TSMC's C.C. Wei is "really jealous" of memory makers' margins but won't charge memory-style rent; our call is that memory is collecting it

Key takeaways

Two things pointing in opposite directions happened on TSMC's July 16, 2026 earnings call. On the results side, TSMC's second-quarter revenue of US$40.2 billion landed at the very top of TSMC's own guidance range, the operating margin of 60.3% beat the top of that guidance by 1.8 percentage points, and net income rose 77.4% from a year earlier. Without a one-time gain from selling shares in Vanguard International Semiconductor, growth is about 63%. Those are the numbers of a company with no capacity left to sell. On the Q&A side, Chairman C.C. Wei was asked whether demand exceeds supply by 30% to 50%. Wei said the gap was very big, began to compare it with something, and stopped: "Sorry, I don't want to make a comment on the memory, but a very big gap." Asked about pricing, Wei said TSMC would not suddenly raise prices 4x or 5x, because customers could not survive that kind of increase. Asked whether TSMC would copy customers that invest in their own customers, Wei answered: "So far, no, TSMC don't do this kind of financial arrangement." The textbook says whoever holds the bottleneck collects the rent. The holder of the deepest bottleneck in the industry turned down, in public, both standard ways of collecting it. One is a multiple-fold price increase. The other is an equity tie to customers' customers.

This piece answers three questions. First, how wide is that refusal really? TSMC is in fact raising prices, and TSMC's gross margin is already at a record. Second, where does the rent TSMC declines to charge end up? Mostly with the memory makers, and the flow can be measured. Third, for anyone weighing the risk in AI's circular financing, what is it worth that the most upstream supplier sits outside the circle? Checked against the published maps of that circle, the foundry appears in none of them.

The four passages from the call are rechecked against TSMC's official transcript; the version that circulated after the call had three distortions, and one of them changes the scope of the argument. Three audits then attack the claim that TSMC is "forgoing rent." Next, the evidence that memory and packaging are each collecting rent is laid out and measured. Finally, three competing explanations are weighed against each other, and the conditions under which this piece fails are written down.

What does this add beyond the daily brief's item on the same call? Until now, this publication held Wei's remarks only through a single source, a reporter's live transcription. This piece matches them for the first time against TSMC's official transcript, in full context. The circulated version turns out to be the reporter's grammatical cleanup, and the original contains a line almost nobody quoted: "So yes, I'm really jealous about memory company's 86% gross margin. 86 -- 68, I'll be happy about that." This piece also attacked its own claim, and half of "forgoing rent" did not survive. At the same utilization rate, TSMC's gross margin is 19 percentage points higher than in 2018-19, by analyst Sravan Kundojjala's comparison, and TSMC is raising prices 5-10% across about 75% of its wafer revenue, according to reporting by Tim Culpan. "Charges no rent" does not hold. What holds is a narrower version: TSMC refuses memory-style rent. The piece ends with the conditions that would overturn the call: any single list-price increase of 20% or more on advanced nodes, or a disclosed equity arrangement between TSMC and an AI customer (or that customer's customers), sends the call back for a new ruling.

One earnings call, two confessions that pull against each other

The evidence base comes first, because it changed. This publication first had the four passages through live posts on X by Dan Nystedt, a semiconductor journalist who transcribes TSMC's calls as they happen. That was a single source, and comparison afterward showed the spoken English had been smoothed. This piece instead uses the call transcript posted on TSMC's investor-relations site, the edited version produced by LSEG StreetEvents, a transcript service. It can be cross-checked against the Motley Fool transcript and TrendForce's same-day report. All four passages are there, and three of them are worth rereading word for word.

First passage: the gap. A Deutsche Bank analyst noted the common assumption that unconstrained demand at 3 nanometers and below exceeds supply by 30% to 50%, and asked whether the real gap was much larger. Wei answered: "No, we don't have a number to share because of -- let me say that the gap is very big. Bigger than… Sorry, I don't want to make a comment on the memory, but a very big gap." Wei braked at "bigger than." Wei declined the comparison, but everyone on the call could tell what the gap was about to be compared with.

Second passage: pricing. A JPMorgan analyst asked about wafer pricing strategy. Wei's full answer runs well past the circulated version. Wei said that "our customer got to be successful. I don't want to squeeze them out from the market." Then: "So we don't suddenly increase our price by which I like to have 4x or 5x." TSMC, Wei went on, earns its value and makes sure "our gross margin is enough for our long-term sustaining expansion," which Wei described as good for customers and for TSMC alike. Then came the line the circulated version left out: "So yes, I'm really jealous about memory company's 86% gross margin. 86 -- 68, I'll be happy about that."

The news is not the refusal itself. Wei does not deny wanting the rent; "which I like to have" sits inside the sentence. Nor is Wei unaware of the going rate. The 86% Wei blurted out is about one percentage point from the gross margin Micron actually reported for its latest quarter. Wei knows the rent is there and says plainly that TSMC will not collect all of it.

Third passage: investing in customers. This publication's earlier record of this exchange got the scope wrong. The analyst from Arete, an equity research firm, did not ask whether TSMC would take stakes in its direct customers. The question was explicit: TSMC's customers are investing in their customers, the end customers; would TSMC consider doing the same? The call's moderator, Jeff Su, checked with the analyst on the spot that the question meant the customers' customers. Wei answered that every company has its own considerations, then: "So far, no, TSMC don't do this kind of financial arrangement, because of we think we are working with current customer, with the current model smoothly and also successfully."

So the question Wei faced was about exactly the kind of deal NVIDIA has with OpenAI, or AMD's warrants to OpenAI. The analyst put the circular-financing template in front of Wei, and that template is what Wei turned down by name. That is narrower than the earlier record, "refuses to invest in customers," and more useful: it is a direct statement on whether TSMC will join the circular-financing loop.

Fourth passage: welcoming a rival into packaging. A monopolist publicly wishing a competitor well is unusual. It makes sense only when the bottleneck in that line of business is so tight that it limits shipments of the main business. A Macquarie analyst asked about adoption of EMIB-T, an Intel advanced-packaging technology that competes for the same AI-accelerator assembly work as TSMC's CoWoS line. Wei answered: "our packaging capacity is so tight that now it limits my customers' growth. So we welcome that additional flexibility in the market and so that will help TSMC's frontend wafer business growth." Wei added that the front-end wafer business is the bulk of TSMC's business, and closed: "we hope they will be successful and so that share some of the loading from TSMC." The customers whose growth is capped are the AI-accelerator makers waiting for CoWoS capacity.

Pressed later on whether back-end competition would eat into TSMC's value, Wei answered a second time. Wei said front-end wafers and back-end packaging are different businesses, and that "since our capacity in the backend is still in shortage mode, the gap is bigger. And so I welcome that the competitor offer some of the flexibility to my customer." One detail is worth logging: Wei never said "Intel" during the exchange. Wei said only "they" and that "the technology looks good, according to the newspaper." The analyst who named Intel was a Bank of America analyst, later in the call.

The financial side of the same call makes the stance look stranger. Per TSMC's press release, revenue was US$40.20 billion, up 33.7% year on year, right at the top of the US$39.0–40.2 billion guidance range. Revenue at the top of guidance means TSMC was limited by capacity, not by demand: it sold out. Gross margin was 67.7% and operating margin 60.3%, both above the top of TSMC's own guidance; the operating margin beat by 180 basis points. Net income rose 77.4%. That figure includes a one-time gain of about NT$63 billion from disposing of shares in Vanguard International Semiconductor, a Taiwanese foundry, worth NT$2.24 of earnings per share. Without it, growth is about 63%.

Chief financial officer Wendell Huang attributed the margin improvement in full as "primarily due to cost improvement efforts and a higher capacity utilization rate, partially offset by margin dilution from overseas fabs." Price increases are not in that margin bridge. That is the mechanism that lets "declines the rent" and "record profit" both be true: the extra earnings come from a falling cost curve and full fabs, not from turning the shortage into price.

"Forgoing rent" has to survive three audits

Up to here, this is TSMC's version of the story. The next step is to attack it. Is "forgoing rent" a false claim dressed in investor-relations language? Three audits go after it, and half of it falls.

Audit one: TSMC already collects a lot more. In July 2026, Sravan Kundojjala, a semiconductor analyst who publishes foundry margin analysis on X, put out a damaging comparison: TSMC's gross margin was 66.2% in the first quarter of 2026, while "the same utilization level in 2018-2019 produced gross margins of 47-48%, a gap of 19 points at identical capacity." Kundojjala's conclusion: "Peak margins no longer require peak loading." Whether those 19 points come from advanced-node mix, yield learning or pricing, TSMC now takes far more value per wafer than it did five years ago. If "forgoing rent" is read as "charging no rent," this audit kills it outright.

Audit two: TSMC is raising prices across three-quarters of its wafer revenue. In an exclusive in late June, former Bloomberg reporter Tim Culpan reported that TSMC is raising prices on more than 75% of its wafer revenue base, covering every advanced node at 7 nanometers and below, by 5-10%, with the increases already taking effect. Tom's Hardware's report on the same story puts the share at 74%. A Nikkei exclusive adds that even mature nodes will go up, by as much as 10%, effective January 2027, the first such increase in more than three years.

TrendForce's tracking gives the ceiling of current public reporting: up to 15% for 3 nanometers in the second half of 2026, then another 5-10% in 2027. For 2 nanometers the reported range is 10-20%; a rumor of 50% was knocked down by supply-chain sources TrendForce cited. There are also reports of a 10-15% surcharge for high-performance-computing customers that over-order.

This publication has to log something here. When the original call was written, its 'what would prove this wrong' test said the call would fail if a double-digit-or-larger price increase on advanced nodes appeared in 2027. Read literally, the 15% for 3 nanometers already trips it. That test was set too tight, and that is itself a finding: the call wrote "refuses multiples" as "refuses double digits."

Audit three: TSMC has never collected the full rent in the past either. The last global shortage, the 2021-22 chip crunch, works as a control experiment. TSMC's largest price increase of that cycle was about 10% on advanced nodes and up to 20% on mature nodes. The most aggressive move in TSMC's history stopped at a fifth.

Back in early 2024, Ben Thompson at Stratechery treated the question of why TSMC doesn't take higher margins as a puzzle to solve. Thompson's three reasons were a Taiwanese cost-plus pricing culture, the deep conservatism of a company that has survived many industry cycles, and the habit of a long-time challenger whose biggest fear is losing volume, with the arrival of Intel's foundry business as a further check. Pricing restraint, then, is not 2026 news; it is twenty years of the company's behavior. The news is that for the first time the shortage is large enough to put a measurable price on that habit.

Taken together, the three audits force a narrower claim. What TSMC forgoes is not price increases; it is memory-style rent, meaning three specific things TSMC says it will not do, or visibly does not do. One is multiple-fold price increases on the back of the shortage: TSMC's increases run 5-15%, and memory's run on a different order of magnitude, as the next section shows. Two is turning the shortage into money customer by customer, according to bargaining power: TSMC's increases are uniform list-price changes by process node. Three is tying customers into TSMC's own capital structure through equity: "So far, no." Once narrowed, "forgone rent" has a clear definition: the difference between the actual increase and the increase the market would bear. How much would the market bear? Wei named it: an 86% gross margin, the number Wei is jealous of.

The control group: memory is collecting exactly the rent TSMC says it won't

The comparison Wei braked on is public, number by number. Lined up, the figures show what Wei had in mind.

Prices. Start with contract prices. TrendForce's tracking shows server DRAM contract prices rising 13-18% quarter on quarter in the third quarter of 2026. A separate industry report puts the rise in mobile memory prices at 78% in the second quarter alone, as makers moved mobile supply over to fill the server and HBM gap; legacy parts rose even faster. HBM3e supply prices for 2026 went up by nearly 20%. Gartner estimates that the combined price increase for DRAM plus SSDs reaches 130% cumulatively by the end of 2026.

On TSMC's side, the comparable figures are 5-10% as the base and 15% at the single highest point. The two rulers differ. TSMC's numbers are one-time list-price changes across a year; memory's are quarterly increases, and several quarters compounded produce Gartner's cumulative 130%. Even with that difference in measurement allowed for, the two ways of turning one shortage into money are an order of magnitude apart.

Profits. For its latest quarter, fiscal third quarter 2026, Micron reported a non-GAAP gross margin of 84.9%, a company record, and guided to about 86% for the next quarter. That is where Wei's number comes from. SK hynix's second quarter showed revenue up 257% year on year, operating profit up 557% and an operating margin of 76%, all records. The same official release says average DRAM selling prices rose about 30% in the quarter and NAND about 50%.

TSMC's gross margin in the same quarter was 67.7%, and TSMC expects the 2-nanometer ramp to dilute next quarter's margin by 3-4 points. The revenue-growth comparison is harsher still: TSMC up 33.7%, SK hynix up 257%. Price is the main driver of SK hynix's growth. The volume part, the HBM ramp in particular, is not broken out in this piece, so this is a directional observation, not a precise attribution.

Choosing whom to charge. The memory makers do one more thing that TSMC's list-price system does not: they price by customer. The detail in TrendForce's server DRAM report is that large US cloud providers locked a ceiling on increases through long-term agreements, while customers without long-term agreements pay the full increase. Separately, NVIDIA has been reported to get supply terms far below market price because of its purchase volume. The weaker the buyer's bargaining power, the more the buyer pays. That is textbook price discrimination with monopoly rent, and TSMC's uniform list-price changes give up that layer by design.

The downstream ledgers are already paying this tax. Amazon raised its 2026 capital spending plan by about US$20 billion. The reporting attributes the increase to a single input cost: memory. The same reports estimate memory cost per server rising from about US$6,000–6,500 to US$15,500–16,500. SemiAnalysis estimates, as relayed by others, that memory's share of large cloud providers' capital spending rises from about 8% in 2023-24 to about 30% in 2026; the original is paywalled. In the move from NVIDIA's GB300 rack to the VR200 rack, the next generation of NVIDIA's AI server racks, SemiAnalysis puts the GPU cost increase at about 57% and the memory cost increase at about 435%. The consumer side is more direct. Gartner estimates that memory costs push PC prices up 17% and phone prices up 13%, and that global PC shipments fall 10.4% in 2026 as a result. Memory's rent is moving down every product line toward the end buyer.

And packaging? The call this piece is anchored on originally said cost increases were driven by memory and packaging. That second half needs fixing. The evidence that packaging is a bottleneck is very strong. Wei said packaging capacity is so tight it limits customers' growth. CoWoS is sold out through the end of 2026, with a supply-demand gap of about 20%, and orders are spilling over to UMC (expanding silicon-interposer capacity), ASE (its largest-ever fab construction plan) and Amkor.

The evidence for packaging rent is weak. This piece's search found no independent report of a price increase for the CoWoS packaging service itself. A claim that advanced-packaging average prices rise 10-20% a year comes only from low-credibility sources, and this piece does not use it. And most of CoWoS sits inside TSMC, so it is governed by the same pricing restraint. Packaging's correct place in this piece is therefore a volume bottleneck (waiting lists, lead times over a year, capped shipments), not a price rent. The tax packaging collects is paid in time, not in price. The price driver of rising AI hardware cost is memory, singular.

Outside the circle: the circular-financing contagion chain breaks at the foundry

That settles where the forgone rent goes. The second call this piece has to establish is about risk: if the equity-for-compute loop in today's AI industry goes wrong, TSMC is not on the contagion chain. This section draws the circle clearly, then lays out the evidence that TSMC is not in it.

What does the circle look like? Bloomberg built an interactive graphic this year titled AI Circular Deals: How Microsoft, OpenAI and Nvidia Pay Each Other. Economist Noah Smith, writing about that graphic, summarized the mechanism: many circular deals involve an "upstream company buying the stock of a downstream company, who then uses that cash to buy products from the upstream company." The names on the map: NVIDIA, OpenAI, AMD, Oracle, xAI, CoreWeave, Microsoft.

Most of the money and equity flows among these companies have been checked against primary documents. AMD gave OpenAI warrants for 160 million shares at an exercise price of US$0.01, tied to 6GW deployment milestones and equal to about a tenth of AMD, as an SEC 8-K shows; Meta signed a near-copy in February. NVIDIA's original letter of intent with OpenAI, US$100 billion of investment for 10GW, fell apart, and NVIDIA instead bought directly into OpenAI's funding round, at a reported US$30 billion. Oracle and OpenAI have a five-year, US$300 billion cloud contract. CoreWeave is the textbook case: NVIDIA invests in CoreWeave, and CoreWeave uses the money to buy NVIDIA chips. Even memory has a foot in the door. Samsung and SK hynix signed letters of intent with OpenAI's Stargate data-center program to supply up to 900,000 DRAM wafers a month, about 40% of global output. These are only letters of intent so far, and actual 2026 volumes are small.

TSMC is not in the circle, and the evidence comes in three layers. TSMC does not appear anywhere in Bloomberg's map. Noah Smith went through the map company by company, and "TSMC is not mentioned anywhere." TSMC was not discussed and ruled out; it was never on the board. Wei turned the arrangement down in public. That is the line quoted above, "So far, no, TSMC don't do this kind of financial arrangement," and, as shown earlier, the analyst was asking precisely about the circular template of investing in customers' customers. TSMC's history is consistent with that answer. In 2024 Sam Altman came with a grand plan for US$7 trillion and 36 fabs, and TSMC executives reportedly dismissed Altman privately as a "podcasting bro". Taiwanese commentary lists customer neutrality as one of the two foundations TSMC was built on. Even the US government treats TSMC differently: a White House official said the administration was not looking to take an equity stake in TSMC the way it did with Intel.

What this is worth for risk assessment has to be read against what is wrong with the circle. The structure carries an unresolved debate: Bill Gurley's view that circular investment hides the demand signal, against Jensen Huang's view that there is no overcapacity. The debate cannot be settled here, but both sides accept the mechanism. Once equity and orders interlock, an outside observer cannot tell real demand from demand propped up by insiders.

TSMC's absence gives the system an uncontaminated reading. TSMC takes cash for wafers, holds no customer equity and does not depend on customers buying TSMC's stock, so TSMC's orders are just orders. If any link in the circle breaks, say OpenAI's demand falls short or a neocloud's (a newer cloud provider that rents out AI compute) backstop gets triggered, contagion travels through equity and receivables. When it reaches the foundry it becomes plain order risk: a cancelled order is a cancelled order, with no equity writedown, no voided warrants and no chain of cross-holdings to revalue. If the circle catches fire, the fire reaches TSMC as lost orders, not through a chain of shared stakes.

That call needs an honest boundary. Being outside the equity circle is not the same as being outside the supply chain's risk. If AI demand collapses across the board, TSMC's advanced-node orders fall hard too. TSMC's 2026 capital spending, raised to US$60–64 billion, and TSMC's statement that capital spending over the next three years will be significantly higher than over the past three, are bets on AI demand lasting. "Outside the circle" in this piece means one precise thing: the contagion chain through the capital structure breaks at TSMC; the business cycle does not.

Why not collect? Three explanations that can all hold, with different predictions

TSMC knows the rent is there and does not collect it. That needs an explanation. Three explanations are tested against each other here. After the evidence, the conclusion is that all three are true, at different levels, and that they predict different futures. That difference is what lets the call be settled later.

Explanation one: rational play in a repeated game. This piece gives it the lead role. In a one-shot game, a monopolist should price right up to the point where the buyer barely survives. TSMC and its customers are in a repeated game with no last round. Collect the full rent today, and tomorrow customers put twice the effort into building up a second source, or move designs to Intel's foundry. In game terms, Wei's line that customers have to succeed means keeping customers alive and loyal is an asset worth more than one period's rent.

Welcoming EMIB works the same way. The packaging bottleneck chokes shipments of TSMC's own front-end wafers, so giving up a lower-margin step to protect the volume of the high-margin core business is arithmetic that adds up. Ben Thompson's three cultural factors (cost-plus tradition, scars from past cycles, the challenger's habits) are the historical sediment under this logic. This explanation predicts that the restraint is structural: no matter how big the shortage gets, multiple-fold price increases will not appear.

Explanation two: political insurance. This is true, and this year it gained a concrete exchange mechanism. On the same call, TSMC announced another US$100 billion for Arizona, bringing its total US investment to US$265 billion across twelve planned fabs. By the CFO's account, building in the US costs four to five times as much as in Taiwan. That seemingly bad deal paid off in January. The Section 232 action signed by the Trump administration put a 25% tariff on certain advanced semiconductors, but the US-Taiwan framework announced the same day gave companies building in the US tariff-free import quotas: 2.5 times planned capacity during construction and 1.5 times new capacity after completion. Taiwanese officials called it most-favored treatment with "zero tariffs within the quota."

Under that structure, gentle pricing for US customers and more US capacity are two sides of one thing: the forgone rent buys policy room. This explanation predicts that the restraint tracks geopolitical pressure. If the tariff framework shifts or political pressure rises, pricing softens further; if the political risk fades, the ceiling on price increases quietly rises.

Antitrust is a shadow version of this explanation. So far only think-tank-level advocacy pieces have called TSMC monopolist-like. A search of US Justice Department and Federal Trade Commission, Taiwan Fair Trade Commission and European Commission records turned up no formal case. That is background pressure, not an active constraint, and this piece downgrades it accordingly.

Explanation three: public restraint, private collection. Partly true, with an upper bound. The skeptic's version: Wei says no 4x on the call, but TSMC collects through three structural channels. One is mix moving upmarket, as the most expensive node, 2 nanometers, starts volume production, with 7 nanometers and below already 77% of wafer revenue. Another is packaging and over-order surcharges. The third is the established fact of a 19-point higher margin at the same utilization.

This explanation is true; audit one is its evidence. But it has a measurable upper bound: the margin bridge. If private collection were the main engine, price and mix would show up in the attribution of margin gains. Instead the CFO's bridge is cost improvement plus utilization, and next quarter's guidance cuts gross margin by 3-4 points for the 2-nanometer ramp. A company turning a shortage into money does not look like that. Private collection exists, but its size stays on the same order of magnitude as the 5-15% list-price increases. There is no second set of books.

All three explanations share one observable point: a behavioral ceiling. Whatever the motive, repeated-game logic, political insurance or image management, TSMC in 2026 has pinned its own room for price increases in a band: 5-10% as the base, 15% at a single point, 20% as the historical extreme. That band is the number downstream cost models actually need.

What this changes, and for whom

The three explanations agree on the ceiling, and the ceiling changes how three kinds of work get done.

• Companies that model AI compute costs: The three big items on an AI hardware bill of materials (logic wafers, memory, packaging) now carry three entirely different risk shapes. The wafer segment has a public behavioral ceiling (5-15%, from the supplier's own words and twenty years of behavior) and can be treated as close to a constant. The memory segment has no ceiling: cumulative increases are measured in multiples, Micron's gross margin of about 85% is still guided upward, and weaker buyers pay more under customer-by-customer pricing. Memory is the main open risk on the whole cost curve. Packaging's tax is priced in time (lead times, allocations), not in price. Our read: the leverage of purchase hedging and long-term contract negotiation sits entirely in the memory segment. That is this piece's direct answer to which segment to lock in. One more effect follows. NVIDIA, reportedly able to get below-market terms during the memory price surge, gains another layer of cost advantage. Second-tier buyers without long-term agreements pay the full increase.
• Firms assessing AI circular-financing risk: Every company on the circle map carries some demand propped up by insiders in its financials; only the foundry's order book is a clean reading. TSMC's advanced-node wafer starts and capital spending guidance are among the few demand thermometers in this cycle not contaminated by the equity loop. Cross-checking the circle's narratives against them is more reliable than relying on the backlog of any company inside the circle. In contagion models, the chain of revaluation through equity, warrants and cross-holdings can be cut off at the foundry.
• Analysts following memory as the mirror case: The mirror question to this piece is why memory dares to collect in full. Three makers form an oligopoly, the products are standard, customers have nowhere to go, and none of the three carries a customer-neutrality brand. In the same shortage, TSMC and the memory makers take opposite approaches to rent. That is a live lesson that market structure decides behavior: whether to collect rent is not a moral choice; it is a function of the game's structure. TSMC can afford not to collect, because customers can't leave and the volume stays with TSMC. The memory makers have every reason to collect in full, because a cyclical business has to fill in the next downturn's hole during the upturn.

Our call

The price driver of rising AI hardware cost is memory, not wafers. The most upstream logic foundry, facing the largest shortage it has ever acknowledged, has pinned its own price increases in a band from single digits to 15% (20% as the historical extreme), and in statements recorded in TSMC's official call transcript has refused multiple-fold price increases and circular-financing-style equity arrangements. The rent TSMC forgoes is being collected by memory, through cumulative contract-price increases of more than 100% and gross margins around 85% (Micron's latest quarter: 84.9%), and is passed down to end buyers through PCs, phones and cloud capital spending. Gartner estimates that combined DRAM and SSD prices rise 130% cumulatively by the end of 2026. The capital contagion chain of circular financing breaks at the foundry: TSMC takes cash, holds no customer equity and is not held by customers, so TSMC's order book is one of the few thermometers of AI demand in this cycle not contaminated by the equity loop.

What would prove this wrong:

1. Any single list-price increase of 20% or more on an advanced node (above every range in current public reporting; tracked through TSMC's earnings calls and TrendForce price reports, an observation window this publication set for itself and reviews each quarter). The behavioral ceiling gets revised, and explanation one gives way to explanation three.

2. A disclosed equity, warrant or similar arrangement between TSMC and an AI customer or that customer's end customers (verifiable through 6-K filings or earnings calls). The "outside the circle" call is void.

3. Credible reporting that TSMC has started pricing by customer bargaining power. The second pillar of "refuses memory-style rent" falls.

4. Memory contract prices falling for two consecutive quarters, with Micron's and SK hynix's margins falling markedly. "Memory is the driver" expires, and this piece becomes a historical record of the 2026 shortage cycle.

Written from the same research and judgments as the Traditional Chinese edition. Sources are linked; we distinguish original documents from reporting and mark what we could not verify.