Too Cheap

The Falling Price

In late May, an index of the price actually paid per token of artificial intelligence reached a peak of about $2.07 per million tokens. By early October it had fallen by 52%.¹

The index, compiled by Silicon Data, weights prices by usage rather than taking published price lists, so it captures what buyers really pay after discounts, caching and switching to cheaper models. Some of the fall is that switching: a change in the mix of models used, not only cuts in the price of any one of them. A fall of more than 50% in five months, in a product whose use is growing explosively, is either the best news in the history of computing or a problem for anyone who borrowed to produce it. It is both.

Parts I and II of this series dealt with the first two legs of the Catch. Part I, Too Slow, showed that the plans call for far more capacity than measured demand requires, built from scarce inputs on a critical path measured in years. Part II, Too Soon, followed the bill: costs that ratchet up with delay, three real projects that distribute the same shock very differently, and a credit market that has begun pricing the tenant rather than the building.

This final part asks the third question. Suppose everything works. The buildings are finished, the power arrives, the customers come. Who gets paid when intelligence keeps getting cheaper?

The answer matters because the capital and the commercial advantage have very different lives. The buildings last decades and the debt five years or more. The commercial edge of the chips inside them may last two or three. And the price of what they produce is falling faster than almost anything that has ever been sold.

Three Prices

There are three prices of intelligence, and only one of them settles the question.

The published price is what a provider advertises per million tokens. The realised price is what customers actually pay, after discounts, caching, routing to cheaper models and contract concessions. The capability-adjusted price is what it costs to complete a given task to a given standard of accuracy, speed and reliability. Only the third measures the economic price of intelligence, and no one yet discloses it consistently.

What can be measured points in one direction. At OpenRouter, an independent platform that routes requests across many models, open-weight models accounted for about 72% of tokens in mid-September and 27% of spending; closed models for 28% of tokens and 73% of spending.² These are one platform's measures, not global shares, and the token and spending figures come from slightly different weeks. On those figures, closed models earned roughly seven times as much per token as open ones. The comparison is loose, since the workloads differ, but it shows why market share measured in tokens misleads. Part I noted the same pattern at Vercel's gateway.

The frontier laboratories feel it too. According to Ramp data cited by Reuters Breakingviews, Anthropic's token consumption rose about 90% between early July and late September, while customer spending with it rose about 19%.³ If the two series are comparable, revenue per token fell by more than a third in a quarter.

And a growing share of the tokens is not new thought at all. By September, almost 90% of the tokens consumed by agentic systems on OpenRouter were cached prompts, text the model had already seen and was being fed again, with output and reasoning only about 1%.⁴ Token counts are growing fastest precisely where they say least about work done.

The Arithmetic of Abundance

Revenue is price multiplied by quantity. When the price of a unit of intelligence falls, volume must rise faster just to stand still.

The third line is the danger: volumes treble and revenue falls. The last line is the bull case's reply: a tenfold rise in usage can absorb a 90% price cut. Whether cheaper intelligence pays depends entirely on how much new demand each price cut releases, and on whether costs fall faster than prices.

This is Jevons's paradox again, from Part I. Cheaper steam engines raised Britain's coal consumption. Cheaper intelligence may well raise the world's use of computing by more than the price falls. But Jevons describes the coal burned, not the profits of the colliery. A paradox that raises volumes need not raise returns on the capital that produces them.

Call it two clocks. The price of a unit of intelligence moves in months; the equipment that produces it is ordered years ahead. Silicon Data's token-price index halved between late May and early October, though it is weighted by usage, so part of that fall reflects customers moving to cheaper models rather than any model getting cheaper. Over the same months, a large power transformer still took up to three years to arrive. The faster clock sets the revenue. The slower one sets the cost. The gap between them is the risk.

Deutsche Bank’s own figures capture the squeeze. The price of renting a GPU for a year has risen by about a third over the past year; the price of the intelligence it produces has been almost halving every quarter, by Epoch’s measure about thirteenfold a year.⁵ Scarce inputs and deflating outputs are a margin squeeze, and it falls on whoever is paying for the capacity.

Who Keeps the Surplus

Economic value is not revenue. That sounds obvious, and it is the most frequently forgotten point in the AI debate.

Consider a task that once cost $100 of human labour. Suppose an AI system performs it, to the same standard, for a total cost to the customer of $15: $8 to the software provider and $7 for integration and checking. The customer saves $85. Of the $8, only a part reaches the developer of the underlying model, and only a part of that pays for the data centre.⁶ The largest immediate gain goes to the customer, even though the suppliers invested the capital. That is how technological competition usually works.

The evidence so far is consistent with that pattern. Researchers at the Stanford Digital Economy Lab, surveying American adults on what they would need to be paid to give up AI chatbots for a month, estimate that consumer surplus from generative AI rose from about $116 billion a year in July 2025 to $172 billion by March 2026, and their tracker puts it at about $183 billion by July 2026.⁷ That is the value users place on access, well above what they pay for it. The authors' own conclusion is the one that matters here: the surplus substantially exceeds generative AI's revenue in the United States, so consumers are capturing most of the gains.⁷ In business, the picture is uneven. A study of almost 6,000 executives found 69% of firms using AI, yet about nine in ten reported no effect on employment or productivity over the previous three years.⁸ In McKinsey's latest global survey, published in August, only 37% of respondents attributed any of their earnings before interest and tax to AI, about the same share as a year earlier.⁹ Where the gains are real they can be large; customer-service agents resolved about 14% to 15% more issues an hour.¹⁰ And customers are not loyal. Ramp reports that the share of firms switching AI provider each month, averaged over three months, reached a record 8% in September.¹¹

GENAI_CSGENAI_CS · Jul 2025 to Jul 2026 · Stanford Digital Economy Lab

The Customers Keep the Surplus

Estimated annual US consumer surplus from generative AI, $ billion. Jul 2025 to Jul 2026.

Estimated

In August, Dario Amodei's case for the size of the prize rested on AI capturing part of the world's labour income, an addressable market Anthropic has put above $30 trillion a year. The question that number begs is how much of it reaches the suppliers.

Stijn Van Nieuwerburgh calculates that the US data-centre capacity built by 2032 in his central case would need about $3.7 trillion of revenue a year to earn a 10% return.¹² Set that against the table. It requires the suppliers to capture well over 10% of the entire $30 trillion pool, every year, in a market where prices are halving and customers are switching. Gartner reaches a similar place from another direction: it estimates the industry needs $8.5 trillion of revenue through 2030 to earn a 12% return on AI infrastructure, against $3.6 trillion it expects.¹³

The Macroeconomic Paradox

There is a further twist, which the August letter raised. A company that replaces expensive labour with cheaper intelligence improves its own margins. If enough companies do so at once, wages and employment in the affected occupations can weaken, and with them the consumer spending that ultimately pays for everything else. The gains flow first to shareholders and to customers through lower prices. But a dollar of extra profit for a wealthy shareholder is spent less readily than a dollar of wages for a lower-income household.

The evidence is early and it cuts both ways. Researchers at Stanford find employment among 22- to 25-year-olds in the occupations most exposed to AI about 19% below where it would be had it kept pace with similar workers in less-exposed jobs, with no comparable gap among experienced workers, though without establishing that AI alone caused it.¹⁴ The executives surveyed for the National Bureau of Economic Research expect AI to raise productivity over the next three years and to cut employment by about 0.7%.⁸

For credit, the point is narrower and sharper. Falling prices for intelligence are good for the economy. But debts, leases and guaranteed payments were contracted at yesterday's prices, and they do not fall with them. If technological success lowers the nominal revenue of the businesses that financed it, while their obligations stay fixed, success itself becomes a source of strain. Irving Fisher would have recognised the mechanism. And it is one more reason the authorities, as Part II argued, have every incentive to keep nominal growth, and long-term borrowing costs, under control.

The Reliability Question

There is one more reason the timetable matters. Much of the forecast demand rests on agents: systems that carry out long sequences of tasks on their own. Reliability compounds. A system that gets each step right 95% of the time completes a twenty-step task without error only about 36% of the time. At 99% per step, it is about 82%.⁶ Errors can be caught and retried, and humans make mistakes too, but the arithmetic explains why impressive performance on short tasks does not translate directly into dependable autonomy.

Yann LeCun has long argued that today's language models are not, by themselves, the route to human-level intelligence; Andrew Ng has argued for judging AI by the practical work it does. Neither is a counsel of despair. Both are questions about timing: how soon useful, reliable autonomy arrives, and therefore how soon the demand the buildings were financed for becomes paying demand.

The Laboratories

None of this means the model developers lack customers. Anthropic reportedly told investors it earned $16.8 billion of revenue in the third quarter, and an adjusted operating margin of about 11% for the year to date, a measure that excludes share-based compensation among other costs.¹⁵ These are private, unaudited figures from an investor briefing. Even so, they are the economics of a business with genuine demand, and they are improving.

But the growth is slowing as it grows; third-quarter revenue rose about 45% on the second quarter, after about 147% the quarter before.¹⁵ And the commitments run far ahead of the revenue. Anthropic's draft prospectus, as Reuters reported it, showed $518 billion of compute commitments, about 80% of them non-cancellable, against $4.6 billion of revenue in 2025.¹⁶ Part II called that the Catch-22 in one document.

ANTH_REVANTH_REV · Q1 to Q3 2026 · Reuters (14 Oct 2026)

Anthropic: Growing Fast, Slowing as It Grows

Anthropic reported quarterly revenue, $ billion. Q1 to Q3 2026.

Reported (private)

OpenAI's numbers show the same pattern from the other side. On its own projections, as the FT reported them in September, revenue rises from $36 billion this year to $350 billion in 2030, about $840 billion in all, while spending on computing and infrastructure reaches about $856 billion. Free cash flow is negative throughout, by $278 billion over the five years, and the $122 billion it raised in March is on track to run out by 2028.¹⁷ The spending is largely contracted; the revenue is the part that has to arrive. A revenue miss does not shrink the bill. It widens the gap that someone else must finance.

Two cautions make the comparison harder than the headlines suggest. The first is that the two companies count differently. Anthropic books the full value of sales made through cloud partners such as Amazon and Google, and records the partners' share as a cost; OpenAI books only its own share of comparable sales. The FT put OpenAI's September run-rate at about $50 billion on its own basis, against the roughly $70 billion that had been circulating in the market, a figure closer to Anthropic's way of counting. The $20 billion difference is largely the partners' share, not customers that disappeared.¹⁸ On a like-for-like basis, Anthropic's run-rate, which SemiAnalysis estimates at about $74 billion, and OpenAI's roughly $70 billion are close to level, not the 50% lead the headlines imply. Neither figure shows what each company keeps once its partners are paid, and that is the money that has to service the commitments. Anthropic's $74 billion is also below the $100 billion-plus exit rate some investors had expected.¹⁸

The second is that the forecasts have risen with the commitments. Anthropic's 2028 revenue forecast was reported at $70 billion in November 2025 and at $190 to $200 billion in August 2026. OpenAI's 2030 forecast has doubled in fifteen months, from $174 billion to $350 billion, while its projected compute bill rose from about $600 billion in February to $856 billion.¹⁹ When the forecast justifies the commitment and the commitment requires the forecast, the two move together. Soros would recognise the loop.

LAB_MATHLAB_MATH · Run-rates Sep 2026; forecast vintages Apr 2025 to Aug 2026 · FT; Reuters; The Information; SemiAnalysis

Bigger Forecasts, Bigger Bills

Anthropic and OpenAI: revenue counted the same way, and how the forecasts and the compute bill have grown. $ billion.

Reported; estimates and projections

A · Counted the same way

B · Each forecast bigger than the last

OpenAI's own plan, 2026–2030

Revenue

~$840bn

Compute and infrastructure

~$856bn

Free cash flow

−$278bn

Spending is largely contracted; revenue has to arrive.

The infrastructure's own economics give a sharper test. Van Nieuwerburgh estimates that his capacity needs to earn about $5.50 per GPU-hour at full use, or $6.90 at 80% utilisation, to make its return.¹² Silicon Data's standardised rental index for Nvidia's B300 stood near $6.96 per GPU-hour on 16 October, though the two figures are not built the same way: one is a modelled hurdle for a whole facility, the other a market rental benchmark.²⁰ If those measures are comparable, today's rental prices sit almost exactly at the hurdle. The capacity earns its return only if prices hold where they are, while the chips that set those prices are superseded every year and the price of what they produce is falling.

Gartner has done the arithmetic for the whole industry. It expects about $7 trillion of AI infrastructure investment between 2026 and 2030, roughly $1 trillion this year and $6 trillion more after it, which is why its numbers dwarf Morgan Stanley's $2.9 trillion to 2028. To cover that spending, its financing costs and a 12% return on capital, the owners would need about $8.5 trillion of revenue by 2030. Gartner forecasts no more than $3.6 trillion of revenue attributable to AI, a shortfall of nearly $5 trillion.¹³ Its conclusion follows from the gap. Assets that cannot earn their cost of capital are written down, so more than $3.5 trillion of the investment, about half of it, is at risk of impairment by 2030, rising above $4 trillion by 2032.¹³ And more than 80% of neoclouds, it expects, will cease operating independently by 2030: not necessarily by failing, but by being bought, merged or absorbed by the companies that can carry the loss.¹³

Corporations, Not Laboratories

The industry calls them laboratories, and the word does a great deal of quiet work. A laboratory pursues knowledge. A corporation pursues profit, owes duties to its shareholders, signs contracts, carries debt and answers for what it sells. OpenAI completed its restructuring into a public benefit corporation a year ago. Anthropic is one already. One has filed a draft prospectus and the other is reported to be preparing to.^{16 21} They are corporations, raising capital at record valuations, and they are not absolved of the responsibilities and liabilities that come with that.

Which makes their behaviour towards government all the more curious.

The Backstop That Was Not Asked For

In November last year, OpenAI's chief financial officer, Sarah Friar, was asked at a Wall Street Journal event how the company would finance its commitments. She described an ecosystem of banks and private equity and a role for government. Asked whether she meant a federal backstop for chip investment, she replied: "Exactly."²² Within hours she wrote that OpenAI was not seeking one and that the word had "muddied the point". Sam Altman followed: OpenAI does "not have or want government guarantees" for its data centres.²² David Sacks, then advising the White House on AI, was blunter: "There will be no federal bailout for AI." If one frontier company fails, he added, others will take its place.²² Senator Elizabeth Warren, the senior Democrat on the Banking Committee, nonetheless wrote to the White House asking what plans existed for loan guarantees, tax credits or other support, noting that OpenAI had also floated guarantees under the Defense Production Act.²³

A request withdrawn within hours is still a request that was considered. Markets have a long memory for implicit guarantees, because they have been expensive.

What Implicit Guarantees Have Looked Like

Fannie Mae and Freddie Mac were private, shareholder-owned companies for decades, yet they borrowed as though the government stood behind them, because investors assumed it did. The assumption let them borrow more cheaply and more heavily than any private competitor. In September 2008 it was proved right; both were placed into conservatorship and Treasury committed capital to keep them solvent, close to $190 billion of it in the end. The guarantee had never been written down. It did not need to be.

The internet companies and telecoms of the 1990s received something different. They were not bailed out; when the telecom bubble burst in 2001 and 2002, WorldCom and Global Crossing went bankrupt, bondholders lost heavily and fibre changed hands for cents on the dollar. What they received instead was leniency, written into law. Section 230 of the Communications Decency Act of 1996 shielded online platforms from liability for what their users posted. The Internet Tax Freedom Act of 1998 barred taxes on internet access, a moratorium later made permanent. The Federal Communications Commission classified broadband as a lightly regulated "information service" rather than a common carrier, a decision the Supreme Court upheld in 2005. And the Universal Service Fund, later joined by large broadband subsidy programmes, paid to extend networks where the economics alone would not.

None of that was a bailout. All of it lowered the cost and the legal risk of building an industry.

The AI industry is already receiving some of the same treatment: a 9.9% government equity stake in Intel, federal land offered for data centres, tax credits for chip manufacturing, and a federal push to override state AI laws in favour of a single national framework.²⁴ A single federal framework, in particular, can work as a liability shield as much as a constraint. That is worth remembering when a corporation asks to be regulated.

Super Intelligence, by Decree

The puzzle deepened at the end of September. On 29 September the President signed an executive order instructing federal agencies to stop calling the technology artificial intelligence and to call it "Super Intelligence" instead.²⁵ The same day, the chief executives of Google, Anthropic, Meta, xAI and Nvidia, and OpenAI's president, signed a "White House Accord on Super Intelligence", a voluntary "Joint Commitment on Frontier Responsibilities" covering safety controls, independent audits and board-level oversight, which the President described as "morally binding". The accord itself adds: "Over time, it may make sense to codify these steps into laws or regulations."²⁵

A technology renamed by decree, and a voluntary code its leading producers signed with the government that may one day bind everyone else: whatever else that is, it is not a market left to price itself. It is also, notably, a code whose signatories "will meet regularly to establish standards and best practices", standards their smaller rivals may one day be required to meet.

Why Ask?

The four investors who host the All-In podcast, Chamath Palihapitiya, Jason Calacanis, David Sacks and David Friedberg, have argued about this all year with more wit than most, and they are closer to it than nearly anyone. Two of them sit on the President's Council of Advisors on Science and Technology. Their answer is mostly regulatory capture. Sacks has accused Anthropic of running "a sophisticated regulatory capture strategy based on fear-mongering" and of being "principally responsible for the state regulatory frenzy that is damaging the startup ecosystem".²⁶ Friedberg has made the subtler point that the technology is "already everywhere in the world", so the United States "cannot simply step back and pause". The implication cuts both ways: the earlier the rules are fixed, the more they favour whoever is already established. By late September the panel was discussing whether Anthropic's listing was at risk, from its safety messaging, its dependence on a few large customers and its lobbying.²⁶

I confess I am not sure capture explains it all. It is a strange moment to warn the world about the dangers of your product when you are on the cusp of selling shares in it at a record price. Several explanations are possible, and they are not mutually exclusive.

Conviction. Some of the people involved believe what they say about the risks, and have said so for years, long before there was anything to sell.

A moat. Regulation that requires licences, testing or approval raises the cost of entry. It falls hardest on open-weight models and smaller rivals, and on foreign competitors whose models, Intel's chief executive conceded in Singapore this month, are now "good enough".²⁷ Both laboratories have accused rivals of illicitly distilling their models at industrial scale.²⁸

A shield. A corporation facing open-ended liability for what its product does has every reason to prefer a single federal standard, written with its help, to fifty state laws and the courts. The internet got Section 230. The question is what the AI industry hopes to get.

A cheaper cost of capital. An implicit guarantee, however denied, lowers the cost of financing commitments that run to hundreds of billions of dollars. Friar said so in terms before withdrawing it.

The S-curve. This is the explanation that fits this series best, and it is my own suspicion. If the price of intelligence is falling by half in five months, if open-weight models carry most of the tokens, and if customers switch providers at record rates, then the frontier corporations' cost structure is under threat from the third leg of the Catch: too cheap. A business that fears commoditisation has a strong incentive to seek rules that slow it. The panel itself argued in September that converging open-weight models are eroding the frontier companies' pricing power.²⁶

Which of these is right? Those closest to it do not seem to know, and I remain somewhat baffled. The evidence that would discriminate between them is observable: whether proposed rules fall hardest on open-weight and foreign models, whether liability safe harbours appear in the federal framework, whether financing terms improve on any hint of state support, and whether the companies' pricing power holds up as they list.

What is clear is that a technology of enormous promise has acquired party lines. One side hears safety and suspects capture; the other hears acceleration and suspects recklessness. That is a pity, because the benefits are real and, in places, extraordinary. The 2024 Nobel Prize in Chemistry went in part to the AlphaFold team for predicting the structure of proteins, and AI is already reshaping drug discovery. My own view is that the biosciences and medicine may be where this technology does the most good, and soonest. None of that depends on any one corporation's valuation, and none of it requires taxpayers to stand behind its debts.

Paid on Delivery

The semiconductor companies are paid to provide the equipment. The infrastructure owners must make the equipment earn its keep. Those are different businesses.

That distinction runs through the most profitable part of the chain. Nvidia reported quarterly revenue of $96.2 billion in August, up 106% on a year earlier, with a gross margin of about 75%. TSMC, the memory makers and Broadcom have reported exceptional results. The profits are real, and anyone who has dismissed the boom as fictitious has been wrong about them.

But a supplier recognises revenue when it delivers. The buyer's return is proved, or not, over the following years. The supplier's earnings can be entirely genuine while the return on the capital that paid for them remains uncertain. The question is not whether the chips were delivered. It is whether the structure that financed them can earn an acceptable return after operating costs, depreciation, interest and replacement.

The connections make the question harder. Nvidia supplies CoreWeave's chips, has invested in it, and has committed to buy up to $6.3 billion of its unsold capacity through 2032.²⁹ Broadcom guarantees financing vehicles that buy chips for Anthropic, including one of $60 billion now in the market, and could fund Anthropic directly, through up to $42 billion of convertible notes if needed, while backstopping its chip leases up to $29 billion.³⁰ At the end of July, Nvidia disclosed $366 billion of future commitments, $279 billion of them for supply and capacity, up from $119 billion a quarter earlier, and in August it agreed to provide conditional credit support, capped at $105 billion, for OpenAI's 20-year leases at SB Energy's 4.25-gigawatt campus in Ohio, taking effect as the leases begin, the first expected in its 2029 fiscal year.³¹ None of this is improper. But when a supplier supports its customers' demand, the supplier's earnings become partly dependent on the financing of companies whose economics it is supporting. Hock Tan, whose company is among the largest beneficiaries, called many AI investors gamblers in a casino this month, and capital the industry's biggest constraint.³²

NVDA_COMMITNVDA_COMMIT · 26 Jul 2026 · Nvidia Q2 FY2027 CFO commentary

Nvidia Beyond the Factory Gate: $366 Billion Committed

Nvidia future commitments by category, $ billion, at 26 Jul 2026.

Reported

The Inconvenient Fact

The strongest evidence against the bear case belongs here. Posted prices for renting Nvidia's newest accelerators rose sharply this year; on-demand rates for the B300 roughly doubled in about six months, to around $9 an hour, and one large neocloud listed $9.50 on 1 October.²⁰ That is hard to reconcile with a glut.

CoreWeave's own September investor presentation adds to that side of the ledger; management said prices across its products were about 70% higher in July, and that contracts signed in the second quarter carried expected contribution margins 5 to 10 percentage points higher, with strong demand even for older chips.³³ That is genuine contrary evidence, though it is management's account of signed contracts, not proof of a return after financing, depreciation and replacement.

And a posted price is not a realised average, and one provider is not an index. The standardised B300 index stood at $6.96 per GPU-hour on 16 October and barely moved over the week. And unit economics turn on utilisation as much as price. A GPU rented at $10 an hour and billed 85% of the time earns about $74,000 a year; at $7 and 70%, about $43,000.²⁰ A 30% fall in the posted price becomes a 42% fall in revenue. When capacity is scarce, owners enjoy high prices and high utilisation together. When new supply arrives, both can fall together. That asymmetry is why a leveraged owner of GPUs is so much more exposed than the company that sold them.

Memory: Real Shortage, or a Place in the Queue?

Memory is the clearest test, and it raises a question that applies to the whole chain. The shortage of high-bandwidth memory is real, and its makers are earning extraordinary margins. Memory prices rose more than five times in a year.³⁴ But how much of the demand behind that is consumption, and how much is buyers ordering everything they can now to secure a place in the queue?

The buyers are more numerous than the headlines suggest. The hyperscalers (Microsoft, Google, Amazon, Meta and Oracle) order for themselves. The laboratories (OpenAI, Anthropic, xAI) commit through them and increasingly directly. The neoclouds (CoreWeave, Nebius, Lambda, Nscale and others) buy against contracts. Behind them stand the server makers that assemble the racks and hold inventory in between, sovereign buyers in the Gulf and Asia building national capacity, enterprises building their own, former bitcoin miners turned AI hosts, and Chinese buyers stockpiling ahead of export controls. Not every one sees the same lead times; suppliers serve their largest, longest-standing customers first, which Cushman's cost guide documents for power equipment.³⁵ So every one has a reason to order early, and the smaller ones a reason to order more than they need.

That is the classic recipe for what supply-chain engineers call the bullwhip effect, first described by Jay Forrester at MIT in the 1950s and 1960s.³⁶ When lead times lengthen and supply is rationed, each buyer orders more than it needs, because it expects to be allocated only part of what it asks for. Each tier up the chain sees a larger and more volatile order book than the final demand beneath it. The semiconductor industry has lived through this recently. During the 2021 chip shortage, customers double-ordered across suppliers; when supply caught up, orders were cancelled, inventories swelled and memory makers swung to heavy losses in 2023.

Morgan Stanley’s semiconductor analyst sees the same risk: in a pause, the most disruption would come “at the tail of the bullwhip”, in memory, optics and other small components most likely to be over-ordered, and if spending slows, “there will be no safe havens in the sector”.³⁷

My own working assumption, from conversations and from the order books, is that buyers are front-running demand and overstocking chips. That is a judgement, and it should be tested rather than asserted. The evidence that would distinguish real consumption from queue-jumping is observable:

• Chips waiting for power. If accelerators are being delivered faster than they can be energised, some of the order book is ahead of the buildings. Microsoft's chief executive, Satya Nadella, said late last year that the company may have "a bunch of chips sitting in inventory that I can't plug in". The problem, he said, was not a supply of chips but a lack of "warm shells to plug into".³⁸ Jupiter's finished, unpowered halls in New Mexico make the same point physically.
• Inventory in the middle of the chain. Rising inventories and receivables at server makers and chip suppliers, relative to sales, would show stock piling up between the factory and the data centre.
• Prepayments and long-term agreements. Buyers paying in advance or signing multi-year supply contracts to secure allocation are paying for a place in the queue. That is rational, and it is also exactly how double-ordering looks.
• Utilisation and rental prices. If installed capacity were scarce everywhere, rental prices for existing chips would be rising broadly. The standardised B300 index was flat in mid-October, even as some providers raised posted prices.
• Cancellations and deferrals. The first sign of a bullwhip turning is not a cancelled order but a deferred one: a campus postponed, a delivery rescheduled, an option not exercised.

The strong counter deserves its weight. AI memory is more specialised and more contractual than in past cycles, yields are hard to raise, the makers may be more disciplined about adding capacity, and falling prices could unlock new demand in devices and local inference. Prices could normalise while volumes keep growing.

But the history of every component shortage is the same. Scarcity looks permanent while it lasts, because it makes buyers order more, which makes it look worse. A shortage can be real and a peak can still arrive.

The Stack

It is worth going through the chain layer by layer, because the rents, and the risks, are not evenly spread. The useful questions for each layer are the same: what does it sell, where does its pricing power come from, how far are today's earnings above a normal level, and what would hurt it?

Accelerators

Nvidia remains the centre of gravity. In the quarter to late July it reported revenue of $96.2 billion, up 106% on a year earlier, of which $89.0 billion came from data centres, at a gross margin of 75%. It guided to about $108 billion for the following quarter.³⁹ On the day after those results its market value rose by more than $400 billion, to about $5.5 trillion.⁴⁰

Two details in the results matter more than the totals. The first is who is buying. Nvidia now splits its data-centre revenue between the hyperscalers, about $48.7 billion in the quarter, and a category it calls AI clouds, industrial and enterprise, about $40.3 billion and growing faster, up 138% on a year earlier.⁴⁰ That second group includes the neoclouds, the sovereign buyers and the enterprises: on average, less creditworthy than the hyperscalers, and more dependent on borrowed money. The fastest-growing part of the most profitable company's revenue comes from its least creditworthy customers.

The second is the margin. Management guided gross margin down to about 74%, and indicated it would bottom at 71% to 72% before recovering, partly because memory has become so expensive.⁴⁰ When a monopolist's margin is squeezed by its own suppliers, the rent is moving up the chain.

Around Nvidia, the competition is redistributing rather than reducing the rent. AMD sells a competing range of accelerators. The hyperscalers increasingly design their own chips (Google's tensor processing units, Amazon's Trainium) with partners such as Broadcom and Marvell. Broadcom's AI semiconductor revenue more than tripled in a year, to $16.7 billion in its latest quarter.⁴¹ Custom chips may take share from Nvidia without reducing demand for the foundry that makes them all.

Foundry

That foundry is, overwhelmingly, TSMC. In the third quarter it reported revenue of $46.7 billion, up 41% in dollars, with a gross margin of 66.4% and an operating margin of 58.7%. High-performance computing, the chips behind AI data centres, made up 68% of revenue. It guided to $52.3 billion to $54.3 billion for the fourth quarter and raised its capital spending for 2026 to between $60 billion and $64 billion.⁴²

TSM_NODESTSM_NODES · Q3 2026 · TSMC 3Q26 Management Report, p.2

TSMC: 79% on the Leading Edge

TSMC wafer revenue by process node, % of total. Q3 2026.

Reported

TSMC's position is the strongest in the chain, because it is paid by every designer, Nvidia, AMD and the custom-chip makers alike, and because no one else can yet make the most advanced chips at its scale. Samsung and Intel, now part-owned by the US government, are the alternatives, a long way behind. TSMC's risk is not competition but concentration; its fortunes now rest heavily on one end market, and its own capital spending is a bet that the AI orders keep coming for the years it takes a new fab to pay back.

Memory

Memory is where the cycle shows most clearly, and where the earlier question about queues applies most directly. Micron's revenue in the quarter to early September was $54.2 billion, almost five times a year earlier, at a gross margin of 87%.⁴³ For a memory company, that is without precedent. It has sold most of its high-bandwidth memory for 2027, at significantly higher prices, and guided to about $61.5 billion for the next quarter.⁴³ SK Hynix and Samsung are riding the same wave.

Analysts themselves do not expect it to last. The consensus forecasts for Micron see its gross margin falling from the mid-80s towards about 65% by the end of the decade.⁴⁴ Even that would be extraordinary by the standards of the industry's history: in its 2023 fiscal year, after the last shortage turned to glut, Micron lost money. The share price has already noticed, closing about 12% below its late-June high on the eve of the latest results.⁴⁴

Equipment

The machines that make the chips sit one step further back. ASML, the sole supplier of extreme ultraviolet lithography, said this month that some customers were already expressing interest in placing orders for 2029 and even 2030.⁴⁵ Applied Materials, Lam Research, KLA and Tokyo Electron supply most of the other tools. This layer is the strongest evidence for the bull case: customers who expect demand to collapse do not ask to order for 2030. They are also exactly what a bullwhip looks like from the top of the chain, where every buyer below has an incentive to secure a place in the queue.

Networking, Optics and Packaging

Tens of thousands of accelerators must be wired together, and the networking layer has grown with them: Broadcom and Marvell in switching chips, Arista in switches, Coherent and Lumentum in the optical components that carry data between racks. Advanced packaging, chiefly TSMC's CoWoS, and the specialised substrates beneath it have been bottlenecks in their own right. These businesses share the accelerators' fortunes with a lag, and with less pricing power.

Servers

The server makers, Supermicro, Dell, Hewlett Packard Enterprise and the Taiwanese manufacturers Foxconn, Quanta and Wistron, assemble the chips into racks. They have the thinnest margins in the chain and carry its inventory; they buy components early and sell systems later. If the bullwhip turns, the stock left in the middle of the chain is where it shows first.

Power and Cooling

The final layer is the equipment that keeps the racks running: Vertiv, Schneider Electric and Eaton in power distribution and cooling, and GE Vernova, Siemens Energy and Mitsubishi in generation. Their order books are long, their lead times longer, and their customers include utilities as well as data centres, which gives them a cushion the chipmakers lack. As Part I showed, they are also part of the constraint.

STACK_MARGINSSTACK_MARGINS · Latest reported quarter · Company filings (Micron 8-K; Nvidia results; TSMC management report; CoreWeave 10-Q)

Paid on Delivery: Margins Down the Chain

GAAP operating margin, latest reported quarter, %, from the chip suppliers to the owner-operator.

Reported

Read down the margin column and a pattern appears. The highest margins in the chain sit furthest from the final customer's return: memory and accelerators earn the most, the server assemblers and the leveraged owners of the finished capacity the least. That is exactly what one would expect in a shortage financed by credit. The suppliers are paid on delivery, at prices set by scarcity. The owners must earn those prices back over years, at prices set by competition.

None of this says the suppliers' profits will collapse. It says they are earned at the peak of the order cycle, and that the market should ask what they look like when the cycle turns, rather than multiply the peak by a multiple. Which brings us to the earnings themselves.

What the Earnings Include

Before turning to valuations, it is worth repeating a point made in August, in A Credit Cycle Wearing AI's Costume, because it bears directly on how much the market should pay for these profits. Accounting conventions can inflate reported earnings in a boom like this one without anyone breaking a rule.

Depreciation. A chip that earns most of its economic value in three years but is written off over six makes reported profits look better than the sustainable return. The hyperscalers spent the early 2020s lengthening the assumed lives of their servers; Amazon extended its own from five years to six in 2024, then cut it back to five in 2025, citing the pace of AI hardware development.⁴⁶ CoreWeave depreciates its chips over six years. Nvidia now launches a new data-centre platform roughly every year. And the short-lived assets are the bulk of the spending; Microsoft said about two-thirds of its $41 billion of capital expenditure in its June quarter went on shorter-lived assets, principally chips.⁴⁷ If the assumed life is too long, today's profits are borrowing from tomorrow's write-downs.

The arithmetic is simple, and large. A gigawatt of accelerators costs about $20 billion. Written off over six years, that is about $3.3 billion of depreciation a year; over four years, $5 billion; over three, about $6.7 billion. Every year added to the assumed life lifts reported profit. Every year it proves too long comes back later as a write-down, booked at once. It helps to distinguish four lives for the same chip: the accounting life management chooses, the technological life set by the next generation, the commercial life set by what customers will pay, and the financing life set by the loan. They need not agree. CoreWeave's latest chip-backed loan, arranged by JPMorgan and MUFG, runs about five years against customer contracts averaging about three.⁴⁸ When the financing life outruns the commercial life, the lenders hold the chips and CoreWeave's guarantee; the shortfall, if there is one, falls on its unsecured bondholders and shareholders.

Interest that never reaches the income statement. While a data centre is under construction, the interest on the money borrowed to build it is generally capitalised into the cost of the asset rather than charged against profit. A project delayed for a year therefore carries a year of extra interest on its balance sheet, not in its earnings, until depreciation releases it slowly over the asset's life. Delay, in other words, does not show up in reported profits when it happens.

Commitments that are not on the balance sheet. Moody's estimates the hyperscalers carry about $970 billion of lease commitments, roughly $660 billion of them off the balance sheet.⁴⁹ Leases on buildings not yet delivered, and purchase commitments for chips and capacity, are disclosed in the notes rather than as liabilities. The obligations are real; the balance sheet does not show them.

Backlog that is the seller's view. Remaining performance obligations, the contracted revenue a supplier has yet to recognise, are disclosed by the seller, not the buyer, and the same demand can appear in several companies' backlogs at once. A large backlog demonstrates demand for a service. It also implies an obligation to finance the equipment to deliver it.

Customer credit and supplier support. In August, Nvidia's accounts receivable had risen from about $38.5 billion in January to $63.1 billion by July, with five customers accounting for about 70% of the total and the company noting extended payment terms on large agreements with certain investment-grade customers.⁴⁶ Its customers are creditworthy and nobody suggests a bad-debt problem. But Nvidia is increasingly also an investor in its customers, a buyer of their capacity and a guarantor of financing for its own products. It records the same revenue whether a chip is paid for out of a profitable customer's cash or by a customer financed against a lease on a building not yet powered. The economic quality of those two sales is not the same.

None of this is fraud, and none of it is new to anyone who reads the notes. The point is cumulative. In a capital-spending boom, each convention leans the same way: costs are deferred, obligations are disclosed rather than recognised, and the supplier's revenue arrives before the buyer's return. The numerator of the price-to-earnings ratio may be helping to inflate the denominator. That is why a low multiple on these earnings may not be what it seems.

Both Sides of the Valuation

A share price is earnings per share multiplied by a valuation multiple. The two are usually discussed as if they were independent; earnings reflect the business, the multiple reflects sentiment and interest rates.

In a reflexive investment cycle they are not independent. High valuations make capital easy to raise. Easy capital funds infrastructure. Infrastructure spending becomes the suppliers' revenue and earnings. The earnings validate the valuations. Both sides of the equation can be inflated by the same credit expansion. That does not make the earnings fictitious. It makes them cyclical in a way most forecasts of technological demand do not capture.

Both sides of the price-to-earnings equation may now be reflexive.

A share can fall while its earnings rise: earnings up 20% and a multiple down 25% still lose 10%. Reverse the earnings cycle and the losses compound. That is why a semiconductor share that looks cheap on current earnings may not be cheap on normal earnings. A low multiple on unusually high earnings is not necessarily a bargain. It may simply be the market questioning the denominator before the analysts have.

The transmission from credit to equity follows a sequence that is now worth watching: credit reprices; marginal financing shrinks; fewer new deployments are ordered; supplier order growth slows; earnings estimates are revised; valuations fall further. No default is needed at any step. A neocloud whose bonds trade at a 15% yield need not cancel the chips already ordered. It may simply postpone its next campus, and a slice of next year's orders quietly disappears.

Positioning

There is one more force on the multiple that has nothing to do with earnings: who owns the shares, and with what money. The most widely owned investment becomes the marginal source of funds when anything goes wrong elsewhere.

South Korea has just given a vivid demonstration. Samsung Electronics and SK Hynix together make up more than half the value of the Kospi, and retail investors had borrowed heavily to own them. Margin loans reached a record of about 38 trillion won in June, as the index passed 9,000, and the margin debt tied to SK Hynix had risen almost fivefold in six months.⁵⁰ Leveraged products built on the two chipmakers added to the exposure.

In seven sessions from 6 October, the Kospi fell about 21%. Samsung fell 23.9% and SK Hynix 31.5%. On 15 October the index dropped 8% and trading was halted, the tenth market-wide circuit breaker of the year, and brokers' forced sales of investors' shares exceeded 80 billion won for a second day.⁵⁰ Korean analysts called the fall excessive, driven by a lack of buyers rather than by any deterioration in the companies' earnings, which remain at records.⁵⁰

In America, Micron's shares closed about 12% below their late-June high the day before it reported the best quarter in its history.⁴⁴ The semiconductor funds most investors use to own the sector, the VanEck Semiconductor ETF (SMH) and the iShares Semiconductor ETF (SOXX), had risen by roughly two-thirds and nine-tenths respectively in the first half of the year, concentrated in a handful of names led by Nvidia, TSMC, Broadcom and Micron.⁵¹ Their direct exposure to Korea is small: SK Hynix made up about 3.8% of SMH on 15 October.

The index data tell the same story more precisely. From their closing peaks on 6 October to 16 October, the Philadelphia Semiconductor Index (SOX) fell 15.7%, the Nasdaq 100 6.5% and the S&P 500 2.0%.⁵² The semiconductors had in fact stopped leading months earlier; the ratio of the SOX to the Nasdaq 100 peaked in June, and the October rally made a lower high before the break. The broad market has barely moved. The part of it closest to the spending has repriced first. The same day, 6 October, the Nasdaq Composite left a one-day island at its high, a pattern the separate note takes up.

SOX_NDX_SPX_DSOX · NDX · SPX · Daily, 1 Jul to 16 Oct 2026 · FRED (Nasdaq; S&P Dow Jones Indices); price index closes

Semiconductors Break From the Index

Philadelphia Semiconductor, Nasdaq 100 and S&P 500 daily closes, rebased to 100 at 6 Oct 2026. 1 Jul to 16 Oct 2026.

Reported

This is the first row of the table above in action: earnings rising, multiples falling, because the marginal holder was leveraged and had to sell. It does not prove the earnings will fall. It shows how quickly the price of strong earnings can change when everyone owns them with borrowed money.

The broader market is not yet pricing that. FactSet put the forward price-to-earnings ratio of the S&P 500 at 19.0 times on 16 October, below its five-year average of 19.8, with blended third-quarter earnings growth of 29%, combining the few results then reported with forecasts for the rest.⁵³ But the index is not the economy; the ten largest constituents made up 37.8% of the S&P 500 at the end of August, the largest alone 8.1%.⁵⁴

Pricing the Fall

The hyperscalers may not come unstuck in credit. Their share prices can. And because almost every investor owns them, through index funds, active funds and hedge funds alike, the question of how far they could fall deserves an answer, even if only a framework for one.

There are three layers.

Normalised earnings. Today's earnings are flattered in the ways described above. Restating them means recalculating depreciation on a shorter, more realistic life for the chips; removing gains from revaluing stakes in their own customers; allowing for write-downs on capacity that is not filled; and testing the core businesses against a downturn. The depreciation calculation alone, for companies spending well over $100 billion a year, could remove a large share of reported profit.

The multiple. Investors pay a high multiple for confidence that today's spending will earn a return. When that confidence fades, the multiple falls before the earnings do. The table above shows how quickly the two compound: earnings down a fifth and the multiple down a quarter take 40% off the share price.

History suggests those are not extreme numbers, even for exceptional companies. Between 2000 and 2002, Cisco lost more than four-fifths of its value and Microsoft roughly two-thirds, without either coming near insolvency. In 2022, with no credit event at all, Meta lost about three-quarters of its value, largely because investors punished heavy spending on the metaverse that it could not show would pay. Meta in 2022 is the closest precedent: a highly profitable company spending tens of billions without being able to say what it would earn. Its chief executive has just described the same thing, at several times the scale.

Ownership. This is what turns a repricing into an overshoot. Index funds do not sell on valuation; they sell when money leaves them, and because they hold by weight, outflows fall hardest on the largest companies. The ten largest now make up nearly two-fifths of the S&P 500.⁵⁴ Hedge funds are crowded in the same names, and when they cut risk they sell what they own most of and can sell most easily. And the largest buyer of these shares for a decade, the companies themselves through buybacks, now competes with capital spending for the same cash, and loses where free cash flow has turned negative, as it has at Amazon and Oracle. The cash that used to support the share price in a sell-off is going into data centres.

Put the three together and the drawdowns can be ranked. The ranking will not match the credit ranking; the company most at risk in credit may not be the one most at risk in its share price. The separate note on the equity charts will set out the numbers, company by company, with current prices.

Four Endings

The three parts of this series point to four broad outcomes through the end of the decade. I would not assign probabilities. We do not yet have reliable figures for the independent profitability of the final customers, and a number would disguise judgement as measurement. But each outcome can be defined, and each can be tested.

Two histories sit behind the last two rows. The investment-cycle correction is 2001: the network was built, the technology won, and the lenders to the builders lost. The credit and liquidity event is 2007: a financing machine built to sell a product, mortgages to sell houses then and what Part II calls Originate-to-Ship now, seizes when its marginal paper stops clearing. Neither repeats exactly. This time the floor of the stack is made of the richest companies in the world, which makes a wholesale 2008 less likely than a tiered market in which the weaker borrowers are repriced and the strongest hold. But the sequence of 2007 is the one being followed. On that clock, as Part II sets out, credit now reads about July and equities closer to October. The August moment, when a fund gates or a funding line is pulled, would mark the move from the third row to the fourth.

The first two are good for technology. The third may be constructive over a long enough horizon. The fourth is a financial disturbance that cannot yet be treated as established.

What to Watch, in Order

A large enough dashboard will always contain evidence for any conclusion. So the indicators are worth ranking by what they tell us and how quickly.

The third tier decides the outcome. But by the time it has conclusively turned, prices in the first will long since have moved. The three have to be read together.

What Would Change the View

A view that cannot be disproved is of little use in markets. The risk thesis would be materially weakened if:

• The refinancing markets stabilise, and CoreWeave and its peers raise capital on terms their projects can bear.
• Major projects convert power agreements into energised capacity, on tolerable schedules and within budgets.
• Model and cloud revenue increasingly comes from customers paying out of useful work, not repeated equity financing.
• Efficiency lowers the cost of compute faster than prices fall, or usage rises enough to offset lower prices.
• Semiconductor earnings stay robust as new capacity arrives, with margins normalising only modestly.

If that combination emerges, the October repricing in the semiconductor and neocloud sectors will have been an aggressive shake-out of a crowded trade, not the start of a downturn. It is also possible that the thesis proves right about leveraged infrastructure owners and too bearish about chipmakers, equipment suppliers and the buyers of distressed assets. That would refine the view, not refute it.

Where the Money Settles

The consequences will not be evenly distributed, and that is where the opportunities lie.

Energised capacity becomes more valuable. A finished, powered, fully let data centre with a strong tenant is worth more in a world of shortages and delays, not less. The constraint that hurts the builder of an unfinished campus protects the owner of a finished one. Applied Digital and Galaxy, which delivered on time because they secured equipment early, show the premium on execution; the established operators of energised capacity hold the same advantage at scale.

The bottleneck owners keep their rents longest. Those who own what cannot be expanded quickly are paid first and paid regardless of whether the buyer later earns a return: powered land and grid connections; transformers, switchgear and turbines (GE Vernova, Siemens Energy, Mitsubishi Heavy Industries, Schneider Electric, Eaton); leading-edge manufacturing and lithography (TSMC, ASML); high-bandwidth memory (SK Hynix, Samsung, Micron), for as long as the shortage lasts; and the metals in Part I.

The credit stack matters more than the sector. Senior debt backed by a hyperscaler's contract and guarantee, such as the bonds for Meta's Hyperion campus in Louisiana, is a different risk from the unsecured debt of a leveraged neocloud such as CoreWeave, even when both finance the same chips in the same kind of building. The market has started to say so. It has further to go.

Strong hyperscalers may gain from weak ones. A financially strong hyperscaler (Google, Microsoft, Amazon, Meta) can buy, lease or take over capacity abandoned by a leveraged competitor, possibly at distressed prices. Microsoft taking capacity at Abilene, Texas, that OpenAI declined, and Meta courted for the same site, shows how readily capacity changes hands. Consolidation is one of the likelier ways this cycle resolves.

Materials stay tight regardless. Whichever scenario prevails, the grid, the turbines and the copper must still be built and mined, and the metals in short supply in Part I are in short supply for reasons that have little to do with AI's fortunes: falling grades, long lead times and supply that is a by-product of other mining.

Two Clocks

Somewhere in the desert outside Santa Teresa, the buildings of Project Jupiter are nearly finished. Four thousand people have built them. The cooling system has been shown to the press. The power is not there yet, and when it comes, the chips it switches on will be competing with others that have not yet been made, to sell intelligence at a price that has halved since the spring.

That is the whole Catch in one site. Too slow, because the electricity and the equipment cannot be delivered on the timetable the financing assumed. Too soon, because the debt began accruing before the revenue could. And possibly too cheap, because by the time the revenue arrives, the product it sells may cost a fraction of what the building was financed to charge.

None of it requires artificial intelligence to fail. It may well prove more useful, more capable and more widely adopted than its most enthusiastic advocates expect. The internet arrived, and the telecommunications companies still had to pay their creditors. Americans still needed houses in 2007; the machine that financed them failed anyway.

The question was never whether the technology is real. It was whether the price of building the future has become detached from the cash available to pay for it. Two clocks are running. The technology clock runs in months. The capital clock runs in years. The distance between them is where the losses will be found, and the people who lent against that distance are the ones who will find them.

The technology can be right. The financing can still be early.

Ben Davies
Co-founder · Hinde Capital · London, 19 October 2026

Sources

1. 1Financial Times, What's really going on with AI token price deflation?, 5 Oct 2026, https://www.ft.com/content/96184ef4-05be-467c-9d9c-745e21dad31a; Silicon Data, LLM Token Expenditure Index (SDLLMTK).↩
2. 2OpenRouter, The Token State of Play, 22 Sep 2026, https://openrouter.ai/data/newsletter/the-token-state-of-play↩
3. 3Reuters Breakingviews, Anthropic's $2trn Dream Faces Deflation Trap, 15 Oct 2026, citing Ramp, https://www.reuters.com/commentary/breakingviews/anthropics-2-trln-dream-faces-deflation-trap-2026-10-15/↩
4. 4OpenRouter, Cache Is Crucial, 8 Oct 2026, https://openrouter.ai/data/newsletter/cache-is-crucial↩
5. 5Deutsche Bank Research Institute (Adrian Cox, with Azeem Azhar of Exponential View), AI boom or doom: what will superintelligence feel like and when will we get there?, Thematic Research, 12 Oct 2026, pp.3 and 18, citing Epoch AI.↩
6. 6Author's illustrations (HindeSight analysis, not published).↩
7. 7Erik Brynjolfsson, Avinash Collis, Felix Eggers, Sophia Kazinnik and David Nguyen, What Is Generative AI Worth?, Stanford Digital Economy Lab working paper, 13 Apr 2026 (surveys Jul 2025, N = 1,491, and Mar 2026, N = 1,908, after cleaning; p.4), https://digitaleconomy.stanford.edu/app/uploads/2026/04/WhatsGenAIWorth.pdf ; publication page, https://digitaleconomy.stanford.edu/publication/what-is-generative-ai-worth/ ; Stanford Digital Economy Lab, GDP-B Surplus Observer (consumer-surplus tracker, Jul 2026 update), https://digitaleconomy.stanford.edu/project/indicators/gdp-b-surplus-observer/ ; AI Economic Indicators, https://digitaleconomy.stanford.edu/project/indicators/↩
8. 8Yotzov and colleagues, survey of nearly 6,000 executives, NBER working paper w34836, Feb 2026, https://www.nber.org/papers/w34836↩
9. 9McKinsey, The state of AI in 2026: On the road to ROI, 25 Aug 2026, https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai↩
10. 10Brynjolfsson, Li and Raymond, Generative AI at Work, NBER working paper w31161, https://www.nber.org/papers/w31161↩
11. 11Ramp Economics Lab, Businesses are changing AI model providers more often, 6 Oct 2026, https://ramp.com/data/businesses-are-changing-ai-model-providers-more-often↩
12. 12Stijn Van Nieuwerburgh, Financing the AI Buildout, Brookings Papers on Economic Activity, conference draft, Sep 2026, https://www.brookings.edu/wp-content/uploads/2026/09/4c_Van-Nieuwerburgh.pdf ; summary, https://www.brookings.edu/articles/financing-the-ai-buildout/ ; Columbia Business School record, https://business.columbia.edu/faculty/research/financing-ai-buildout↩
13. 13Gartner, Gartner Says Over $3.5 Trillion of AI Infrastructure Spending Could Be Exposed to Impairment by the End of 2030, press release, 14 Oct 2026 (Will Sommer), https://www.gartner.com/en/newsroom/press-releases/2026-10-14-gartner-says-over-us-dollars-three-and-a-half-trillion-of-artificial-intelligence-infrastructure-spending-could-be-exposed-to-impairment-by-the-end-of-2030 (investment ~$1trn in 2026 and ~$6trn 2027-30; $8.5trn revenue needed at 12% ROIC v $3.6trn attributable AI revenue; >$3.5trn at impairment risk by 2030, >$4trn by 2032)↩
14. 14Stanford Digital Economy Lab, update of Canaries in the Coal Mine?, 12 Aug 2026, https://digitaleconomy.stanford.edu/news/canariesaug26/↩
15. 15Financial Times, 14 Oct 2026, on Anthropic's third-quarter results, https://www.ft.com/content/11632d03-358c-454b-af95-d06d4072932b ; The Information (corrected to state the 11% margin is for the year through Q3).↩
16. 16Reuters, on Anthropic's draft IPO prospectus, 28 Sep 2026, Anthropic's IPO prospectus shows sweeping AI vision, surging costs, https://www.reuters.com/business/finance/anthropics-ipo-prospectus-shows-sweeping-ai-vision-surging-costs-2026-09-28/↩
17. 17Financial Times, report on an OpenAI company presentation (July 2026), 18 Sep 2026, as reported by Reuters, OpenAI expects to burn through almost $280 billion by 2030, 18 Sep 2026, https://kfgo.com/2026/09/18/openai-expects-to-burn-through-almost-280-billion-by-2030-ft-reports/ (management projections, not audited forecasts).↩
18. 18Financial Times, OpenAI annualised revenues $20bn less than previously signalled, 8 Oct 2026 (run-rate and gross versus net treatment of cloud-partner sales), https://www.ft.com/content/b66a9858-f8fb-46cb-b506-44bfe26fca2a ; SemiAnalysis estimate of Anthropic's September 2026 run-rate, as reported by the FT. Run-rates are annualised figures, not recognised revenue.↩
19. 19The Information, on Anthropic's 2028 forecast, 4 Nov 2025; Reuters, on Anthropic's 2028 projection of $190bn to $200bn, 14 Aug 2026; The Information, on OpenAI's forecasts, 23 Apr 2025; Financial Times, 18 Sep 2026; The Next Web, OpenAI's $856bn compute v $600bn reset, https://thenextweb.com/news/openai-856bn-compute-vs-600bn-reset-burn-improved↩
20. 20Silicon Data, B300 Rental Price Index (standardised daily index; $6.96 per GPU-hour on 16 Oct 2026), https://www.silicondata.com/products/silicon-index/b300 ; Telnyx, B300 GPU rental price per hour 2026, reviewed 12 Oct 2026, https://telnyx.com/resources/b300-gpu-rental (posted on-demand prices; a vendor compilation).↩
21. 21OpenAI, completion of recapitalisation as OpenAI Group PBC, 28 Oct 2025, https://openai.com/our-structure/ ; Anthropic, public benefit corporation; Reuters and FT on Anthropic's draft prospectus and OpenAI's IPO preparations, 2026.↩
22. 22CNBC, OpenAI CFO Sarah Friar says company isn't seeking government backstop, 6 Nov 2025; The Register, 6 Nov 2025; David Sacks on X, 6 Nov 2025, https://x.com/DavidSacks/status/1986476840207122440 ; Sam Altman on X, 6 Nov 2025.↩
23. 23Senator Elizabeth Warren, Ranking Member, Senate Banking Committee, letter to David Sacks and Michael Kratsios, 18 Nov 2025, https://www.banking.senate.gov/newsroom/minority/warren-presses-trump-administration-on-plans-to-prop-up-openai-and-big-tech-with-taxpayer-dollars-at-the-expense-of-working-class-americans↩
24. 24Intel, Intel and Trump Administration Reach Historic Agreement, 22 Aug 2025 ($8.9bn for 433.3m shares at $20.47, a 9.9% stake, funded from CHIPS Act grants), https://www.intc.com/news-events/press-releases/detail/1748/intel-and-trump-administration-reach-historic-agreement-to ; Executive Order 14318, Accelerating Federal Permitting of Data Center Infrastructure, 23 Jul 2025, https://www.whitehouse.gov/presidential-actions/2025/07/accelerating-federal-permitting-of-data-center-infrastructure/ ; US Department of Energy, sixteen federal sites identified for data centres, Apr 2025, https://www.energy.gov/articles/doe-identifies-16-federal-sites-across-country-data-center-and-ai-infrastructure ; advanced manufacturing investment credit, 26 U.S.C. §48D (CHIPS and Science Act, 2022; 25%, rising to 35% for property placed in service after 2025), https://www.irs.gov/credits-deductions/advanced-manufacturing-investment-credit ; Executive Order 14365, Ensuring a National Policy Framework for Artificial Intelligence, 11 Dec 2025, https://www.whitehouse.gov/presidential-actions/2025/12/eliminating-state-law-obstruction-of-national-artificial-intelligence-policy/↩
25. 25The White House, Executive Order 14434, 29 Sep 2026, https://www.whitehouse.gov/presidential-actions/2026/09/inaugurating-the-era-of-super-intelligence/ ; White House Accord on Super Intelligence: Joint Commitment on Frontier Responsibilities, 29 Sep 2026, signed by the President, Sundar Pichai, Dario Amodei, Mark Zuckerberg, Greg Brockman, Elon Musk and Jensen Huang, https://www.presidency.ucsb.edu/documents/white-house-accord-super-intelligence , full text as published by The Week, 30 Sep 2026, https://www.theweek.in/news/sci-tech/2026/09/30/ai-safety-white-house-accord-analysis.html , and White House copy, https://d3i6fh83elv35t.cloudfront.net/static/2026/09/accord.pdf ; "morally binding": the President's remarks, 29 Sep 2026, at about 00:57, https://www.whitehouse.gov/videos/president-trump-delivers-remarks-sep-29-2026/↩
26. 26All-In podcast (Chamath Palihapitiya, Jason Calacanis, David Sacks, David Friedberg): episodes of 18 Jul (Can the AI Industry Regulate Itself?), 24 Jul (Fight Over Open Source AI) and 26 Sep 2026 (Anthropic IPO at Risk… Open Source Gains Share); Friedberg clip, https://x.com/theallinpod/status/2107497528350408979 ; David Sacks on X, 14 Oct 2025; White House announcement of PCAST members, 25 Mar 2026 (Sacks co-chair; Friedberg member); PCAST order, https://www.whitehouse.gov/presidential-actions/2025/01/presidents-council-of-advisors-on-science-and-technology/↩
27. 27Reuters, Intel CEO says Chinese frontier AI model adoption has been very fast, 12 Oct 2026, https://www.reuters.com/world/china/intel-ceo-says-chinese-frontier-ai-model-adoption-has-been-very-fast-2026-10-12/↩
28. 28OpenAI, Disrupting a coordinated model-distillation campaign, 30 Sep 2026, https://openai.com/index/disrupting-a-coordinated-model-distillation-campaign/ ; Anthropic, Detecting and preventing distillation attacks, 23 Feb 2026, https://www.anthropic.com/news/detecting-and-preventing-distillation-attacks ; Anthropic, Strengthening our defenses against distillation attacks, 12 Oct 2026, https://www.anthropic.com/research/distillation-defenses↩
29. 29CoreWeave, Inc., Form 8-K, 15 Sep 2025, https://www.sec.gov/Archives/edgar/data/1769628/000176962825000047/crwv-20250909.htm↩
30. 30Ares Management, In the Gaps, Alternative Credit Newsletter, Fall 2026, https://www.ares.com/content/dam/aresmgmt/Documents/In-the-Gaps/Ares_InTheGaps_Newsletter_Fall-2026.pdf ; Broadcom Inc., Form 10-Q, quarter ended 2 Aug 2026, Note 10 (conditional convertible notes up to $42bn, none issued at 2 Aug; lease backstop up to $29bn)↩
31. 31Nvidia, CFO Commentary on Second Quarter Fiscal 2027 Results and Form 10-Q, quarter ended 26 Jul 2026, https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/nvda-20260726.htm (SB Energy guarantees: ~4.25 GW IT load at the PORTS campus, 20-year leases to OpenAI, capped at $105bn in aggregate, effective on lease commencement, first expected in fiscal 2029).↩
32. 32CNBC and Bloomberg, 12 Oct 2026, on Hock Tan's remarks in Singapore. CNBC, https://www.cnbc.com/2026/10/12/intel-broadcom-ceos-warn-capital-crunch-could-become-ai-bottleneck.html↩
33. 33CoreWeave, Inc., investor presentation, Sep 2026, filed with the SEC, https://www.sec.gov/Archives/edgar/data/1769628/000176962826000429/crwv2026convertpresentat.htm (management statements).↩
34. 34Bloomberg (Lynn Doan), The Staggering Costs of a Single AI Data Center, 12 Oct 2026, https://www.bloomberg.com/features/2026-ai-data-center-build-costs/ ; Bloomberg Law, https://news.bloomberglaw.com/international-trade/the-staggering-costs-of-a-single-data-center-from-land-to-chips↩
35. 35Cushman & Wakefield, Data Center Development Cost Guide 2026, lead-times overview: buyers with existing supplier relationships and a large annual spend "often secure significantly shorter lead times compared to new market entrants". Report: https://www.cushmanwakefield.com/en/united-states/insights/data-center-development-cost-guide ; digital edition, https://digital.cushmanwakefield.com/datacenterdevcostguide-amer-content-pds-datacenter-/ ; release, https://ir.cushmanwakefield.com/news/press-release-details/2026/Cushman--Wakefield-Releases-2026-Data-Center-Development-Cost-Guide-Citing-21-Rise-in-Per-MW-Construction-Costs/default.aspx↩
36. 36Jay W. Forrester, Industrial Dynamics, 1961 (the bullwhip effect), https://books.google.com/books/about/Industrial_Dynamics.html?id=KRzHAAAAIAAJ↩
37. 37Morgan Stanley Research (Nicholas Lentini, Michael Wilson et al.), Following AI's $1.6 trillion Money Trail: Mapping the Revenue Generated by the Capex Boom, US Equity Strategy, North America Insight, 15 Oct 2026, pp.17-18 (Joseph Moore, semiconductors).↩
38. 38Satya Nadella, Bg2 Pod interview, Nov 2025, as reported by DatacenterDynamics, https://www.datacenterdynamics.com/en/news/microsoft-has-ai-gpus-sitting-in-inventory-because-it-lacks-the-power-necessary-to-install-them/↩
39. 39Nvidia, results for the second quarter of fiscal 2027 (quarter ended 26 Jul 2026; released 26 Aug 2026): Form 10-Q, https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/nvda-20260726.htm ; Form 8-K, https://www.sec.gov/Archives/edgar/data/1045810/000104581026000073/nvda-20260826.htm ; results release, https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000073/q2fy27pr.htm ; Nvidia investor relations, https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Announces-Financial-Results-for-Second-Quarter-Fiscal-2027/↩
40. 40Nvidia earnings call and CFO commentary, 26 Aug 2026, https://s201.q4cdn.com/141608511/files/doc_financials/2027/Q227/Q2FY27-CFO-Commentary.pdf ; Form 10-Q (Hyperscale $48.710bn; AI Clouds, Industrial & Enterprise $40.313bn, +138%, on recast comparatives); earnings-call transcript, 26 Aug 2026, p.7 (Colette Kress: gross margin 74% in Q3, trough 71-72% in Q4, 72-73% in FY2028, citing memory costs); market value from closes of $209.66 (26 Aug) and $227.98 (27 Aug) on 24.147bn shares, a rise of about $442bn to about $5.51trn (author's calculation, Yahoo Finance prices).↩
41. 41Broadcom Inc., Third Quarter Fiscal Year 2026 Financial Results, 2 Sep 2026, https://www.sec.gov/Archives/edgar/data/0001730168/000173016826000076/avgo-08022026x8kxex99.htm↩
42. 42TSMC, third-quarter 2026 results (Form 6-K) and earnings call, 15 Oct 2026, https://investor.tsmc.com/english/quarterly-results/2026/q3 ; Form 6-K, https://www.sec.gov/Archives/edgar/data/1046179/000104617926000726/tsm-20261015x6k.htm↩
43. 43Micron Technology, Record Fiscal Fourth-Quarter and Full-Year 2026 Results, 30 Sep 2026, https://www.sec.gov/Archives/edgar/data/0000723125/000072312526000018/a2026q4ex991-pressrelease.htm (gross margin 86.8% GAAP, 87.0% non-GAAP)↩
44. 44TIKR, Samsung and SK Hynix Fell Over 5%, Sep 2026 (consensus: Micron gross margin from about 84% towards 65% by fiscal 2030), https://www.tikr.com/blog/samsung-and-sk-hynix-fell-over-5-heres-what-microns-september-30-report-must-show ; Micron closes: $1,213.56 on 25 Jun 2026 (intraday high $1,255) and $1,065.08 on 29 Sep 2026, a fall of about 12.2% (Yahoo Finance historical prices).↩
45. 45Reuters, ASML forecasts Q4 sales above expectations on AI chip demand, 14 Oct 2026, https://www.reuters.com/business/asml-forecasts-q4-sales-above-expectations-ai-chip-demand-2026-10-14/ (chief executive: "some customers are expressing an interest in placing orders with us for 2029 and even 2030").↩
46. 46HindeSight, A Credit Cycle Wearing AI's Costume, Aug 2026, citing Amazon and Nvidia filings (Nvidia 10-Q to Jul 2026). https://hindesight-letter.netlify.app/↩
47. 47Microsoft, FY2026 fourth-quarter earnings call, Jul 2026, https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q4 ($41bn includes finance leases; cash PP&E purchases $35.8bn).↩
48. 48CoreWeave, Inc., Form 8-K, 7 Aug 2026 (DDTL 5.5: maturity 1 Sep 2031; parent guarantee), https://www.sec.gov/Archives/edgar/data/1769628/000176962826000357/crwv-20260807.htm↩
49. 49Moody's, as cited in Stijn Van Nieuwerburgh, Financing the AI Buildout, Sep 2026, https://www.brookings.edu/wp-content/uploads/2026/09/4c_Van-Nieuwerburgh.pdf ; summary, https://www.brookings.edu/articles/financing-the-ai-buildout/ ; Columbia Business School record, https://business.columbia.edu/faculty/research/financing-ai-buildout↩
50. 50Korea Herald, 19 Jun 2026, on record margin debt in Samsung Electronics and SK Hynix (syndicated copy), https://v.daum.net/v/20260619182119174 ; Korea Times, 16 Oct 2026, https://www.koreatimes.co.kr/economy/20261016/seoul-stocks-rebound-over-2-on-eased-chip-rout-fed-bets ; Seoul Economic Daily, 15 Oct 2026, https://en.sedaily.com/finance/2026/10/15/kospi-plunges-76-percent-to-six-month-low-in-black-thursday and https://en.sedaily.com/finance/2026/10/15/kospi-plunges-as-forced-sales-hit-leveraged-retail-investors ; Seoul Economic Daily, 17 Oct 2026, https://en.sedaily.com/finance/2026/10/17/forced-stock-sales-top-80-billion-won-as-kospi-plunges-76 ; Startup Fortune, 15 Oct 2026, https://startupfortune.com/south-koreas-kospi-sinks-757-as-chip-fears-compound-a-brutal-week/↩
51. 51VanEck, Semiconductor ETF (SMH), holdings (SK Hynix 3.80% at 15 Oct 2026), https://www.vaneck.com/us/en/investments/semiconductor-etf-smh/ ; iShares, Semiconductor ETF (SOXX), https://www.ishares.com/us/products/239705/ishares-phlx-semiconductor-etf ; Benzinga, 6 Jul 2026, https://www.benzinga.com/etfs/sector-etfs/26/07/60290857/semiconductor-etfs-ai-2026-psi-smh-soxx-soxq-xsd↩
52. 52FRED (Nasdaq; S&P Dow Jones Indices), daily price-index closes, 1 Jul to 16 Oct 2026: PHLX Semiconductor (NASDAQSOX), https://fred.stlouisfed.org/series/NASDAQSOX ; Nasdaq 100 (NASDAQ100), https://fred.stlouisfed.org/series/NASDAQ100 ; S&P 500 (SP500), https://fred.stlouisfed.org/series/SP500 ; checkpoints verified against Nasdaq, Yahoo Finance and ChartExchange.↩
53. 53FactSet, John Butters, S&P 500 Earnings Season Update: October 16, 2026, https://insight.factset.com/sp-500-earnings-season-update-october-16-2026↩
54. 54S&P Dow Jones Indices, S&P 500 index characteristics, as of 31 Aug 2026, https://www.spglobal.com/spdji/en/indices/equity/sp-500/↩

Talk Your Book

The small print. Read it anyway.

This is not investment advice. It is one person thinking out loud about markets, usually with a position on, so assume I may be talking my own book. That is not a confession, merely useful information. Views can change quickly, occasionally before the ink is dry, and I am not going to defend a trade simply because I once wrote about it with great conviction. Nothing here is an offer, solicitation or recommendation to buy or sell anything.

I write HindeSight primarily for me. The process forces me to test the thesis, join a few dots, throw some away and, with luck, understand the problem rather better than when I started. If it sparks a thought in someone else, even better. There is no publication schedule. I write when I have something to say or when the urge becomes sufficiently annoying that writing it down is easier. I can be tactically bearish and medium-term bullish, long one week and short the next. Markets rarely reward consistency for consistency’s sake. My objective is simpler: find the best risk/reward and try to extract as much upside as possible while giving as little back as possible.

Hinde Capital no longer manages external capital, so there is nothing to sell you. Past performance is no guide to future returns and, frankly, neither is this letter. Do your own work, size risk so you can still sleep, and seek regulated advice where appropriate. No liability is accepted for losses arising from reliance on this material, including those incurred after forwarding it to your sister- or brother-in-law. © Hinde Capital 2026.

Hinde Capital · hindecapital.com · Not investment advice. Utterly talking my own book.