruns new workload creation and a business process requiring 100 units of compute today needs ten by the time an enterprise finally puts it into production.

Both can be true across different workloads. Today’s infrastructure is therefore financed against the intersection of two curves that are moving violently: how much intelligence the economy wants and how much compute is required to produce it. A 2026 accelerator does not compete only with other 2026 accelerators. It competes with 2029.

China is attacking the clearing price

Then there is China. The American strategy has effectively been to spend vastly more money, build the best models on the best hardware, and recover the investment because superior intelligence commands a premium. That works exactly as long as the premium survives.

The numbers are getting uncomfortable. Stanford says the U.S.–China performance gap between leading models had fallen to just 2.7% by March 2026, while its economic data say U.S. private AI investment was about 23 times China’s. The technical comparison is here; the investment comparison is here. Those are different measures, not a return-on-capital calculation. The juxtaposition is still hard to ignore: twenty-three times the private investment for a frontier whose measured performance lead is now in the low single digits.

Price differences are much larger than capability differences. DeepSeek currently lists V4 Pro at $0.435 per million uncached input tokens and $0.87 per million output tokens. Its current API price sheet is public. OpenAI’s GPT-5.5 is $5 per million input and $30 per million output tokens. OpenAI publishes those prices here. They are not identical products and raw token pricing is not a quality-adjusted benchmark. But on output price alone the gap is more than thirtyfold, and most corporate workloads do not need the world’s best model at any price.

A German manufacturer, Brazilian bank or Indian retailer does not get paid for winning an AI benchmark. It gets paid for producing an economic result. If almost all the useful capability for a task can be purchased at a small fraction of the price, frontier intelligence stops being a strategic necessity for that task and becomes a procurement choice.

China also sits on a very different physical base. It generated about 10,578 TWh of electricity in 2025 against 4,520 TWh in the United States. Pew’s compilation of Ember data shows the comparison directly. China remains constrained by advanced chips and high-bandwidth memory, but those constraints have also given its labs unusually strong incentives to squeeze more intelligence from the compute they can access.

Elon Musk made the geopolitical consequence unusually blunt in his July interview with The Economist: the two hard constraints are chips and electricity; China has a credible path to leadership; and Washington can stop American companies using Chinese models but cannot stop the rest of the world using them. The relevant part of the interview is transcribed here.

The strong conclusion is not “China wins AI.” Chinese open weights can run in an American data centre, which could damage the U.S. model provider while leaving the American power connection, fibre and building economically valuable. Model weights can cross an ocean in an afternoon. A substation cannot. The narrower and harder claim is that China can help set the global clearing price of intelligence without owning the whole stack.

That clearing price is what today’s buildout eventually has to earn against.

5. Put the two sides together

The financing side is pulling more future cash flow into the present while the monetisation side is being pushed the other way. Enterprises may take years to reorganise enough of their businesses to produce the full economic return. When they finally do, better models may require less compute for the same task. When they buy that intelligence, Chinese and open competitors may have pushed its market price far below the assumptions embedded in today’s investments. And during the waiting period, some of the largest buyers of infrastructure are still financing their spending with capital raised from the same ecosystem.

None of this requires AI demand to collapse. That is the point. The dangerous outcome for a leveraged buildout is strong adoption that arrives late, uses less expensive compute and pays lower unit prices than the capital structure assumed. Users can win, AI usage can explode and capability can keep improving while the infrastructure’s first owners earn terrible returns.

In fact, AI may create its own financing problem precisely by succeeding at its core economic mission: making intelligence abundant.

That is how Buildout eventually turns into Digestion. It rarely happens because everyone simultaneously decides a technology was useless; it happens because the next investor refuses to finance yesterday’s assumptions at yesterday’s price. A refinancing comes back wider, a customer signs three years instead of ten, a lender reduces the collateral value of used accelerators, a project requires more equity and gets postponed. Higher financing costs then reduce project values, weaker collateral reduces borrowing capacity, fewer projects reduce marginal equipment demand, and falling equipment values make lenders more conservative again.

Nothing in that sequence requires the models to get worse. This is why “AI bubble” is too crude a description. A bubble says the thing was mostly an illusion. A capital cycle says the thing was extremely valuable and financed at the wrong price.

Those stories have very different endings.

And importantly, we are not at that ending yet. Capital remains available, hyperscaler capex has not rolled over and physical bottlenecks still bind. CoreWeave closed its financing; Nvidia assembled new pools of capital; Oracle just demonstrated that customer prepayments can materially reduce its own financing requirement. The window is getting more expensive and more inventive, but it is still open.

That makes the funding migration useful as a clock, not as a crash call.

6. A reset would not hit “AI infrastructure” equally

Calling the whole buildout “AI infrastructure” hides a crucial difference. A data centre is not one economic asset. Part of it can last decades: land with the right permissions, transmission access, grid connections, substations, fibre routes, cooling systems and sometimes generation. Another part lives on a technology treadmill: accelerators, servers, networking generations and software layers whose advantage can disappear in a few hardware or model cycles.

A reset separates those layers brutally. A 2026 GPU whose economic life turns out to be three years instead of six can lose most of its value while the site housing it becomes more valuable because getting another powered site takes five years. A lab can lose its model premium while the electricity supplier keeps collecting rent. Chinese weights can run on American compute. The software becomes abundant; the physical permission to run it does not.

That is why the handoff from Digestion to Diffusion matters more than the crash itself. When railway equity collapsed, the track stayed. When telecom companies failed, the fibre stayed. If the first AI capital cycle breaks, the useful physical layer does not disappear either. Some unfinished projects will die, but enormous quantities of functioning infrastructure will simply acquire new owners and new prices.

A data centre that was mediocre at a $20 billion valuation can be excellent at $8 billion. Compute that cannot support the first owner’s financing cost can become an extraordinary input for the second owner. An application that makes no economic sense when inference costs $30 can become obvious at $1.

The first wave pays to manufacture abundance. Digestion removes the return assumptions attached to manufacturing it. Diffusion inherits what survives at a lower cost basis.

That is not a side effect of the capital cycle. It may be the mechanism by which the technology becomes cheap enough to transform everything else.

7. There is one more balance sheet underneath all of this

Normally a private credit cycle has one final source of optionality: the state. The United States is entering this one with less room than usual. Federal debt crossed $40 trillion in August, including roughly $32.3 trillion held by the public. Treasury figures reported by Reuters show the milestone and the composition. The Congressional Budget Office projects a $1.9 trillion deficit for fiscal 2026, equal to 5.8% of GDP, with debt held by the public rising from about 101% of GDP this year to 120% by 2036 under current law. The CBO baseline is here.

The same capital market is therefore being asked to finance two extraordinary things at once: the accumulated promises of the American state and the largest private infrastructure programme of the AI era. The government has options, but none make the underlying arithmetic disappear.

Treasury has already increased its liquidity-support buybacks in longer-dated securities. Its August announcement raised the planned long-end operations from $2 billion to at least $4 billion each; in September it increased the first 10-to-20-year operation to as much as $6 billion. Reuters reported the expansion here. These are debt-management operations intended to improve market liquidity, not quantitative easing. That distinction matters. What also matters is that the functioning of the long end has become something Treasury is actively spending balance-sheet capacity to improve.

If a private AI financing problem eventually becomes large enough to demand intervention while inflation remains elevated, the policy menu becomes uncomfortable. The state can guarantee infrastructure, become a customer or alter regulation so balance sheets can carry more of it. Fiscal policy can cut spending or raise taxes. The central bank can eventually ease if inflation permits. And if market-clearing rates become politically or fiscally unbearable, some combination of monetary accommodation and financial repression can move the adjustment away from defaults and toward purchasing power.

None of these choices makes an economically bad asset produce an economically good return. They change who absorbs the loss.

Ray Dalio’s debt-cycle framework is useful here in this narrow mechanical sense. When debt service consumes more income and the supply of claims outruns buyers willing to absorb them at prevailing rates, either rates rise enough to choke activity or money is created to prevent the adjustment. His 2026 HBR discussion lays out the mechanism directly. His broader argument about changing powers should be treated more cautiously, because China has severe debt, demographic, property and capital-market problems of its own and the dollar has no obvious reserve-scale replacement.

But there is one uncomfortable intersection between the two stories. America could finance the world’s most expensive first generation of intelligence abundance while Chinese competition helps determine how little the final product sells for. The United States can still benefit enormously from that outcome: its enterprises get cheaper intelligence, consumers get better products, and scarce American power and infrastructure may remain valuable. What can disappear is the premium expected by some of the capital that financed the frontier.

The country can win. The technology can win. The customers can win enormously. The original investors can still lose.

That is not an exotic outcome. It is what happened before.

8. The clean escape route is also the hardest one

There is an elegant way for the system to work without a serious reset: AI simply works fast enough. Enterprise productivity accelerates, falling intelligence prices unlock so many workloads that total compute demand outruns every efficiency gain, and normal businesses start sending operating cash into the AI stack faster than the stack needs fresh financing. Hyperscaler cash generation catches up with capex, frontier labs become increasingly self-funding, refinancing spreads stabilise and infrastructure earns its cost of capital even as intelligence gets cheaper.

That outcome could also help the sovereign side. A sufficiently large productivity shock raises real output, lowers production costs, expands the tax base and creates the technological deflation needed to offset some of the inflationary pressure from public borrowing. Today’s extraordinary infrastructure spending would then look less like an excess and more like the bridge that arrived just in time.

It is possible. It also requires several clocks to line up unusually well: capital markets move quarterly, hardware depreciates over years, models improve every few months, large enterprises reorganise over several years and transmission infrastructure can take close to a decade. The system does not need perfect synchronisation. It does need the cash to arrive before enough providers of capital decide they have waited long enough.

The strongest ways this essay could be wrong are therefore straightforward:

1. Cash flow reclaims the buildout. If operating cash flow begins covering more of capex and reliance on debt, project structures and supplier-supported capital falls, the financing clock gains years.

2. Outermost economic demand accelerates. If enterprises and consumers increasingly fund AI from their own operating cash rather than frontier companies continuing to depend on fresh capital, the payer-provenance problem resolves in the bullish direction.

3. Jevons wins decisively. If intelligence becomes ten times cheaper and that creates one hundred times more compute demand, efficiency strengthens the infrastructure thesis rather than weakening it.

4. Enterprise restructuring happens unusually fast. AI distributes through software and cloud infrastructure that already exist, giving it an enormous advantage over railway construction or factory electrification.

5. The cycle grinds rather than breaks. Private credit can amend and extend, hyperscalers can cancel marginal projects before credit deteriorates, strong balance sheets can absorb weak ones and governments can move private risk onto public balance sheets.

6. China fails to turn low prices into global economics. Chip controls, memory constraints, regulation, security concerns and geopolitics could preserve a much larger premium for the Western stack than current model prices suggest.

The most likely way to be wrong, however, is the boring one: the mechanism is right and the calendar is early. Telecom vendor finance did not collapse the moment it appeared. It extended the cycle. Being right about a capital cycle and wrong about when it turns is still being wrong.

9. Five numbers matter more than another benchmark

The first number is who funds the marginal dollar of capex. If infrastructure spending moves back toward operating cash flow, the system is becoming healthier; if it continues migrating toward credit, project structures, prepayments and supplier-supported finance, the first clock is running further.

The second is where the customer’s dollar originated. How much incremental AI demand is funded by external enterprise and consumer cash, how much by customers themselves consuming fresh capital, how much is internal circulation through the stack, and how much remains impossible to classify from public disclosure?

The third is actual enterprise cash return. Counts of people using copilots are becoming almost useless. Margin improvement, revenue per employee, operating savings, throughput and cycle-time reductions matter much more.

The fourth is the race between efficiency and consumption. If useful intelligence becomes ten times cheaper while compute consumption rises one hundred times, today’s infrastructure bulls win. If consumption rises three times, a lot of underwriting has to be revisited.

The fifth is the global clearing price of intelligence. Chinese model quality, open weights, inference pricing, accelerator efficiency and electricity availability matter because American AI does not ultimately compete only against another American model. It competes against the cheapest acceptable way to produce intelligence anywhere a customer can legally and practically buy it.

And underneath all five sits the cost of long-duration capital.

Those measures will tell us more about the time path than another argument over whether AGI arrives in 2028 or 2032.

The strange way technology becomes cheap

We like technological history to have a clean sequence: someone invents something useful, companies adopt it, productivity rises and everyone becomes richer. The real sequence is much less polite. Capital sees a future before the future exists and races ahead of it. Equity turns the story into companies; credit turns the companies into physical infrastructure; competition keeps building after the obvious opportunities have already been funded. Expectations that existed only in spreadsheets eventually become steel, copper, silicon and debt.

Then the future actually arrives. It arrives later than somebody promised, with better technology than somebody modelled, using less of some inputs and more of others, and at prices that were not in the financing documents when the debt was issued. That is when the capital structure finds out whether being right about the future was enough. Usually it wasn’t. Assets get repriced, owners change and the useful infrastructure remains, leaving the second generation with something the first generation never had: abundance without having to pay the original cost of creating it.

That is why this story is more interesting than “AI bubble” versus “AI boom.” AI can work. Enterprise demand can become enormous. Models can continue improving. Intelligence can become one of the most important factors of production in the economy. And the people who financed its first version can still lose money.

Railway investors financed transportation abundance and failed to capture much of it. Telecom investors financed bandwidth abundance and failed to capture much of it. Their losses were real, and so was everything that became possible because they had overbuilt. AI may repeat the pattern: the people financing intelligence abundance today may discover that abundance is precisely what destroys some of the economics they financed.

If that happens, the reset will not mean the AI buildout failed. It will mean it worked well enough to move into the next phase. The first owners finance scarcity; Digestion discovers what their assets are actually worth; Diffusion gets what survives.

Which leaves the same contradiction we started with, and perhaps the one worth remembering:

AI can become one of the best things that ever happened to the economy and one of the worst investments made by some of the people who paid to build it.

Those statements can both be true.

They have been before.

When Intelligence Is Free is independent research on where economic value moves as intelligence becomes abundant. Explore the framework, see how the thesis is expressed in the public investment book, or apply the same reasoning to a company or market decision through advisory.

In August, CoreWeave asked lenders to make a similar judgment. It closed a $2.6bn facility lasting roughly five years, supported by customer contracts averaging about three. The financing extends beyond the average contracted customer commitment. Renewals, new customers and the prices they pay therefore matter to the outcome. CoreWeave’s 10 August announcement.

The analogy is about timing, not a claim that AI equipment has the same useful life or resale value as a building. Those economics have to stand on their own. This was the development I found most revealing in August. The buildout was finding ways to buy time for future earnings to arrive. Whether that time would prove sufficient remained open.

When Intelligence Is Free researches how AI reshapes the economy: where profits disappear, where new opportunities emerge, and how to invest through the transition. These monthly reviews follow its capital cycle—from the rush to build, through the test of returns, to AI’s spread through the wider economy. August showed how the first phase could last longer.

The lender had a reason to say yes

There is an easy version of this story in which every extension of financing signals trouble. The evidence does not justify it.

CoreWeave’s facility was oversubscribed. The company said shorter customer contracts could command higher prices and bring in a broader enterprise customer base. A lender can reasonably accept a gap between contract length and loan maturity if it expects the capacity to keep earning at attractive rates. The question is whether that expectation proves right.

The company was also putting capacity to work. Its second-quarter results, published in August, reported 1.5GW of active power at 30 June, after adding nearly 500MW during the quarter. This was an operating expansion as well as a financing story. Installed capacity is not proof of good lifetime returns, but it is more concrete than a promise to build. CoreWeave’s results.

In the successful version, customers find profitable uses for AI, utilisation stays high and contracts renew on attractive terms. Earnings grow into the investment. Financing has done the useful job it was meant to do. That possibility matters to anyone expecting an imminent reset. A business does not have to fund every new asset from today’s cash flow for the investment to make sense.

The supplier offered time too

Another part of the same story appeared in Nvidia’s quarterly filing. Revenue for the quarter ended 26 July reached $96.2bn, up 106% year on year. Equipment demand remained strong. The filing also described payment terms ranging from 90 days to one year for certain investment-grade customers making large data-centre purchases. Nvidia’s quarterly filing, published 26 August.

The company selling the equipment was also helping finance the interval before payment. Such terms can sensibly match cash payments to deployment. They also mean the supplier carries more of the wait for cash; the filing explicitly linked financing arrangements to the timing of operating cash flow.

CoreWeave’s lenders and Nvidia were not taking identical risks. One arrangement depended in part on business beyond existing customer contracts. The other gave selected customers longer to pay. But both helped the expansion continue by changing when cash had to arrive.

The distinction between selling more equipment and earning an adequate return from using it therefore remained central. Supplier revenue is real. It is not, by itself, proof of the buyer’s eventual economics. Nor can revenues at several stages of the supply chain simply be added together as independent end demand.

The wider financing conditions mattered

Alongside these company decisions, Treasury moved to support liquidity in the government-bond market. On 19 August, it announced that maximum liquidity-support buybacks in the 10-to-30-year nominal sectors would rise from $2bn to at least $4bn per operation, beginning 9 September and running through 4 November. Treasury’s announcement.

This was a separate development, with an indirect connection to AI. Long-term government debt helps anchor borrowing costs for projects expecting cash flows years away. My inference is that conditions in that market could affect how financial pressure reaches the buildout. Treasury’s stated purpose was market liquidity—not support for AI—and the action guaranteed neither cheaper corporate borrowing nor successful projects.

At the August close, my cycle judgment remained Buildout. Capacity was coming into service, equipment sales were growing and financing remained available. A broad turn toward Digestion, when returns on installed capacity constrain further investment, was not established.

August told us more about how the expansion might continue than when it would end. Its lenders and suppliers were willing to finance more of the wait. The next stage still depended on what customers could earn.

What this meant for the public book

For the month ended 31 August 2026, the public investment book returned +9.95%, against +4.18% for the Nasdaq-100. Macro Scenario Bets contributed +8.84 percentage points, while Reserve & Optionality contributed −1.35 points.

The result came predominantly from the monetary and real-asset side of the book. Attribution identifies those exposures’ contribution; it cannot prove that a policy announcement caused the gain. These remain risk positions that can lose alongside equities in a reset.

At 31 August, Macro Scenario Bets represented 35.4% of the book and Physical Scarcity 17.2%. Compute & Platforms stood at 12.9%, compared with 16.0% at 31 July. Moonshots moved from 0.0% at 31 July to 2.2% at 31 August. The strong return left the gap between the prominence of physical infrastructure in the research and its weight in the book unresolved.

Time has to become earnings

The next evidence I want is less about the size of a new financing announcement and more about what happens after it. Are customers turning AI spending into recurring earnings? Are suppliers collecting cash on the terms they extended? Are projects securing finance on assumptions about utilisation and prices that remain credible?

More cash generation covering investment would make the expansion more durable. Weaker project returns, tighter lending and postponed capacity would move it toward Digestion. Physical assets with lasting scarcity may remain valuable through either path, but ownership price and financing terms still determine the investment case.

A loan can carry the landlord beyond the end of today’s leases. It cannot make tomorrow’s tenants pay enough. August gave the AI buildout more time to make that part of the story work.

When Intelligence Is Free is independent research on where economic value moves as intelligence becomes abundant. Explore the framework, see how the thesis is expressed in the public investment book, or apply the same reasoning to a company or market decision through advisory.

Concentrated, high-volatility strategy; capital may be lost. Past performance does not guarantee future results. Nothing here is investment advice or an offer.

Think of a hotel owner whose rooms are full. There is an obvious case for adding another wing. But the construction bill arrives before the new guests do. Bookings can make the expansion look attractive without providing the cash to build it. The more ambitious the owner becomes, the more important the financing becomes.

That was July’s problem in miniature. The evidence for building grew stronger. So did the question of who could afford the wait.

When Intelligence Is Free researches how AI reshapes the economy: where profits disappear, where new opportunities emerge, and how to invest through the transition. These monthly reviews follow its capital cycle—from the rush to build, through the test of returns, to AI’s spread through the wider economy. July’s chapter was still about expansion, with a more demanding financing story underneath it.

The customers were waiting

Microsoft’s July results gave the builders a reason to keep going. Azure and other cloud services grew 43% year on year in the quarter ended 30 June, and demand still exceeded available capacity. That is a difficult environment in which to argue that the industry has already built enough. It is also a cloud-wide measure, not a clean reading of AI profitability. Microsoft’s earnings call.

Even an apparent spending retreat needed a second look. Microsoft’s lower headline capital-spending forecast reflected a change in lease classification; the company said its underlying calendar-year investment expectation was unchanged. The headline moved. The intended investment did not.

My judgment at the July close therefore remained Buildout: the phase in which capital races to create capacity. A broad turn toward Digestion, when returns on installed capacity begin to constrain further investment, was not established. The customers were still waiting for the rooms.

The financing bill was growing

Oracle showed the other side of that expansion. On 9 July, S&P lowered its rating to BBB−, the lowest investment-grade tier. The concern was the upfront investment needed to serve an expanding AI business, its pressure on cash flow and uncertainty about eventual returns.

Oracle’s latest annual accounts made the tension concrete. It had generated $32bn of operating cash flow in the year ended 31 May 2026. After investment, however, it reported negative free cash flow of $23.7bn. These figures were already public; July’s downgrade sharpened their credit implications. A substantial cash-generating business was undertaking an even larger expansion. Oracle’s year-end release.

This is where an investor’s experience can diverge from a customer’s. The customer sees more capacity coming. The investor has to ask what financing it will cost, how long the cash will take to arrive and what return will remain after that wait.

If lenders demand more compensation, a project can become less valuable without losing a single prospective customer. The technology need not disappoint for the capital financing it to be repriced. The financial strain can arrive before an operating slowdown.

Who is paying for the expansion?

There is another reason to be careful with the growth figures. A customer pays an AI provider; the provider pays a cloud; the cloud buys equipment. Each transaction produces real revenue for its recipient. Adding all those revenues together does not produce an equal amount of independent demand. Some of the same customer money has been counted at several stages.

The unanswered question is how much cash enters from end customers, what margins are earned serving them, and how much investment is needed to keep doing so. If the original payer is itself spending money raised from investors, access to new capital remains part of the story. July strengthened the evidence that customers wanted capacity. It did not settle the return on all the capital being committed to supply it.

Yet there was a credible way for the expansion to work. Microsoft reported $19.6bn of company-wide free cash flow for the June quarter, after its large investment programme. It also said all sequential growth in commercial contracted obligations came from customers outside frontier-model companies. Neither disclosure isolates AI’s profitability, but both matter: a builder with other profitable businesses and a broader customer base has more ways to finance the wait. Microsoft’s financial commentary.

The hotel owner may complete the new wing, fill it at attractive prices and repay the loan comfortably. In AI, ordinary businesses may keep increasing paid usage while the largest suppliers fund expansion from their broader earnings. Builders can also defer weaker projects. A long expansion with selective disappointments remains a serious alternative to a broad reset.

What this meant for the public book

For the month ended 31 July 2026, the public investment book returned −0.32%, against −6.61% for the Nasdaq-100. Reserve & Optionality represented 28.0% of the book at that close, while Physical Scarcity represented 18.0%.

The allocation helped explain why the book behaved differently from an equity index during the selloff. It does not establish that the cycle view was correct. For the month ended 30 June 2026, the book had returned −14.43%, against −0.19% for the benchmark; July’s resilience did not undo that loss. Physical infrastructure also remained a smaller part of the book than its prominence in the research would suggest.

The next chapter depends on the cash

The next test is whether financial strain changes what gets built. Spending cuts driven by weaker expected returns, postponed projects and easier access to previously scarce capacity would move the evidence toward Digestion. Stronger customer cash generation, improving project economics and stable financing terms would favour a longer Buildout.

July left us with a distinction worth carrying forward: a business can become more useful to its customers while becoming more demanding of its investors. The task is to separate durable earning power from dependence on generous financing—and to ask what price compensates for the difference.

The queue outside the hotel tells us there is demand. It does not tell us what the next wing will earn after paying for its construction.

When Intelligence Is Free is independent research on where economic value moves as intelligence becomes abundant. Explore the framework, see how the thesis is expressed in the public investment book, or apply the same reasoning to a company or market decision through advisory.

Concentrated, high-volatility strategy; capital may be lost. Past performance does not guarantee future results. Nothing here is investment advice or an offer.

Much of civilisation was organised around that limitation: who received an education, who could afford expertise, which problems deserved the attention of the few people capable of solving them. An extraordinary amount of human ambition died waiting for someone with the time and ability to pursue it. We are beginning to change that. Something we once had to raise can increasingly be manufactured. What happens to an economy built around its scarcity?

My starting conviction is that intelligence will become abundant faster than the world can reorganise around it. That gap can create fortunes. It can also ruin investors who mistake a temporary shortage for a permanent right to charge. When Intelligence Is Free is my investment research into how this transition unfolds, which parts of the emerging economy are worth owning, and how to invest in them through public markets.

Subscribe now

The most expensive thing we learned to copy

We have been copying the products of intelligence for centuries. A book lets a million readers benefit from one person’s discoveries. Software repeats a procedure as often as we need it. But when the next problem required fresh judgment, we usually had to return to a person. Knowledge could travel much further than the mind that produced it; the supply of people able to put it to work remained a constraint.

AI begins to separate useful cognitive capability from the person supplying it. Developing a model can cost a fortune, but serving its next customer does not require another childhood, education and career. It requires computing resources. Better chips and more efficient models reduce the resources needed for a given result. The frontier can become more expensive to create while yesterday’s frontier becomes cheaper to use.

The decline is already measurable. In The plunging price of thought, published on 22 September 2026, Epoch AI estimated that the cost of reaching a given performance level had fallen about 47% per quarter since 2023, roughly thirteenfold per year. These are benchmark results, not the full cost of dependable work inside a business. But the speed at which expensive capabilities become inexpensive inputs is extraordinary.

Competition brings that improvement to the buyer. In May 2026, Epoch estimated that leading downloadable models trailed frontier closed models by about four months on its aggregate benchmark measure. A customer does not need the best model ever created for every assignment. Once several systems can do the job, yesterday’s miracle becomes something buyers can shop around for. Epoch’s cost curve assumes buyers keep choosing the cheapest suitable model, while real users switch less often: cheaper intelligence can arrive long before the saving reaches their software bill.

My working assumption is that over the next decade AI will do essentially all the work we now do at a computer better and far more cheaply: writing software, analysing accounts, designing products and running everyday operations. As that intelligence moves into robots, much physical work should become cheaper too. “Almost free” describes a world where skilled work can be bought for a fraction of today’s cost. That changes what people can afford to attempt, and what businesses can charge to do it for them.

Which part of the invoice survives?

A company adopts AI, lowers its costs and expects to keep the saving. Then its customer acquires the same capability. So does its competitor. What looked like a margin improvement inside one business becomes a renegotiation across an industry. The work still matters. The old price no longer commands the same authority.

There is an intellectual inheritance here. Joel Spolsky explained why firms benefit when their complements become cheap. Christensen and Raynor described how profit opportunities migrate as parts of a value chain become commodities. Those ideas tell us where to look. But profits lost by one business do not have to become profits for another. I expect a substantial share of the gains to reach customers as lower prices, or make previously unpaid activities possible, without becoming durable corporate profit anywhere. The research question is which part of the invoice survives, and which new activity can earn one.

Substitute starts with the threatened invoice. On 3 February 2026, RELX shares fell about 14% during a selloff following Anthropic’s new legal tools. By July, RELX was reporting 10% underlying revenue growth in its Legal division for the first half, with adoption of its own AI-enabled platform supporting growth. The market feared that generating legal answers would destroy the business. Customers were still paying for a product combining AI with verified legal content and an established workflow. Those results do not settle the long-term threat. They show why replacing a task and replacing the business that makes its result usable are different propositions.

The opportunity can belong to a new provider taking over the function, or to an incumbent whose valuable position survives the destruction of its old product. RELX is a case to investigate, not proof that every information business is protected. If the customer can obtain the content, reliability and workflow elsewhere, the remaining invoice comes under pressure too.

Amplify begins when the customer uses cheaper intelligence to do more. Occasional analysis becomes continuous analysis; software agents carry out more transactions; robots enter work that could never justify a robot before. All that activity needs power, cooling, equipment, connectivity and ways to establish identity and permission. In the quarter ended August 2026, Oracle reported 121% growth in cloud-infrastructure revenue and $664 billion of remaining performance obligations. The first measures revenue already earned; the second is contracted revenue still to be recognised. Both explain why so much capital is chasing the physical means of supplying intelligence. Neither tells us what return the last unit of new capacity will earn.

Reprice concerns things cheap intelligence cannot copy. You may be able to summon a castle in virtual reality. You cannot summon another stretch of Mediterranean coastline. If AI makes people richer, more money can chase the places and experiences they still want in the physical world. Yet the virtual castle could also change what they want to pay for. And if work no longer forces them to live in an expensive city, they may leave. Some places will discover that their scarcity was partly captivity. Monetary assets involve another mechanism: how governments finance the transition and how much confidence survives in the money used to hold wealth. I explore that mechanism in the section on the sovereign balance sheet in Every Technology Boom Borrows From the Future.

Unlock goes beyond redistributing existing spending. Some demand never appears in today’s accounts because serving it has always cost too much. A useful service requires more expert attention than its customers can afford. A neglected problem never gets a team. Cheap intelligence changes which problems can receive sustained attention.

Consider a rare disease whose genetic cause remains unexplained. In a case described by Google DeepMind, researchers at the Broad Institute used AlphaGenome Atlas to prioritise a variant in the DNM1 gene. Its prediction of a faulty splice site supplied supporting evidence that helped resolve the case. The model had given researchers a promising place to look in an enormous field of possibilities. A prediction still needs investigation, and a diagnosis still needs a treatment. But making that search affordable can bring previously neglected problems within reach. The investment question is who can turn the new capability into a service people can use and pay for. DeepMind offers Atlas free for academic research and has announced commercial access through Google Cloud, so the business collecting the new fee may be the model owner itself.

These four movements happen together. The service replacing an old fee consumes infrastructure, releases spending and makes another activity possible. They take the opportunity far beyond the companies selling models. They also bring us to the next difficulty: identifying what could become valuable is much easier than knowing when it will become a business that pays.

A condition in the plan is not a condition delivered

Across my career in transformation consulting, only one project ran to plan: a bank launching an insurance company from scratch. It was a greenfield business, with no dependencies on existing systems, one owner rather than a collection of shareholders, and that owner deeply involved in the project.

In every other case, delays ranged from six months to three years. Their scale depended on how many of the client’s commitments and preconditions failed to materialise. These were assumptions we had discussed at the start, not obstacles nobody had thought to mention. The exceptional project worked because the conditions behind its timetable could actually be delivered.

That experience shapes where I look for the first beneficiaries. A new business with an involved founder can build around the capability from the beginning. An established company has to change systems, incentives and decisions that already belong to somebody. The bank’s new insurer shows that an incumbent can create the same freedom, but has to give it real authority. My inference is that the ability to act without renegotiating the whole organisation may matter more than the size of the AI budget. It gives Unlock a head start and makes the defence described in Substitute a transformation project of its own.

Now lend money to the infrastructure waiting for those transformations. The model can improve on schedule while the organisation that intends to use it is still trying to make its own plan true. Interest, meanwhile, arrives according to the contract.

Shareholders finance the wait too. In the quarter ended August 2026, Oracle sold $20 billion of new shares through its at-the-market programme. The buildout gained cash; each existing share represented a smaller fraction of the company.

Customers have made promises of their own. The latest filings from Microsoft, Alphabet, Amazon and Oracle, covering June and August 2026, together disclosed more than $2.3 trillion of remaining contracted revenue. That total spans software and cloud services, includes amounts already billed, and cannot be treated as AI debt. It shows how much future business has already been promised.

Some of those customers have committed to paying for capacity whether they use it or not. CoreWeave’s filings describe take-or-pay contracts that begin on a fixed date or when the specified capacity is delivered. Its large customers include AI labs and cloud providers. Part of the revenue they need to earn from this capacity must come from thousands of organisations like those I helped transform. A strong balance sheet can fund the wait, but cannot make those organisations adopt faster. For delivery-triggered contracts, the bill runs on the infrastructure timetable. The earnings needed to cover it run on the adoption timetable. In all but one of my transformations, that second timetable slipped by six months to three years. A contract can make the customer pay for an idle machine. It cannot make the machine profitable for the customer.

This is the AI capital cycle: Buildout → Digestion → Diffusion. During Buildout, expectations about tomorrow’s demand finance today’s capacity. Suppliers can earn before their customers have discovered how to earn from the equipment. The orders are real. What remains uncertain is whether the eventual revenue will reward all the capital committed along the way.

Digestion begins when those promises meet operating results. Adoption takes longer, new capacity weakens prices, or a better model changes the economics. Investment slows, contracts change and assets are repriced. In a gentle adjustment, demand grows into the capacity. In a harsher one, creditors become owners and the person who financed the future is removed from it. The technology need not fail for its original financing to fail.

Diffusion is the wider economy reorganising around capability that is dependable and affordable. A later business can rent capacity cheaply, acquire an asset after a restructuring or begin with a better generation of equipment. It benefits from work already done without paying the original bill. Its attractive economics may depend on somebody else’s disappointing return. These phases overlap: a component can remain scarce while its buyer’s finances deteriorate, and useful services can spread while another part of the industry digests excess capacity.

That is why being right about what the world will need is insufficient. The original shareholders may not survive to own it, or today’s price may already demand too much of tomorrow. I develop this conflict in Every Technology Boom Borrows From the Future. Here it supplies the time dimension of the investment: what becomes necessary, when it starts paying, and who still owns it by then.

When the map stops working

The argument so far describes the world I consider most plausible: Powered Diffusion, where intelligence keeps improving and spreading while the rest of the economy adapts unevenly. The four movements identify where to investigate; the capital cycle explains why the opportunity changes along the way. But an investor can understand both and still be wrong about the world they are entering.

In Plateau, models remain useful assistants but fail to become reliable autonomous workers. The demand financed into today’s capacity never fully arrives. In Discontinuity, a handful of laboratories achieve such a large advantage that they absorb the businesses investors expected to prosper around them. The supplier becomes its customer’s competitor, and the biggest winners may remain private. Knowing who won would not guarantee that an ordinary investor could buy a meaningful share of the victory.

Abundance Shock attacks the thesis through success. Better algorithms reduce computing needs, energy supply expands and robotics makes physical production much easier to scale. The world catches up with its new intelligence faster than expected. The machines have helped build their way out of the scarcity you bought. New constraints may emerge without lasting long enough, or remaining accessible enough, to reward public shareholders.

Politics can change the result even if the technology develops exactly as expected. A government losing payroll revenue must find another way to fund services and sustain household purchasing power. It can broaden ownership, tax profits to purchase provision, or make essential systems public utilities with capped returns. A humane, orderly settlement may leave citizens richer in things they can use and investors poorer in profits they can collect. The claim to the output matters as much as the output itself. That is the question I explore in Who Owns the Machines That Replace You?

The framework assumes systems remain sufficiently aligned with human purposes and under human control, governments hold together, contracts and property rights retain meaning, and some businesses earn profits their shareholders can receive. Redistribution and regulation fit inside those assumptions; the disappearance of investable private profits does not. A cleverer stock list cannot repair every possible future.

Owning the transition

When Intelligence Is Free turns this argument into research on public transition beneficiaries and a public investment book with positions, returns and reasoning. A business must matter in the economy emerging around it, retain a worthwhile share of what it creates, and be available at a price that leaves something for its owner. A company can satisfy the first condition and fail the other two.

RELX shows how I apply those three conditions. Reliable legal work remains necessary; the next question is whether customers still need RELX to obtain it. The Legal division’s growth gives me evidence of value retained, but I look for renewals and pricing that survive competing AI products. Then I ask what the share price requires me to believe. I value the cash the business could produce if legal answers keep getting cheaper, and compare that with the price of the whole group. A share-price fall creates an opportunity only if the business has lost less value than the market assumes. RELX is in my public investment book at publication; the book and its reviews are where I explain and update the case for owning it.

My monthly and quarterly reviews show how new evidence changes the investment case and the book. If competitors keep passing each productivity gain to customers, I stop treating that gain as something shareholders own. If capacity keeps arriving while customers postpone deployment, I put more weight on who can finance the wait. If supply expands faster than demand and pricing power disappears, I stop valuing yesterday’s shortage as a permanent asset. These are reasons to change an investment, even while the technology keeps getting better.

I expect the uneven world to offer factories to build, organisations to change, services to invent and problems worth solving. Some will make excellent investments. Others will improve life while disappointing their owners. An investor cannot make every holding a wish for some human problem to remain unsolved.

For most of history, ambition had to wait for enough capable people. We are beginning to remove that limit. The next constraint may be power, the means to act on an idea, permission to proceed or a claim on what the machines produce. Each has different owners and a different life expectancy. The task is to recognise when something useful becomes worth owning, and when the world has learned to do without the scarcity we bought.

When intelligence is free, I want to own what the world is trying to do next.

When Intelligence Is Free is independent research on where economic value moves as intelligence becomes abundant. Explore the framework, see how the thesis is expressed in the public investment book, or apply the same reasoning to a company or market decision through advisory.