🌐 Read this issue on the web

This issue arrived about 3 hours later than usual today — apologies for the delay.

This issue arrived about 4 hours later than usual today — apologies for the delay.

At a glance

1. For the first time, Nvidia's quarterly report adds up its own guarantees in a single statutory table: $108.5 billion. The $530 billion circulating in the market is the sum of three tables that measure different things.
2. To get the same gigawatt of compute built, the money Nvidia's three backstop tools put on its own balance sheet falls from $59 billion to $9.4 billion — and each cheaper step hands the protection to a different party: first the provider that keeps the option to resell, then the site's tenant, then mezzanine investors.
3. Oracle delivered 300,000 GPUs in a single quarter, and its finance chief says most of the new contracts behind that growth do not need Oracle's own money.
4. Two unrelated voices — a Princeton computer scientist and the chief executive of Box — say general-purpose agents have swallowed the task-specific harness layer; what is left standing is your own workflow yardstick, and that is what to build before buying an agent tool.

This issue draws on the research report written in the small hours of September 12; the material spans August 19 to September 11. Last night's sweep covered 228 pieces, of which 14 clickable receipts made it into this issue. This is the email edition; the full edition of this issue is the archive of record.

Today's main line

1. [Evidence update] (filed August 26) Nvidia's quarterly report adds up its guarantees in a statutory table for the first time — and the total the market is passing around is not in that document

Nvidia's quarterly report, filed August 26 for the quarter ended July 26, does something it has never done before: it rolls its guarantee exposure into a single statutory table, and the number is $108.5 billion. The table has two rows. The existing land, power and shell guarantees for AI clouds come to $3.5 billion, and a new guarantee signed in August 2026 for SB Energy, the energy developer owned by SoftBank Group, comes to $105 billion. A guarantee means "if my customer cannot pay you, I will." It sits off the balance sheet in normal times and counts only when something goes wrong.

The $530 billion figure being relayed around the market is not in this document. We pulled the original from the US Securities and Exchange Commission's database, converted it to plain text, some 170,000 characters, and searched for four terms one at a time. "Off-balance sheet": zero hits. "$530 billion": zero. "$184 billion", the prior-quarter baseline the outside analysis chose for itself: zero. "Take-or-pay", the clause that makes you pay whether or not you use the capacity: zero (Nvidia 10-Q, 2026-08-26). What the filing actually publishes is three tables built on three different definitions: "future commitments" of $366 billion, "additional commitments" of $56 billion, and "maximum gross exposure related to our guarantees" of $108.5 billion. Add them and you get $530.5 billion. So the problem is not the arithmetic — it is the label. The largest block inside that $366 billion is $279 billion of supply and capacity commitments, mostly memory, which is material Nvidia has agreed to buy for itself. Another $25 billion is equity investments it has agreed to make. Calling a purchase order and an equity commitment "guarantees" turns a procurement schedule into credit exposure.

There is a timing problem too. The notional amount of the land, power and shell guarantees moved from $3.530 billion on January 25 to $3.529 billion on July 26, essentially flat across the entire first half. The $105 billion was signed after the quarter closed; the filing itself says "signed in August 2026." So "contingent liabilities exploded this quarter" is not what the document says.

What is genuinely new in the filing is two sets of terms, and they are far more useful than any total. The first describes how the $105 billion comes into force: the guarantees take effect and step up as each of nine phases of data center construction is completed, the first of which is expected in fiscal year 2029, and each then steps down over a 20-year lease term. In the filing's words, the guarantees "are limited to defined portions of lease and power payments and not the full cost of the site or all of the tenant's obligations." In exchange, the site will exclusively host Nvidia AI infrastructure. The second describes the $36 billion of "AI cloud agreements," and it runs the opposite way from the common reading: the AI clouds "can unilaterally stop providing to us and sell to third-party customers at more advantageous rates," Nvidia's commitments "decrease as capacity is used by third-party customers," and Nvidia participates in the upside from those third-party sales only "if certain criteria are met."

Verification: this is Nvidia's own statutory quarterly report, backed by the certification provisions the regulator requires. We read the document itself rather than anyone's account of it, and every word-for-word passage above comes straight from the original. ⚠️ One thing we are not claiming: that anyone concealed anything or disclosed improperly. Every term discussed today was broken out by Nvidia itself in a statutory filing. ⚠️ One limit that still carries weight: the categories in that commitments table are the company's own choice. We can point out that the labels and the categories do not match; we cannot redraw the table.

Judgment update: The correct description is a third thing, and both sides get it wrong: residual demand absorption — Nvidia pays for the capacity nobody else wanted. When we recorded in July that Nvidia offers cloud providers take-or-pay minimum revenue guarantees, the basis was an industry analysis's framing, not a primary document. On August 20 we went looking for a primary document and found something odd: a counterparty, in a single statutory filing, described three other contracts as take-or-pay in plain words, and only for the Nvidia arrangement switched to "revenue share and credit support" (our August 20 issue). Today Nvidia's own statutory filing produces the same result: zero hits for take-or-pay anywhere in the text. Two different filers, two different documents, the same wording gap appearing once in each. That is enough to record as a pattern. ⇒ But flipping to "it is just credit support, not a payment obligation" is too weak. Nvidia puts it in its commitments table with a dollar amount and an annual schedule running past fiscal year 2032. That is a payment obligation scheduled on the books.

Investor note: the prevailing narrative assumes the notional total of the backstops can serve as a risk gauge. This evidence weakens that assumption, and the gap is precise: of the $530.5 billion most often quoted, only $108.5 billion is credit exposure. The rest is procurement and investment commitments Nvidia will pay for itself.

2. [Today] (published September 11) To get the same gigawatt built, Nvidia's three backstop tools book contingent-liability amounts on its own balance sheet that differ by a factor of 6.3 — and each cheaper step moves a layer of protection off Nvidia's books

SemiAnalysis, an independent research house covering semiconductors and compute, published a comparison on September 11. It takes Nvidia's three tools for supporting customers and converts each into the contingent liability Nvidia must carry on its own balance sheet to get one gigawatt of compute built. A contingent liability is the kind of guarantee described in item 1, the kind that only counts when something goes wrong. The three numbers are $59 billion, $25 billion and $9.4 billion, and the most expensive tier is 6.3 times the cheapest (SemiAnalysis, 2026-09-11). Read alongside main-line item 1: the numerators in these tiers are the figures on that statutory table.

Two of the three tiers we reconciled ourselves against primary figures. The cheapest one cannot be reconciled.

For the two reconcilable tiers the arithmetic is in hand. $105 billion divided by 4.25 gigawatts gives $24.7 billion. One announced case under the floor-rent program is in Batam, Indonesia: 360 megawatts, or 0.36 gigawatts, at a deal value of $21.1 billion, which works out to $58.6 billion per gigawatt. Both sit close to the round numbers in the table without matching them digit for digit. The third tier's $9.4 billion cannot be checked, because that program's details are not public and the figure is the research house's own estimate, reasoned by analogy from another chipmaker's structure. Mezzanine investors are the layer of the capital stack that ranks behind bank debt and ahead of equity; "first loss falls on the mezzanine" means they get hit before the banks do. ⚠️ This research house also sells an analytical model on this very topic and due-diligence services to lenders, and is hiring a credit team. It has a direct commercial stake in the story that financing structure is the central issue. That does not make it wrong. It means anyone citing it has to check the work themselves.

Verification: the tier framework is the research house's. The numerator and denominator for the two reconcilable tiers come from the original statutory filing and from announced project capacity respectively, and we did the division ourselves. ⚠️ We read only the passage in front of the paywall; we did not see the estimates for the remaining items. ⚠️ The chronological order of the three tiers runs from expensive to cheap, but the causal claim that Nvidia switched tools because they were cheaper rests on one relayed account and nothing stronger.

Judgment update: the pressure shaping seller backstops is not "how much to backstop" but the balance-sheet price per gigawatt. Once that ruler becomes the selection criterion, risk gets pushed outward systematically. Two observable consequences will arrive together: the notional total of the backstops keeps rising while the thickness of protection per unit falls. When both happen at once, "total backstop" — the most-quoted monitoring variable — misleads in the optimistic direction.

⚠️ Three pieces of counter-evidence we are keeping on purpose. First, Broadcom moved the opposite way in the same period: its protection runs on two legs, a residual-value shortfall guarantee of up to $29 billion plus up to $42 billion of convertible notes covering a customer's rent, which is far thicker (Broadcom 10-Q, 2026-09-10; we covered that thread in our September 11 issue, and today it serves only as the counter-anchor). ⇒ Per-gigawatt efficiency is not an industry law but Nvidia's choice, and the divergence between the two companies is itself an experiment with a checkable answer. Second, the research house's own caveat: the implicit backstop from the hyperscalers is more than 35 gigawatts of third-party leases by 2028, far larger than Nvidia's roughly 6.5 gigawatts, so even if our call here is entirely right, it may not be the largest component of systemic risk. Third, the third tier cannot be computed.

⚠️ What would prove this wrong: four tests, any one of which overturns the call. One, Nvidia's next support program carries a higher contingent liability per gigawatt and explicitly folds the data center rent leg into the guarantee. Two, the floor-rent program resumes, meaning Nvidia officially denies a pause or restarts signing. Three, the formal filing for the residual-value program shows undisclosed rent support beyond the 25%. Four, lenders' pricing does not reflect thinner protection. Verdict date: December 31, 2026, an observation window we set ourselves that does not correspond to any party's scheduled filing.

Investor note: conventional wisdom treats a bigger backstop as bigger risk. Today's comparison rewrites that assumption into a harder shape to read: scale and thickness are moving in opposite directions, so as the total grows, the share landing on Nvidia may shrink while the share landing on lenders grows.

3. [Today] (earnings call September 10) Oracle delivered 300,000 GPUs in one quarter, and its finance chief says most of the new contracts behind that growth do not need Oracle's own money

On Oracle's September 10 earnings call, chief financial officer Hilary Maxson and co-chief executive Clay Magouyrk disclosed a set of figures. The company delivered 300,000 GPUs in the quarter. Of the $26 billion added to remaining performance obligations last quarter, Maxson said the "vast majority" of contracts were "prepay or bring your own hardware or similar mechanic, so will not require incremental capital from Oracle" (Data Center Dynamics, 2026-09-11). Remaining performance obligations are contracts already signed that have not yet turned into revenue, the official measure of a cloud vendor's order book. Magouyrk listed three mechanisms for not spending the company's own money: financing arrangements with suppliers that let Oracle, in his words, "pay for the capacity as the customers pay us"; customers buying their own hardware and renting only the building and the operations; and customer prepayment. Read alongside main-line item 2: there, a chipmaker uses ever-thinner credit instruments to get someone else to pay; here, a cloud vendor uses prepayment and bring-your-own hardware to get someone else to pay. Entirely different tools, exactly the same direction.

⚠️ Two numbers need their denominators stated separately. The executive said the 300,000 GPUs were "almost three times what we delivered in all of Q4, and 73 percent of the total capacity delivered last fiscal year." The first denominator is a quarter, the second a fiscal year, and this cannot be written as "tripled." Likewise, the 121% year-over-year growth belongs to cloud infrastructure revenue, not total revenue. Oracle's capital spending guidance for fiscal year 2027 is $90 billion to $95 billion. The AWS figure the article sets beside it, about $220 billion, includes the retail business, which is not a like-for-like cloud number; Microsoft's for the same period is about $175 billion.

Verification: everything here is verbal disclosure on an earnings call, relayed by industry media, and we have not checked it against a primary document. The upgrade path is clear: once Oracle files its quarterly report for the first quarter of fiscal 2027, every figure can be checked directly. ⚠️ The same call disclosed a data point on the residual-value side: utilization was 97.9%, and the GPUs that came up for renewal in the quarter were "renewed or resold at a 20 percent premium to prior contracts," most of them four years old or older. This cuts against our own main line, which is why we keep it: residual value is the load-bearing assumption behind two of the guarantee tiers in item 2's table. ⚠️ But it cannot serve as a quantitative indicator: the call did not say whether three thousand cards or three hundred thousand came up for renewal, and a renewal premium on a small population at high utilization is very likely a selection effect, with the popular models renewing first.

Judgment update: our September 11 issue covered Oracle funding its operating cash flow with customer prepayments. What is new today is the delivery volume and the complete mechanism behind "will not require incremental capital." The same pressure to push capital needs off one's own balance sheet is showing up in an entirely different role with entirely different tools. ⇒ Pushing risk outward is not one company's strategy; it is the general shape of this build-out, and Oracle's capital spending guidance, far below its peers, is that shape's observable consequence.

Investor note: investors have been reading a cloud vendor's capital spending guidance as a proxy for how much risk it carries. The earnings call weakens that assumption: cloud infrastructure revenue is growing 121% while the spending guidance is half a peer's, and the difference lies not in efficiency but in whose books the spending sits on.

4. [This quarter] (events August 21 to 23) General-purpose agents swallowed the layer that "decides the process for you": the layer left behind is your own yardstick

Within three days, two unrelated people pointed at the same structure.

The first is an academic. Arvind Narayanan is a professor of computer science at Princeton and co-author of AI as Normal Technology, known for years of arguing that the scale-solves-everything view is over-applied, and holds no product interest. On August 21 he volunteered an example that cuts against his own position: a year ago, "Deep Research" tools that run multiple rounds of search and produce a report seemed astonishing, and that kind of task-specific harness was swallowed whole by general-purpose agents — AI systems that plan their own steps, call tools and act in sequence — within a year. A harness is the framework wrapped around a model that decides what to search first, how many rounds to run and how to synthesize. His reason for the general-purpose win is not that the general model is stronger, but that you can instruct it your own way. In his words: "I can get them to do my searching for me without trying to do my thinking for me" (Arvind Narayanan, 2026-08-21).

The second sells software. Aaron Levie is co-founder and chief executive of Box, the enterprise content management company. On August 23 he argued that the bottleneck on AI diffusion is the yardstick. He split the yardstick into two layers and said the industry has been watching the wrong one: the public evaluations that accompany every model release are useful but only show the shape of general progress, while the far bigger space is evaluations of every major workflow an enterprise runs, down to the specifics of a single company. His reason fits in one sentence: "you can't automate what you can't assess the progress on" (Aaron Levie, 2026-08-23).

Verification: both are single-source, unmeasured, one example each. ⚠️ Narayanan's "about a year" is recollection, not statistics, and he explicitly concedes only the core truth of the principle without withdrawing his standing criticism. Do not read this as a reversal. ⚠️ Levie's post is vendor narrative of the highest order: Box sells exactly the software that wires AI into enterprise workflows, and "enterprises need workflow-level yardsticks" is the one-sentence version of why its product exists.

Judgment update: connecting the two yields not the conclusion that agents are powerful, but the finding that value is moving from "deciding the process for you" to "measuring whether you did it right." The boundary between those layers is clean: who holds the input. The process can be written into a general-purpose model. Your company's acceptance criteria cannot, because scale cannot buy your private inputs. ⇒ The implication is a reversed order of operations: the bet right now is not on buying an agent tool. It is on building your own yardstick first. ⚠️ Ceiling on tone: we have noticed a shape, not measured one. ⚠️ And this call carries the mechanism that would kill it: work traces are exactly what you need to generate enterprise-grade yardsticks automatically. If harness vendors use accumulated traces to grow workflow yardsticks for their customers, then the "layer that cannot be swallowed" gets swallowed by the same mechanism, just one step later.

Investor note: most investors watch model capability as the bottleneck on enterprise AI adoption. These two posts move that bottleneck onto the buyer, and the gap is this: the money the buyer has to spend is not on the procurement end but on the acceptance criteria it has not yet built.

Also happened — not verified by us yet

1. [This quarter] (event August 19) Local opposition to US data centers: for the first time we saw a technology reporter counting "how many projects were actually delayed or blocked" rather than how many building moratoriums passed. ⚠️ We are not citing the number he gave, because whether it counts projects or sites, what threshold counts as a delay, which three months, and whether the population is the US or the world are all unknown — and we do not have the chart that would carry those definitions. The one thing worth remembering: opposition is already producing project-level consequences, magnitude unverified (Brody Ford, 2026-08-19).
2. [This quarter] (event August 21) Vercel, the front-end deployment platform, launched is-agentic.com, turning "can AI agents read your website" into a scored audit category, which amounts to porting twenty years of search-engine tooling over to agents. Its chief executive says they ran their own scorer against their own site in a loop until it hit a perfect score. ⚠️ Scoring yourself full marks on your own ruler only proves the site fits the ruler, and that ruler is the thing they are selling to everyone else (Guillermo Rauch, 2026-08-21).

Chips & semiconductors

[This quarter] (event August 20) A hard-sounding first-principles challenge, checked line by line against material we already hold: the most categorical sentence does not survive. An inference chip is a custom part that handles only the stage after training, when a model actually answers questions, giving up generality for efficiency. On August 20 an account posted a three-part challenge to Etched, a startup in that lane: go the high-bandwidth-memory route and you cannot beat GPUs and TPUs; go the on-chip static memory route and you cannot beat Groq and Cerebras; and, in his words, "Low-voltage inference is BS: you need energy to move data; this is physics" (Bing Xu on X, 2026-08-20). ⚠️ We have not verified who is behind this account. This is our first time recording it. Do not treat it as an insider because it sounds like one. What carries the weight is two pieces of material we already hold, not him. The third sentence does not hold: moving data does cost energy, but that is a floor, not a constant, and pushing the floor down is exactly what the industry is doing. d-Matrix, another inference chipmaker, says its 3D-stacked memory test chip reaches roughly 0.4 picojoules per bit in the worst case against 3 to 4 picojoules for high-bandwidth memory, an order of magnitude apart (d-Matrix blog, 2026-03-16; ⚠️ a vendor's first-party claim, not independently measured). The second sentence is actually strengthened, in a direction he may not have counted on: at its March 2026 conference Nvidia unveiled a next-generation inference chip integrating Groq's technology, with 500 MB of static memory per chip and 1.2 petaflops at FP8 (More Than Moore, 2026-03-16). ⇒ The incumbent on the static-memory route is no longer just Groq, it is Nvidia, and a startup on that route is in a worse spot than he described.

Named commentary

[This quarter] (retrospective, originally posted August 21) A viral AI-layoffs number was recomputed by a named person on the day it peaked and cut by nearly two-thirds — and we accept neither figure. This column has no new named view this week; the pick is an older item that matters. On August 21 a post went viral claiming that American tech companies had laid off 170,000 people in the past year "with the explicit claim that these cuts are due to AI," against 1,547 new data center jobs (original post, 2026-08-21). The same day Susan Zhang, who previously led training of Meta's OPT-175B large model, took it apart in two moves: the number's source website is tied to a crypto token, and her own recount came to about 50,000 (Susan Zhang, 2026-08-21). ⚠️ Her recount method was "having two AI agents fix the numbers": not reproducible, no public procedure, no third-party check, so as evidence it is no stronger than the 170,000, and she did not claim precision herself. ⚠️ There is also a problem neither side raised: 170,000 is a year of layoffs across the entire national tech sector, and 1,547 is data center jobs in the operating phase, excluding construction. Dividing one by the other never meant anything. Which of our calls it supports or rebuts: it runs in the same direction as an existing record of ours: that attributing layoffs to AI is a systematically exaggerated phenomenon. The strongest evidence for that record is New York State's mass-layoff filings: more than 160 filings in the first year, with the share ticking the AI box close to zero. ⚠️ But the two differ by an order of magnitude in evidentiary strength, so this item cannot be used to vouch for that one. What is actually useful to a reader is not a better number but two checks you can run: who maintains the data source, and how far the correction travels. Here the correction reached about one forty-eighth of the original post's audience.

Model watch

No model watch item this issue. Nothing new landed in this section, and we would rather leave the column empty than pass off an evergreen idea as this week's news.

Product moves

[Today] (published September 11) A compute cloud made its case to lenders in public that "renting GPUs" and "an AI cloud platform" are two different categories; it timed that case right after Oracle said its customers are bringing their own hardware. CoreWeave's official blog post on September 11 opens with the sentence "I spend a lot of time talking with the investors financing the AI buildout." It goes on to argue that renting hardware alone leaves the customer handling software and operations itself, and that orchestration, observability (the monitoring that shows where a system is failing), security, developer tools and expert teams are what turn a rack of chips into a usable environment. It offers its customer mix as support, naming Jane Street, Zonos and Mercado Libre, and adds its own caveat: "Customer concentration can have many causes, particularly early in a company's growth, so diversification alone is not proof of platform depth" (CoreWeave, 2026-09-11). Why it belongs next to main-line item 3: Oracle has just said most of its new contracts are customers buying their own hardware and renting only the building and operations, and this post's entire argument is that such a business is a commodity. ⚠️ This is the company's own post, addressed to its financiers, and we have no third-party data to check it against.

From the archive

No archive pick this issue. The reusable older material in our own back catalog is exhausted.

Sources & accounting

The past 24 hours. September 11 to 12 added 228 pieces. We finished 2 today and ruled out 3, leaving 223 unread — 0.88% read. Those 223 are unread, not read and rejected, and the two are not the same thing. By category: industry newsletters, 7 arrived, 1 read and 2 ruled out; company and personal blogs, 42 arrived, 1 read and 1 ruled out; those two are the only categories anyone actually opened. X posts 121, papers 50, academic papers 4, macroeconomic data 2, industry analysis 1, company filings 1: all at zero. ⚠️ The heaviest statutory document in today's body is not among those 228: after finishing that one newsletter, we went to the Securities and Exchange Commission's database ourselves and pulled the original. The 0.88% above cannot measure the act of going after a primary document, and today's weightiest main-line item is the product of exactly that act. The named part: the one newsletter read to the end was SemiAnalysis; on the X side, the highest-volume accounts were @teortaxesTex at 65 posts, @bhorowitz at 37 and @GaryMarcus at 34.

There was a second set of material last night, and it is not inside that 228. Everything finished last night in that set was X posts from August 18 to 23, older material we are catching up on from August. Main-line item 4, both items in the not-yet-verified section, and one item each in the chips and named-commentary columns all have event dates between August 19 and 23. They are not things that happened today, which is why each carries its event date beside its tag. We publish them because readers have never seen them. We do not present them as today's news, because that would be false freshness.

Older material added back in one pass. We added 0 new sources in one pass today, and that 0 means the long-term roster gained no new names; it is a different number from the 14 external receipts used in the body. Separately, a set of older material from outside the past 24 hours came back in, with event dates between July 1 and August 30, 4,943 pieces in all, dominated by 1,808 academic papers and 891 industry newsletters. That set and last night's list are two different populations, so the same category reads differently in the two places.

Source concentration. Today's flag is that one document carries too much weight: all of main-line item 1 and its judgment update, plus the two numerators in item 2, rest on Nvidia's single quarterly report. If our reading of that document is wrong, all of it falls together. Mitigating factors: it is a primary document the company filed itself under statutory certification, every word-for-word passage comes from the original, and we deliberately kept Broadcom's opposite-direction anchor. The second flag: main-line item 4, both not-yet-verified items, the chips column and the named-commentary column, six pieces of material in all, are each a single post. That is an artifact of our catching up on August material, not a sign that today's world mainly happened on X. Where the independent second view sits: the four main-line items rest respectively on an original statutory filing, an independent research house, an industry outlet's account of a named earnings call, and two unrelated named speakers.

What you are not getting today. Three things. One, we did not read past that newsletter's paywall, so for the third tier's $9.4 billion per gigawatt the body can only say "does not reconcile," not where it goes wrong. Two, we have not read Oracle's quarterly report for this quarter, so every figure in main-line item 3 stops at "a named executive said so." Three, we do not have the chart the reporter in the first not-yet-verified item attached, and the four definitions it would settle are very likely written on it.

The sources we track. 529 named speakers in total. The spread: X 302, podcasts 90, outlets 51, blogs 48, paper authors 48, newsletters 46, earnings calls 26, keynotes 23 and a scattering of others. ⚠️ Those count venues, and one person can occupy several, so the parts sum to more than 529. Representative names: on X, Arvind Narayanan, Aaron Levie and Susan Zhang; among blogs, Lilian Weng, Terence Tao and Dario Amodei; among paper authors, Ion Stoica, John Jumper and Boaz Barak; among newsletters, Dylan Patel, Ben Thompson and Zvi Mowshowitz; on the institutional side, SemiAnalysis, Data Center Dynamics and More Than Moore. Several identically named numbers count different populations. The roster's 302 X accounts are the total we watch over time; last night's sweep actually touched 374 accounts and pulled 537 native posts, of which 121 reached today's material, and not one was read to the end — four counts of four different things. Likewise the roster's 46 newsletters are the long-term total, while only 7 arrived last night and 1 was read, and a further 891 are back-fill from early July, three separate populations again.

I finished today's issue / I didn't finish

This is not a news digest: we hunt each day's AI firehose for the insights that actually matter and the practitioner judgments worth tracking over time, and we show how every item was verified — the point is always "which judgment got harder, and who's been right," never "what happened today."

— SecondSource · generated by our research system · 14 sources · Got a view? Reply and tell us

Written from the same research and judgments as the Traditional Chinese edition; every claim links to a primary document.

SecondSource publishes industry analysis, not investment advice. We do not evaluate, rate, or recommend any specific security, and nothing here should be treated as financial guidance — verify independently and use your own judgment.

EN English edition|繁 中文版 Traditional Chinese →