When does a company sell its product at a deliberately low price? The honest list is short. First, when the product is inferior, the low price serves as an apology to buyers willing to trade features for savings. Second, when the seller expects to lock the buyer into an ecosystem where switching hurts and to collect the relationship's true price over many years. One may append the occasional political calculation, where a government's preference overrides a seller's arithmetic. What the list has never contained is a third entry: the fact that the product comes from China. Yet a remarkable share of commentary still treats national origin as a pricing strategy, as though geography shipped with a discount attached.
Cheapness is not a nationality. It is a transfer, and somebody always pays it. The payer may be an advertiser, an investor, a government, a complementary product, or a future customer who does not yet know he has been volunteered. This summer, across several industries and both hemispheres, payers have been quietly withdrawing, and price-setters are exhibiting behavior inconsistent with generalized perceptions. This letter is about the presumptions we need to be careful about. Not the presumptions about technology, which we audited in The World of Zero Visibility, but the ones about strategy: the quiet assumptions about how companies ought to behave, formed in an era of cheap physics, cheap capital, and patient customers, and applied unchanged to an era with none of the three.
The Sale Signs Never Came Back
We have been writing about the behavioral changes in pricing across global economies and across sectors for a while. To make the point, in an article three years ago, we started with a quiz about a company selling 100 widgets at $100 apiece, and observed that the scholars who would have aced every business examination of the past three decades kept reaching for the price cut, while the world's actual price-setters had stopped reaching for it somewhere during Covid. Our main conclusion, or a viewpoint, was that we live in a world where the default management decision on pricing is no longer a cut but a r(a)ise.
In other words, we live in a world of few Sale signs. Half-empty restaurants standing beside other half-empty restaurants raise prices to recover the volumes they no longer have. Software companies with tiring subscriber growth raise prices to protect the top line, quietly retiring the old religion of land now and expanding later. The sector that spent forty years as the deflation engine of the consumer price index now contains its hottest-inflating category, video subscriptions, rising roughly ten times faster than the basket around it.
The world has since supplied two sharper editions of the rule. Where a seller cannot assume that today's subscriber will remain a subscriber in the next product cycle, buying loyalty with discounts is like buying ice in summer. And where a seller sits in severe undersupply, facing demand that barely flinches at price, a discount is a needless charity. Almost everything that follows is one of those two editions wearing work clothes.
The Logo Has Lost Its Discount
Consider the year's most instructive courtship. Apple has publicly spent months qualifying memory from CXMT for machines sold in China, lobbying Washington for permission and absorbing senatorial disapproval for trying, while HP and Acer were already shipping the Chinese memory abroad. The context made the courtship rational: contract prices for standard DRAM surged by more than half in early 2026 as AI servers pulled capacity away from devices, and Apple raised prices across nearly its entire lineup as the costs rose.
If reports are true, CXMT has recently refused to supply Apple at any discount to the prices of other memory makers that Apple wants to replace. The reason is perhaps the oldest in commerce. A company once dismissed as a subsidized also-ran is now the market darling; it does not consider selling to Apple a privilege that deserves a discount. Even assuming that Apple’s unusual public courting had no hidden motive of pressuring other suppliers, CXMT does not need Apple’s name on the roster to generate more demand for its products.
For decades, a marquee customer's business carried a premium worth paying for, because the logo validated the supplier. In an undersupplied market, demand validates itself. The queue behind the reference customer is validation enough, and the reference customer's discount has quietly gone to zero.
TSMC makes the same point from the incumbent's chair, which is more damning. Negotiations concluded in July for pricing that takes effect in January 2027: base increases of 5% to 10% across advanced nodes, up to 10% on mature ones, and a further 10% to 15% surcharge for high-performance computing orders beyond a customer's original forecast, which stacks toward roughly 25% on exactly the AI work everyone wants. The stated reasons include the cost of building and running fabs outside Taiwan. Read that slowly. Geopolitical de-risking was supposed to be funded by the supplier's margin or by government subsidy. It is being funded by the buyer, and it now appears on the invoice. The most elegant detail is the timing: the increases were deferred to 2027, according to the reporting, to give customers such as Apple, Nvidia, and AMD time to accommodate the changes and adjust their own prices accordingly. Putting it simply, here is a supplier instructing the most valuable companies on earth to go and raise their prices too.
The memory makers have dispensed with the courtesies altogether. All three, by industry accounts, have allocated their entire 2027 DRAM and high-bandwidth output, with buyers receiving perhaps 60% to 70% of what they asked for. By our calculations, well over USD35bn are now sitting on supplier balance sheets as deposits and advance payments.
Suppliers rebuffing big-name customers is not just limited to the AI world. When Walmart approached its consumer-goods suppliers this summer for lower prices in exchange for better placement, most simply said no. Buyer power survives where supply is plentiful and dies where it is scarce, and it turns out the scarcity is not confined to silicon.
There is a side reason. Many of the previous-era giants are no longer the ones growing the fastest or with the same halo in the new world of innovation. When laptop and handphone volumes are crumbling, it is quite obvious that downstream supply chain companies like Quanta or Hon Hai will not spend as much time pleasing clients who are no longer as important as they once were.
Walls With Doors
The geopolitical presumption has fared no better. The comfortable model held that the technology world had split into two sealed halves and that commerce across the wall would mostly be through the breaches of the officially disallowed. The last few weeks featured headlines about US-China corporate alliances that no other country pair can match.
Start where nobody was looking. Ford and Geely have agreed to build cars together at Ford's Valencia plant for the European market. An American brand and a Chinese one are sharing a Spanish factory because an idle shift has no nationality. A year ahead of schedule, GM and SAIC extended their joint venture in China to 2047. AstraZeneca (of course, not a US company) has signed a 51:49 joint venture with CSPC to build a biologics plant in Shijiazhuang, whose stated purpose is to supply global markets. Ford's battery plant in Michigan runs on CATL technology under a licensing and services arrangement in which the Chinese partner takes fees and no equity. Meanwhile, Western pharmaceutical companies keep paying record-high advance payments while signing long-term agreements for molecules originating in China.
After the items above, the semiconductor evidence appears to be a pattern rather than an anomaly. Samsung and SK Hynix were reported in August to have spent some two years evaluating Chinese etching tools for their Chinese fabs, hedging against the day Washington restricts not merely new Western equipment but the servicing of the old. Samsung denies it, which we note without further comment; the report carries three sources. Around them, the numbers move regardless: four Chinese equipment makers each heading for a billion dollars of revenue this year, roughly a third of their home market. And in the fourth quarter, Huawei intends to sell AI chips in South Korea, the home market of the memory giants, at a fraction of the price of the American export version.
None of this means the walls are decorative. Export rules bind, licenses expire, senators write letters, and every door now comes with a tariff, a license, and a queue. By the time this note reaches the investors, there might be more restrictions. The claim is narrower: borders set the cost of the route, not the customer's existence at its end. Every serious participant treats the wall as a constraint to be engineered around and priced accordingly, and industrial policy, on this evidence, creates optionality rather than sealed blocs.
Both Directions at Once
Which brings us to the summer's most under-read exhibit about the Chinese models turning less expensive (our apologies for the double negative).
For two years, the cheapest serious intelligence on earth has come with open weights and a Chinese passport, and a comfortable belief grew around that fact: the Chinese labs would flood the world and hold prices down forever. Today, as this letter goes out, DeepSeek's new rate card takes effect. Prices on its V4 models rise by anywhere from about 50% to more than eleven times, depending on model, token type, and hour of the day, arriving alongside a peak and off-peak structure with the busiest windows charging double the quiet ones, and arriving alongside a funding round of some $7 billion and preparations for a listing. We are not saying that the company that is credited with beginning the price war is ending it, but the trend is not as simple as most of us assumed. Bloomberg's summary is the whole thesis in a headline: prices rising toward the rivals. Moonshot's Kimi increased its output price over a year of releases; Zhipu raised its coding plan by an official 30%; ByteDance began charging consumers for the first time.
Now watch the other direction. In late July, OpenAI cut GPT-5.6 Luna by 80% and trimmed the mid-tier Terra by 20%, and left its flagship untouched. Anthropic launched Opus 5 at half the price of its previous flagship and then scrapped a planned September price increase for Sonnet 5, keeping it at its launch rate. One token-price index has seen American frontier pricing down by nearly a quarter since mid-July. The names doing the cutting are the ones assumed untouchable, and the names doing the raising are the ones assumed permanently cheap.
Both presumptions died in the same fortnight, and they died of the same cause. Nationality never set these prices. Arithmetic did. The American labs cut because enterprise customers began scrutinizing invoices, and open-weight alternatives achieved performance close enough to make the comparison uncomfortable, with DoorDash, Siemens, and Airbnb among those reported to be testing them. The Chinese labs raised because serving capacity is expensive, because their hardware is constrained, and because a company preparing to list needs unit economics rather than applause. And note what DeepSeek's clock-based pricing really says: a single all-hours price is a subsidy from the patient customer to the impatient one, and that subsidy has been withdrawn too. The price is increasingly a timetable dependent on the timezone and not the provider’s country of origin.
Renting the Right to Leave
Time to look inside the contracts, because the fine print is rewriting strategy more quietly than any press release. The largest compute lease signed this summer carries a cancellation clause. The memory sellers, at the other extreme, take deposits and prepayment before parting with 2027 supply, and even their multi-year agreements are reported to carry renegotiation windows every twelve months. One cloud builder disclosed that $75 billion of its contracted business involves customers prepaying or bringing their own chips.
Notice what has happened to the ancient trade between price and commitment. In the old world, sellers from Asia sold certainty cheap and provided goods on credit. Wherever there were long-term contracts, sellers’ prices and margins were generally capped, with buyers having the right to renegotiate prices and volumes. Now, nobody sells certainty at any price, because nobody can manufacture it. What gets priced instead is flexibility. The resolution is tidier than it first looks: scarcity gets annuitised, and uncertainty gets optioned. Inputs that are certainly scarce, memory, power, and land, are locked up long and paid for early. Outputs whose value nobody can see, the compute itself, the model, the workload, are rented with an exit. The same hyperscalers will sign both kinds of contracts on the same afternoon without contradiction.
As we wrote a year ago in the nationality of the nominal world, the altered patterns are also in the non-tech world. Companies now pay a substantial premium per square foot for flexible office space rather than sign ten-year leases. The wind industry has furnished the decade's cleanest natural experiment. Vestas spent two years exiting contracts on unacceptable terms and embedding inflation indexation in the rest. A Mitsubishi-led consortium that had bid aggressively for Japanese offshore projects without such buffers withdrew from three of them, at a cost reported around ¥52 billion. Vestas sells the turbine and maintenance; it declines to assume macroeconomic risk.
In the world of zero visibility, the contract book of the world is turning from an annuity book into an options book, and a chief financial officer who signs a decade of fixed commitments in a field that resets every two quarters has not bought stability. He has sold his company's ability to change its mind.
The Balance Sheet Follows the Eyesight
A second migration is underway, visible mainly in accounting footnotes. Hon Hai told investors that its customized AI-server business now operates primarily on a consignment basis, with customers supplying the costly components, shortening its cash-conversion cycle by eight days per year. Wiwynn had moved first. Downstream, the costs land on whoever cannot pass them along: General Motors now expects $1.5 billion to $2 billion of commodity and memory inflation this year, up from its earlier estimate, while raising its profit forecast for the second time.
When we assemble the pieces, a pattern appears that we have not seen before. Inventory, components, and residual risk are migrating to the balance sheets, with the best information on demand. The hyperscaler consigns the processor because only it knows its own utilization. Nvidia underwrites residual values because it is the only one that sees utilization across every cloud at once. The assembler, who sees nothing but purchase orders, sheds the working capital and is glad to do so. This is the inverse of the last great financial migration, when risk pooled on the balance sheets that understood it least. Those were the days when Asia produced cheap and took on the risks because those who were not willing had neighbors ready to replace them.
Of course, the power to transfer belongs to those with innovation moats or who benefit from supply shortages. The migration is per layer, not universal, and there are intermediaries in this chain carrying concentration risk they cannot see. And the party that sees furthest does not carry risk for love; it either carries the risk because it has no other options or charges for the privilege.
How a CEO Reads a Menu
Which brings us to the reading habits of armchairs. The commentary of the season asks why Nvidia would enter the low-margin business of financing, why the hyperscalers tolerate returns on capital below their gilded histories, and who, in the great chokepoint league table, is making the most money. These are entertaining questions. They are not the questions a chief executive is paid to ask.
A chief executive reads a menu. The first item is opportunity cost: what else could this dollar do? For the largest companies, the menu is brutally short. Cash earns money-market rates. Buybacks return capital at, in the ideal case, the cost of capital. Beyond those sit only the ponds deep enough to matter at their size, and there are very few. A merely good return on an enormous and strategically necessary pond beats a spectacular return on something too small to move a four-trillion-dollar needle, and (one of our most repeated lines) the days of supernormal returns without much investing are over. Grading today's projects against that vanished era is measuring every harvest against the year the frost never came.
In short, for companies the size of Nvidia, Amazon, or SpaceX, the list of businesses with potential for meaningful growth is extremely small.
The second item is the one the armchair never orders: what is the return on not doing it? Abstention is not zero. It is watching a rival collect the rent of the era, then paying that rival for access from a weaker seat, for as long as the era lasts. Business history is a museum of large companies that economized their way out of the future. The mainframe king who saw a toy in the personal computer donated the richest industry of the century to two startups, and his returns on existing capital looked excellent the whole way down. Intel, two decades ago, declined to build the chip for a telephone nobody had seen yet, on the sensible ground that the price was below cost and the volumes were imaginary. Its chief executive later said, in as many words, that nobody knew what that phone would become.
Lastly, a CEO cannot run an enterprise on pessimism and hope to gloat the day the world comes to an end. A large company CEO must look at the available evidence at their disposal, embrace complexity, take measured risks, and be ready to pivot if things turn out differently, rather than keep touting tales of past bubbles. This does not mean that a great CEO never decides to sit still or opt out. It simply means that CEOs do not have the luxury of idle pessimism of an armchair columnist or a shorting investor.
The announcements in the supposedly quiet peak summer weeks have been telling. Nvidia signed memorandums with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to mobilize more than $500 billion in third-party capital for compute infrastructure, offering residual-value support of up to 25% on an individual financing basis, project by project, while the financiers retain the underwriting pen. Memorandums are intentions rather than facilities, and we hold the number lightly. Three weeks earlier, a market swoon had spent July asking whether any of this pays. Meanwhile, SpaceX rents out the cluster it built for itself, and Google, the world's most sophisticated infrastructure operator, pays roughly $920 million a month to use somebody else's machines because the binding constraint is time rather than capability. Meta, having spent years building compute infrastructure solely for itself, is reportedly building a business to sell it to others. SK Group is building a two-gigawatt AI factory with Nvidia on its own next-generation memory. Some will call this circular, and a little of it is. But a vendor-financing bubble requires a vendor to be fooled about final demand, and the one company that watches utilization across every cloud on earth is the least foolable participant in the economy.
When the Long Term Arrives
Two kinds of companies are abandoning patience this year, for opposite reasons, and the distinction is worth drawing carefully.
Start with the newcomers, because their arithmetic is the simplest. The venture formula of the last cycle, burn boldly now and harvest the locked-in customer for decades, rested on three subsidies: physics that made each generation of service cheaper, capital that was nearly free to nearly everyone, and customers who, once captured, stayed captured. All three have weakened, and the third is the least discussed. Serving intelligence costs real money in every period; capital is still cheap only for an extremely small club, and when capabilities reset every few months, a subscriber bought with subsidies is a tourist rather than a resident. A young model maker cannot promise investors a decade of customer loyalty when the switching cost is just one afternoon of engineering. So the model makers, American and Chinese alike, now mind revenue and burn in every phase, as their pricing this summer made plain. This is not short-termism. The long game has ceased to exist in the world of zero visibility.
The incumbents are the more interesting case with the arrival of shortages. Their plan always ended in a bill. Free was never generosity; it was an invoice with a later date written on it, and the sender fully intended to collect. What changed this year is that the date moved. The scarcity described in these pages, memory sold out through 2027, compute rationed by the week, power queued for years, arrived on the cost line of every business built to give things away, and the long-term savings they had been saving for turned up before they had finished preparing for it.
The results are visible across every surface where something used to be free. WhatsApp, the last great unmonetized property in the world, now carries advertising in its Updates tab, paid channel subscriptions with the house taking its share, and promoted channels. YouTube has doubled the subscriber and watch-hour thresholds a new creator must clear before earning anything, while requiring existing partners to accept revised terms or stop earning. X has ended creator revenue sharing outright and replaced it with a program that pays only on impressions from other paying subscribers. Google, having spent two decades competing on how much storage it could hand out, now enforces the limit it once advertised, with the upgrade never more than one notification away. Snapchat has surpassed $1 billion in annualized revenue collected directly from users rather than advertisers.
None of these have the form of a price rise. The free tier became a qualification period, the familiar feature moved upstairs to a paid floor, and the payout that once flowed toward volume now flows only toward the audience that pays the bills. A platform that pays creators only for the attention of paying subscribers has stopped buying reach and started buying return, which is the same discipline the model makers are showing, arrived at from the opposite direction and for the opposite reason. One group cannot afford to wait. The other has finished waiting.
The great democratization of information service revolution is under pressure. Every free service always had a payer, whether an advertiser, an investor, a complementary product, or a loss of control over private information of those who did not yet know they had been volunteered. The shortages are forcing service providers to recalibrate the costs of doing business.
The Menu Is Not the Spreadsheet
We have a world full of commentators worrying about investment returns at a time when profit growth across the innovation economy is among the highest on record. They presume the managements committing massive capital to be foolish, or captured by fashion, while the pricing evidence from those same managements shows them more mercenary and more short-term than they have been in a generation. They worry about a shortage of demand while every observable queue is a shortage of supply. It is possible to hold each of these worries on its own. Holding all three at once requires a talent this letter admires but does not possess.
The clash between what is presumed and what is visible ran through every realm this summer touched, and the examples are above. Each presumption was reasonable when made. Each is now contradicted by something in plain sight, which is usually the moment to take stock rather than repeat the presumption more loudly.
Two disclaimers keep this honest. We are not claiming these businessmen are right; several of the arrangements described here will look foolish by Christmas. Nor must critics agree with management, since disagreement is the job. The claim is narrower. The opportunity set in front of the decision-maker is a different document from the one open on the analyst's screen, and the screen has no line at all for the cost of the road not taken.
For a host of reasons, the business of business has turned to business, regardless of nationality, customer loyalty, or general goodwill. It will not stay this way. Capacity arrives, shortages ease, and the courtesies return with them. But while it lasts, it is one more indicator of a changed world, and the least useful response is to keep marking the evidence against a scorecard printed in a different decade.
