Daily Brief SecondSource Morning Brief · August 18, 2026 · Aug 18, 2026
This issue rests on the internal research digest compiled in the early hours of August 18. The material's events fall between July 13 and August 17, and the lead item's thread starts back in January. Overnight we went through 52 long-form pieces and 439 posts → 36 clickable receipts here; the full accounting sits at the end. This is the email edition; the full edition of this issue is the website archive of record.
The core judgment. In January, Mark Papermaster, chief technology officer of AMD — the second-largest supplier of AI accelerator silicon in the US — was asked when the copper cabling inside a rack gets replaced by fibre. His answer was a rolling three-year window: within the next three years, starting with the largest clusters. The flaw in that formulation is that it never comes due. Say "within three years" in January, say it again in July, and a supply-chain planner still cannot make a single buying decision from it. The July 23 launch pinned the window down. The official press release puts the next-generation MI500 series of GPUs in 2027, the launch floor confirmed that the MI500 generation supports "copper and optical," and the Helios rack actually announced stays all-copper (SDxCentral, reporting from the floor). This is the first time a GPU vendor has written optical adoption into a named product generation. NVIDIA's equivalent marker sits in 2028: its own technical blog still describes next-generation in-rack interconnect as copper, with rack-to-rack optics one generation further out.
Why we dug now: our August 16 issue, the one that asked which part of an AI hardware bill is actually getting more expensive, mentioned this pin in a single line in its Also happened section. What is new today is a ruling on it against a sentence pointing the other way. A June report from the research shop SemiAnalysis, which the market took seriously, holds that co-packaged optics — the technique of packaging light engines directly beside the chip — is deferred at scale by yield problems and will not mature until 2028 or 2029 (the publisher's home page; paid subscription, so we take direction only). On the surface the two are two years apart. One of them has to be wrong.
Open ?
Verification: this pin comes in four layers of commitment strength, and each supports a completely different decision. The year is written down — it is in the press release in black and white. The optics is spoken — the word "optical" appears nowhere in the full text of that release; optics shows up only on the keynote roadmap, relayed consistently by three outlets in the room, SDxCentral, StorageReview and ServeTheHome. The form factor was explicitly refused — asked whether this means co-packaged optics, in-rack optical interconnect or pluggable modules, Krishna Doddapaneni, the senior vice-president running AMD's networking unit, called it a matter for the future and declined to share more. The supplier line is blank. Two days ago this judgment still sat parked behind four objections, and all four are now resolved: the year is confirmed by the company itself; the original single relay from one analyst's post at the event is superseded by the press release plus three outlets present; the splice between January's words and July's is no longer needed, because the roadmap carries optics on its own; and the mystery of form factor is reclassified from "we don't know" to "AMD is holding it back."
Judgment update: the two-year gap is a false contradiction, and what exposes it is the denominator. The deferral describes mass commercialisation of co-packaged optics in open-market switches — that requires thirty-two light engines in one switch, an industry standard, and a second source. What AMD pinned is an optical option inside its own generation — single-vendor, vertically integrated, its own choice of form factor, its own choice of supplier — and it can ship in only the largest configuration. Two populations. One sentence about "optics being deferred" does not cover both. A second piece of consensus collapses alongside it. We had been using "copper wins near term" as the attack on this pin, and it only half survives. Copper does win this generation, but the 78-layer copper midplane board NVIDIA planned to use in place of cabling in its next-generation rack has been judged unmanufacturable at reliable yield, pushing the whole system to 2028 (Tom's Hardware, relaying SemiAnalysis; not settled). At the densest layer of all, GPU to GPU inside the rack, the real picture is that optics is late and copper is late too. Internally we promote this from candidate to full judgment, and raise the confidence score from 0.55 to 0.65. The scale tops out at 1, and 0.65 means the evidence stands up but has not reached the point of no reservations (how we score).
Investor note: the market currently prices the optical-interconnect timeline as a direction that will arrive sooner or later. This evidence converts it into two dated, expiring markers — 2027 and 2028. That weakens the assumption that "optics is deferred, so anything attached to it can wait." But until the form factor is disclosed, it also supports no specific name in the supply chain. Those two things have to be read together.
What would prove this wrong: if optics is absent when MI500 formally launches, or slips to the generation after, the pin is falsified. And if the optics turns out to be a co-packaged design drawing on an open supply chain, the "two different populations, so no contradiction" reading is void and the two sides become a real contradiction. Verdict date: the nearest marker is the fourth quarter of 2026, the deadline OCP — the cloud-hardware industry consortium — set itself for a first silicon-photonics specification, after twenty companies signed on August 13, though an OCP specification and an industry-wide standard being settled are two different bars; the heaviest one is 2027. The procurement side produced its first round of reconciliation readings today: laser shipments scheduled for the first half of 2027, first shipments of in-rack optics for the second half of 2027, and first revenue guidance from an optical-interconnect acquisition also landing in the second half of 2027 — all lining up neatly on 2027 (the Taiwan-based market-research firm TrendForce's assessment). ⚠️ But those procurement chains run around NVIDIA. Timing that lines up is not a customer list that lines up.
The filing NVIDIA made to the US Securities and Exchange Commission on August 17 discloses that it signed multiple residual value guaranties with the energy developer SB Energy, covering leases for roughly 4.25 gigawatts of IT load at the "PORTS campus" in Pike County, Ohio. The cumulative payment obligation is capped at US$105 billion. The tenant is OpenAI, on a 20-year lease. SB Energy is SoftBank's energy development arm and the lessor here, on a build-own-operate model. A residual value guaranty is an instrument that guarantees what an asset will be worth at some future date; when one is triggered, the guarantor pays the gap between the guaranteed value and what the asset actually fetches, not the whole value of the asset. Four details in this contract decide how you should read it. There are exactly two triggers: OpenAI's insolvency causing a lease default, or OpenAI failing to pay. What gets paid is only the shortfall, because the lessor has to try to re-let and then try to sell before any remaining hole becomes NVIDIA's. Whatever NVIDIA does pay, OpenAI pays back; the filing states plainly that OpenAI has agreed to reimburse NVIDIA for any amounts NVIDIA actually pays. Termination comes in four conditions, and one of them is unusual: the obligation ends when "OpenAI achiev[es] a satisfactory credit rating" (the other three being the 20th anniversary of lease commencement, OpenAI terminating the lease under its own terms, and other customary events). What NVIDIA signed is a contract whose exit condition is its customer growing up into investment grade. The same day's press release adds three more figures: NVIDIA invests US$1.5 billion in SB Energy, SoftBank and SB Energy will build at least 10 gigawatts of new generation and put at least US$4.2 billion into the regional grid, and OpenAI adds US$40 million to the community fund. None of that sits inside the US$105 billion. It is separate money.
Verification: three things need to be kept apart. First, US$105 billion is a cap, not an expected loss. Nobody can currently calculate the gap between the two, because the actual schedule of guaranteed minimum values is pending an exhibit to the next quarterly report. Second, there are three different capacity denominators, and mixing them turns an option into a commitment: 4.25 gigawatts is signed, roughly 3.8 gigawatts is an option NVIDIA may support at its own discretion, and 8 gigawatts is the press release's full-scope number. Yesterday we could not derive that 8. Today we can: chief executive Jensen Huang's blog post that same day states the remaining capacity is 3.75 gigawatts, and 4.25 plus 3.75 comes to exactly 8.00 (the filing writes 3.8 where the blog writes 3.75; we go with the filing). Data Center Dynamics, the specialist trade publication, wrote the same day that the 10 gigawatts of generation — 9.2 gigawatts of it natural gas — can support up to 8 gigawatts of total IT load (the report, 08-17). Third, there is a structural objection to this whole framing, and we accept half of it. The technology outlet The Next Web argued that day that "Nvidia is not backing OpenAI's rent. It is backing the buildings." (the piece, 08-17). We take half and reject half: the object of the guarantee really is the asset, and the filing's definition of the shortfall supports that reading — but the "nothing to do with the tenant" half is contradicted by the filing itself. The start key is hard-wired to OpenAI's insolvency or non-payment, and the off switch is hard-wired to OpenAI earning a rating. The accurate formulation is the object is a thing, the trigger is a person. ⚠️ And there is a fourth layer we only read today. Huang's own account is broader than the filing. He writes that the support is "limited to defined portions of lease and power payments, along with a specified residual-value commitment." If that holds, then "this does not cover rent" is true only of the one arrangement the filing discloses, and only half true inside his own narrative. Settling that needs the quarterly exhibit.
Judgment update: the market talks about "NVIDIA backstopping OpenAI again" as one thing. Today we split it into two contracts with opposite structures. One is a guarantee against demand — it guarantees minimum revenue, the object is generic capacity, and the trigger is the market failing to rent it. The other is a guarantee against an object — this case, where the object is the residual value of specific buildings and the trigger is a credit event at a single tenant and the failure of both re-letting and sale. Their risk even points in opposite directions. Paying out on the first means the whole market went cold. Paying out on the second means one customer went under and nobody wanted that campus. The second can happen while the market is red hot. So the watch list needs different entries. Tracking "total backstop committed" carries little information, because that is a ceiling. What to track instead is, first, OpenAI's credit-rating trajectory, because the contract writes that in as its own exit key; and second, comparable re-lettings and resales of large AI campuses around 2028, because the size of the shortfall is set by how deep that market is. One box is usable at the negotiating table straight away: ask your counterparty whether their backstop is against demand or against an object, because the cash-flow timing under stress is completely different. And if you are the tenant, a guarantee of this kind ends when you get your own rating, which means the guarantor has an incentive to push you towards getting one. That is leverage, not friendly advice. ⚠️ This is our own candidate judgment rather than anyone's quoted view, with an internal confidence score of 0.6 out of 1 — the direction holds up but it is not settled, because we have exactly one case (how we score).
Investor note: the prevailing narrative reads every guarantee of this kind as a chip vendor underwriting its customer's demand. This evidence says the one in the filing guarantees an asset's residual value and triggers on a single customer's credit. That weakens the assumption that backstop size is a proxy for demand strength, and strengthens the reading that the clause-level detail of a contingent liability is where the real signal lives.
What would prove this wrong: if the schedule of guaranteed minimum values in the quarterly exhibit shows real exposure far below the cap and tightly coupled to rent, the economic substance of the two instruments collapses into one — and Huang's "lease and power payments" phrasing is already a first weak signal in that direction. Or if a second residual value guaranty appears whose triggers are not tied to a single tenant, "the trigger is a person" stops generalising. Verdict date: the contract exhibit to NVIDIA's next quarterly report.
Read this beside item 2. Also on August 17, AMD closed a US$4.75 billion public offering of senior unsecured notes in four tranches: US$1.25 billion at a 4.600% coupon due 2029; US$1.5 billion at 5.000% due 2031; US$1 billion at 5.250% due 2033; and US$1 billion at 5.500% due 2036 (the filing). "Senior unsecured" means the notes rank ahead of subordinated debt in a wind-up but carry no specific asset as collateral; they rest on the issuer's overall credit. Being able to issue unsecured paper at all is itself evidence of credit standing. The stated use of proceeds goes no further than "general corporate purposes, which may include the repayment of debt."
Verification: our second reason for taking this item is that it works as a methodological control for item 2. Two large capital moves in the same week are easily folded into the same "AI capital frenzy" story, but the two documents state their purpose with completely different force. NVIDIA's spells out the object, the tenant, the triggers and the termination conditions. AMD's says general corporate purposes. We hold one more comparison sample on the graph: Oracle's roughly US$22.5 billion to US$25 billion investment-grade corporate bond earlier this year, where the source states plainly that the proceeds fund AI cloud infrastructure. Two debt raises, one specific and one not, cannot be written up as the same event. What a filing does not say, we do not say for it.
Judgment update: there is exactly one judgment this money supports today, and it is the useful kind. The curve from 4.60% at three years to 5.50% at ten is the price at which the investment-grade market is willing to lend to an AI accelerator vendor right now. That is not AMD's weighted average cost of capital, and it is not the rate on its existing debt. It is this week's market quote. The control sits right beside it: the compute providers who have to pledge GPUs as collateral borrow far more expensively over the same tenors. How much of this money reaches an AI production line is a question for the cash-flow statement in the next quarterly report, and this document cannot answer it.
Investor note: the narrative currently reads borrowing by any AI company as a signal of capacity expansion. This evidence says that two filings from one week state their purpose very differently, and that the blank is itself information. The proxy that gets weakened here is issuance size: how much an AI company borrows says nothing about how much of it is AI.
1. [This week] (event date 08-16) The payments company Stripe is reported to have agreed to buy OpenRouter for more than US$7 billion. OpenRouter is a routing shop that gives developers one interface onto more than 400 AI models; three months ago it closed a US$113 million round at a US$1.3 billion post-money valuation. The number worth chasing is not the multiple, it is the position. The routing layer sees who is using which model, how much of it, and what they paid — exactly the metering point a payments company wants. ⚠️ Stripe told reporters it does not comment on rumour or speculation, and there is still no first-party announcement; the eight outlets that followed all trace back to the same exclusive, so they are not an independent second source. One more box is worth holding in mind: our August 9 issue measured how the same open-weight model scores differently across third-party endpoints. The value of a routing layer rests on picking endpoints for developers, and the spread in endpoint quality is both the source of that value and its risk (Bloomberg's original report, 08-16).
2. [This week] (event date 08-17) Market commentary from the research shop SemiAnalysis says Google is working with AMD on its next in-house AI chip, TPU v10. If it holds, this is AMD's first substantive involvement in someone else's custom AI silicon project, which puts it on Broadcom's and Marvell's turf. The same commentary cuts its output estimate for the current TPU generation, from 3.2 million units to 2.7 million, blaming a difficult ramp in advanced packaging capacity. ⚠️ The wording in the original is "market chatter suggests," it reaches us through a media relay, there is no independent second source, and we are not admitting it to the graph today (Data Center Dynamics, 08-17; the original post).
3. [This week] (event date 08-17) The venture capitalist Tomasz Tunguz sets out a framework that would invert the economics of inference: every model today freezes on the day its training ends, whereas "test-time training" keeps changing the weights while you use it. Once a model updates on your prompt it is no longer the model that answered your neighbour, so the batch economies of serving a million people from one set of weights disappear and a provider needs a copy in flight per user. ⚠️ This is trend commentary rather than new evidence; the speed multiples in the piece all sit in footnotes citing third-party research we did not chase to source; and the author is an investor, so "more chips" points the same way as his book (the piece, 08-17).
4. [Evidence update] Item 2 of the main line in our August 13 issue — the one on Claude Code's permission prompts drawing a 97% approval click — carried a post from Arvind Narayanan, the Princeton computer science professor and co-author of AI Snake Oil: AI programs that can run multi-step tasks on their own are splitting into a delegation type and a collaboration type, and what limits delegation is how cheaply you can verify, not what the model can do. Today we established that the post restates a keynote he gave at an academic conference on July 13, rather than being a new statement. The clock on that call moves back a month, and we set the check date at October 13. The keynote version is also sharper: the design split should be pushed all the way back into the tuning stage before a model ships, and "reliability... can actually be a hindrance for a collaboration agent." ⚠️ What we confirmed today is provenance, not content: no study anywhere measures how rare delegatable tasks actually are, so that criterion is a useful frame rather than a verified conclusion — and we do not raise our confidence, because the newly found keynote is the same person speaking. The one adjacent piece of evidence we could find, a June preprint, tests something else: writing an explicit delegation contract buys reviewability, not correctness. ⚠️ Correctness did not improve because it was already saturated, which is not the same as "contracts don't work." The thing to take away today: one scorecard cannot rate every agent. The ones that work alongside you are judged on latency and transparency; the ones you throw a task at and walk away from are the ones where leaderboard scores matter.
[This week] (event date 08-17) Jensen Huang answers the circular-financing charge himself, and the answer is no — OpenAI will pay the lease. On August 17 NVIDIA's founder and chief executive published "Securing the Infrastructure of Intelligence" on the company blog, explaining the guarantee in main line item 2 in a question-and-answer format. "Circular financing" is an outside charge levelled at arrangements of this kind: the chip vendor invests in or guarantees its customer, the customer spends that money back on the vendor's products, and demand loops around on itself. This is the first time the accused party has answered it on the record by name — "Is this circular financing? No. OpenAI will pay the lease." We have run a whole thread on that mechanism, and until now it held analysis and circumstantial evidence but nothing from the party concerned. His chain of argument is worth recording. The software ecosystem makes NVIDIA compute versatile enough to be swapped between tenants, so a site built around it can be re-let if the first one walks; that, he argues, makes it "a productive asset: rentable and financeable." That amounts to an official answer to the question of what holds the residual value up: not the buildings, but whether anyone else wants the compute inside them. Three sets of figures in the same piece went unused today, all of them company estimates of opportunity rather than orders: each generation deployed at the site is roughly 1.5 million GPUs, worth roughly US$150 billion to US$200 billion in NVIDIA revenue; OpenAI's existing and planned commitments run to about 12 gigawatts, extending to about 16; and through 2030 the opportunity is around US$600 billion. The line most worth keeping from this defence is the one that cuts against its author. He writes that frontier labs are "growing faster than their balance sheets and long-term credit profiles can support," and that even with strong demand they still lack the decades-long contracts and investment-grade financing that infrastructure requires. That is an official acknowledgement that OpenAI is not investment grade today. Which is precisely the backdrop to the contract in main line item 2 writing "a satisfactory credit rating" in as its exit condition. ⚠️ The mandatory discount: this is a blog post, not a filing, and carries no legal liability; the author is the guarantor. Cite it as "NVIDIA's chief executive said," never as "the documents show."
No new paper to report today: not one of the four papers that came in overnight has been judged, and the same goes for the 176 arXiv papers backfilled alongside them, so we ran no forced paper sweep. What follows is a system measurement published on August 17 that happens to sit on the reverse side of main line item 2.
[This week] (event date 08-17) Same GPUs, same jobs — change only the order of the allocation decisions and utilisation rises by 33 percentage points. A team reported on the blog of the model hosting platform Hugging Face that their constraint-aware GPU allocator — their own product — was compared against the simplest first-come-first-served scheduler across seven scenarios. Hardware identical, workload identical, and GPU utilisation rose by as much as 33 percentage points (the control baseline sits around 50%, which is half a cluster idling). Priority-weighted output rose in all seven, by as much as 105%. That second ruler measures completed work weighted by job priority, against first-come-first-served in the same scenario. ⚠️ The two are different rulers and neither converts into "105% more efficient." The shape of the problem is worth understanding. Four kinds of work compete for one pool of GPUs. Training, batch inference and quantisation are batch-shaped — once started, each wants a whole GPU continuously until it finishes. Real-time inference is the opposite, with a demand curve that moves at every timestep. Real-time inference cannot wait, so the only way to guarantee availability is to reserve for the daily peak all day long. An application that needs six GPUs at midday and two at 4am holds six for twenty-four hours, and those four idle GPUs are unavailable to any batch job all day; they are not being used, and they are not free either. Read alongside main line item 2, it stings: on one side, 4.25 gigawatts of compute requires a US$105 billion guarantee; on the other, the same hardware yields another third of itself from a change of scheduling. How much of the "compute shortage" is real scarcity and how much is scheduling nobody fixed is a question with teeth for the capital narrative. ⚠️ That juxtaposition is our inference, not this post's claim. ⚠️ Four more limits travel with it: this is the publishing party's own measurement; the baseline is one they set themselves; there is no independent replication; and there is no peer review.
[This week] (event date 08-17) The AI editor starts hosting code itself: Cursor launches Origin. The AI code editor vendor Cursor announced Origin on August 17 — its own code hosting service, in early beta from that day, available on every paid plan. It launches with repos, pull requests, code browsing and GitHub sync, described as "designed for agent scale" (the announcement). What matters here is the direction, not the feature list. Code hosting has always been GitHub's territory, meaning Microsoft's, and an AI editor vendor integrating downstream into the hosting layer moves it from tool to platform. Its stated reason is that existing hosting services assume a human is using them, and agents access things differently. If an AI program opens hundreds of branches and runs hundreds of test cycles a day, permission models, notifications and review flows built around people become the bottleneck. The usable line today: the company states that pushes still go to GitHub, and GitHub stays the record that counts for anything started there. So this is a parallel experiment rather than a migration decision, and the question to ask when evaluating it is which workflows move across, not whether to switch vendors. ⚠️ This is a vendor's own product announcement, with no third-party verification, no usage data, and early beta status. A single product announcement does not amount to a new judgment either; for it to be worth something, a second and a third company have to do the same thing over the coming months.
No chips & semiconductors item this issue. No archive pick this issue. The chip thread is today's lead item, and the other chip material is at rumour level and sits in Also happened; the archive queue has run dry.
The past 24 hours. The overnight routine pulled in 52 long-form pieces and 439 posts, and this issue uses 36 clickable receipts, every one of them in the brackets above. By category: long-form — 52 pieces: 34 company and personal blog posts, 7 industry newsletters, 4 academic papers, 4 podcast transcripts, 2 company filings and 1 piece of industry analysis. Company filings — the 2 are NVIDIA and AMD; both were judged, both entered the graph, and main line items 2 and 3 come from them. Newsletters — 7, one issue each: ChinAI, Exponential View, Gary Marcus, Import AI, Interconnects, Latent Space and Stratechery (one of the last two is a paid subscription, from which we take direction only and quote nothing). Posts — we pulled from 374 platform-verified accounts and collected 439 originals, all of which went into the analysis layer; 116 accounts produced something; the largest was @teortaxesTex with 33, @TheStalwart 32, @vkhosla 24, @bhorowitz 21, @pstAsiatech 16, @GaryMarcus 12 and @emollick 12. Podcasts — 2 shows, 4 episodes: Dwarkesh with three and No Priors with one. This batch added no one-time sources.
What you are not getting today. Three things, said plainly. One, only 4 of the 52 pieces that arrived overnight were judged, and 2 of those are company filings. Today's substance concentrates in the filings line not because filings matter more, but because they are the only things that got read. Most of the material outside main line item 1 is what we went back and fished out of those 47 unprocessed pieces this morning, and more is still sitting there. Two, Ben Thompson's analysis of the Stripe acquisition is not available to you today: what we fetched was a headline, a one-line summary and a subscription template, with the body behind the paywall. The thread itself did not fall: the acquisition is the first item in Also happened. What fell is his aggregation framework applied to it — and what Stripe bought is precisely a routing layer aggregating more than 400 models. That is the day's most costly gap. Three, all 439 posts went into the analysis layer, but not one appears among the four pieces judged today; none of the 4 new podcast episodes was judged; and four newsletters — Import AI, Interconnects, ChinAI and Exponential View — went entirely unread. We would rather say so.
Backfilled material. No new one-time backfill batch this issue, but one box needs marking clearly: 176 arXiv papers came in overnight whose publication dates fall outside the past 24 hours. They are inventory we missed earlier and have now fetched, and this issue uses none of them. Likewise, the long-running backfill in the index — papers, posts, blogs, newsletters, filings and more, dated between July 1 and August 17 — is a cross-section rather than a continuum. Industry analysis and podcasts are backfilled only as far as July 7, and macroeconomics and supply-chain intelligence hold a single day each. This issue carries no macro section, because the most recent official data detection lands on August 16, outside our own 48-hour window, so there is no new period to write about.
Source-concentration warning. Two things. First, most of today's eggs sit in one basket: main line item 2 and Named commentary cover the same deal, NVIDIA, SB Energy and OpenAI's Ohio campus, and two of the documents behind it are NVIDIA talking about NVIDIA (the press release and the chief executive's blog). The three documents carry decreasing levels of liability: the filing carries legal liability, the press release less, the blog none. So every citation names its speaker, and we have not blended them into "the primary documents all say this." Second, the first and second items in Also happened both relay a single origin point, unverified by us today; both are marked as such where they sit, and readers should apply that discount. No single external source accounts for as much as a third of this issue's citations.
The sources we track. 529 named voices on the roster; channels are counted separately: 302 X accounts (Mark Zuckerberg, Arvind Narayanan, Sergey Levine, Daniel Kokotajlo and others); 90 podcasts (Sam Altman, Dario Amodei, Demis Hassabis and others); 51 media outlets and press rooms; 48 paper authors (Yann LeCun, Noam Shazeer, Ion Stoica, John Jumper and others); 48 blogs (Lilian Weng, Terence Tao, Armin Ronacher and others); 46 newsletters (Zvi Mowshowitz, Dean Ball, Ian Cutress, Eric Topol and others); 26 results and earnings calls (Jensen Huang, Lisa Su, Colette Kress and others); 23 keynotes; plus smaller channels including YouTube, courses, books, open letters and government documents. Channel counts and head counts are two separate ledgers and do not add together.
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 · 36 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.