🌐 Read this issue on the web

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

At a glance

1. Anthropic's chief executive took his own regulatory position down to statute level for the first time: the California bill he cited exempts outright any company below US$500M in annual revenue or model training cost.
2. One wave of anti-AI feeling, three principals, three explanations — and the first test to come due is the one Dario Amodei set for himself: if nothing loud about biomedical results arrives by year-end, that is the answer.
3. NVIDIA is putting money into land that already has its power deal done. The company it backed, Cloverleaf, has separately pulled out of a Wisconsin site after local opposition.

This issue draws on the research digest our system produced on August 23. The events and documents fall between August 15 and August 21, the most recent dated August 21. The overnight routine swept 75 long-form pieces and 354 posts and back-fetched 356 older blog posts; 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. [This week] (event dated August 15; yesterday's issue gave the link in the short items and called it first in line for today — this is the full treatment, after the checking) A rumour its subject has never flatly denied drew out the fullest account Anthropic's chief executive has given of where he stands on regulation

The spark was hearsay. Gavin Baker is chief investment officer at Atreides Management and a longtime voice on X and on podcasts about AI and semiconductor investing. On the All-In podcast he relayed that several people he trusts had told him Anthropic chief executive Dario Amodei had said Anthropic might at some point become the only private company in the world. That relayed line has no primary source we can point you to, so it carries no link here. The denial did not come from him. A researcher at Anthropic denied it in a personal capacity, calling it entirely false. We have that denial only through media relay, with no first-hand post in hand, so we record that it exists and neither quote it nor link it. Baker then posted a long argument on X — "I think safe to say that Dario has lost the argument" (@GavinSBaker, 08-15) — one reason being that "essentially every major company other than Anthropic has signed Jensen's letter," the open letter from NVIDIA chief executive Jensen Huang backing open-weight models. Open weights means the parameters are published for download, so anyone can deploy, fine-tune or run the model offline; the contrast is a closed service reachable only through the vendor's interface. "Essentially every" is his phrasing. He gave no signatory list and no count, so that sentence cannot be read as a ratio at all.

Dario answered the same night in two long posts (1/2, 2/2). Three things in the reply are genuinely new. First, a design principle for regulation, stated at a level you can go and check against statute text: he says Anthropic works hard on proposals that slow frontier AI companies down while advantaging smaller competitors, and he names a checkable basis: California's SB 53, which "completely exempt[s] any company below a certain amount of revenue or model training costs from being covered at all… it was $500M for SB 53." A second claim comes with direction only: by his account the federal testing process is more rigorous for frontier models than for off-frontier ones, and no document exists that a reader could check that claim against. Second, a structural argument: AI, he says, is structurally a technology that concentrates power, and open weights fall nowhere near a sufficient answer because "they simply shift the concentration somewhat to those with the most compute and chips." Third, a relayed policy line: Washington, he says, is reportedly pursuing "pre-deployment testing for frontier models, and also testing of open-weights models when they get closer to the frontier," and he is very supportive of it.

The same exchange settled an old account too. Our August 8 issue covered the open letter arguing for deliberately modulating the pace of automated AI development at the frontier (pacingthefrontier.com); critics read it then as incumbents locking the door behind them. What is new today is that Dario is himself a named signatory, and that he has now said in the first person what his preferred implementation looks like: modulate only the very best models, leave those catching up unconstrained; and, bluntly, "this hurts the business interests of the frontier labs and helps challengers, including open-weights!" That stands in direct opposition to the critics' reading.

Verification: the evidence for the exchange itself is hard enough. Both replies are full long-form X posts rather than 280-character truncations, we hold local copies, and TechCrunch's report the next day is a second confirming source. But it vouches only for the exchange having happened and the quotes being accurate, not for the truth of the rumour. The rumour itself is single-source hearsay that someone at the institution concerned has denied, and we take no position on whether it is true. The three new things also differ in strength: the SB 53 exemption line and the federal process are his own evidence, and we have not checked the bill text clause by clause today; the Washington route rests on one man's account, and until an official text exists it belongs in the assumptions column. Neither principal describes any of this neutrally: one defends his own regulatory position, the other argues the anti-regulation side from the vantage of an investor who has said he holds a position. Testimony from a party to the argument is not a finding: he has reasons to say this.

Where it stands (as of this issue): in the days after the exchange, financial media reported a rumour that Anthropic plans super-voting shares at listing, an arrangement giving founders more votes per share than ordinary holders once a company is public, and one of the few paths that reconciles "will never list" (the rumour this item opened with) with "about to list" (Baker's account, independently unverified). We have not verified that report ourselves. Both claims go on the record today as unconfirmed hearsay, and we do not reconcile them on anyone's behalf.

Judgment update: the premise that the AI industry stands united against regulation has to be rewritten from today as united, minus one; and that one has written its own exception into a statute you can look up. There is a direct use for you here: if you are building a policy scenario table, the SB 53 US$500M exemption line drops straight into it, while "pre-deployment testing at the frontier" belongs in the assumptions column until Washington publishes a document.

Investor note: the prevailing story treats regulation as a cost item that falls equally on every AI company. This evidence says it is differentiated, and that the line of differentiation runs along company size. For the assumption that regulatory cost flattens the whole industry, that is a weakening; for the reading that regulation widens the relative burden between the frontier and everyone chasing it, a strengthening (a limited one, because the Washington route currently has exactly one person relaying it).

2. [This week] (event dated August 15; our August 17 issue carried his "crisis of trust" line — what is new today is that the other two explanations now sit beside it) One wave of anti-AI feeling, three principals, three explanations, and the first test to come due is the one he set for himself

Read alongside main line item 1. That item records the exchange itself; this one is what becomes visible only once three people's words sit side by side. Our records now hold three competing explanations for why the American public has turned against AI and against data centres, all three spoken by principals, and all three make different, refutable predictions about how the mood recedes.

All three work off the same polling figure as their baseline. We cannot print that figure for you today. We read the poll only inside a paid subscription industry analysis, and we have not yet obtained the polling firm's own public page, so we record that it exists and quote no number from it. We do not ask you to believe a number we cannot link. The publicly checkable substitute, one you can open yourself, is main line item 3.

Of the three predictions, the first to come due is the one Dario set himself. In the second of the same pair of replies he wrote: "Anthropic is ramping up its efforts very quickly in biology and medicine, and we hope to have incredible results in the coming years and some early glimmers in the coming months. When we've actually accomplished something real, the whole world will hear about it, as loudly as possible, you have my word on that." (@DarioAmodei, 08-15) The structure of that promise deserves its own note: he committed to a timetable and to an obligation to publicise in the same breath, so silence at the deadline is itself a negative answer, and "we did it, we just didn't say so" is not available. He closed that exit himself. We put the verdict date at the end of 2026.

Verification: not one of the three is neutral: two are parties to the exchange, one a former insider. The three explanations are also not strictly exclusive; they can partly coexist, so the reader is not being asked to pick one: these are three lines that can each be tested separately. Read the wording of the promise closely: "hope" and "glimmers" are not hard commitments, and "the coming months" names no number of months. What keeps it testable is the binding (the world will hear about it, as loudly as possible) rather than the timetable itself. And "ramping up its efforts very quickly" has nothing independent behind it: no dollar figure, no headcount, no project count. Do not read it as anything quantified. Our August 17 issue covered the AI-designed bacteriophages from Arc Institute and Stanford that were called a Wright brothers moment the same week; that item is weak on its own terms: we obtained neither the primary release nor the original report, and what the AI did in the design step, how far it was validated, and what sample size it was tested on are all unknown. It serves one purpose here: this field already has people producing something they can show.

Judgment update: for a promise to be auditable, the test is whether the promiser has sealed off his own escape routes; sincerity does not enter into it. This one is a ready-made template: bind a timetable to an obligation to publicise, and absence becomes the answer. You can hold the same ruler against any company's public commitment — has it sealed off "we did it, we just didn't announce it"?

Investor note: the prevailing story treats AI-delivers-on-its-promises as a long tale with no due date, and both camps talk past each other. This evidence hands it a specific checkpoint, set by the party being tested. That undercuts the assumption that the delivery story cannot be falsified any time soon, a weakening; and it firms up the practice of treating the end of 2026 as a real checkpoint, a strengthening.

3. [This week] (event dated August 21) NVIDIA is putting money into land that already has its power deal done — and Cloverleaf has already pulled out of one Wisconsin site

NVIDIA has taken a minority stake in Cloverleaf Infrastructure and will support its developments with DSX, NVIDIA's own data-centre reference design (Data Center Dynamics, 08-21, a trade publication covering the data-centre industry). Cloverleaf's business is powered land: ground that already has its electricity arranged. The work is finding the site, negotiating grid interconnection capacity and securing transmission permits, then selling the plug-in-ready parcel to a data-centre developer. DSX is NVIDIA's reference design, a guide to how a building has to be put up to carry its latest racks, networking and storage; so NVIDIA has now pushed its own specification upstream, all the way to land and construction. Deals already closed support more than 7GW, including the Wisconsin sites sold to Oracle and OpenAI, and the company claims a further pipeline of more than 10GW. Do not add those two figures: the first is sold, the second is the company's own account of potential business with no third-party check.

Verification: one trade publication's report, and we have read neither NVIDIA's nor Cloverleaf's own announcement directly. Neither the size nor the percentage of the stake was disclosed, so "how heavily has NVIDIA bet" has no answer today; the appraisal quoted in the piece comes from Cloverleaf's chief executive, which is marketing language from an interested party. The piece does carry a counter-fact of its own, and we are not skipping it: it says the company has met local opposition in some markets, and links to a second report from the same publication saying Cloverleaf has withdrawn from a site in Kewaunee, Wisconsin. We neither fetched nor read that second piece today, so we record only that the publication has another report, and quote nothing from it.

Judgment update: read alongside main line item 2. The three explanations there argue over why the public turns against AI, and we could not hand you the polling figure; this item is that same thing on the ground: not a percentage, an abandoned site. Set it next to the TVA tariff our August 22 issue covered, where the Tennessee Valley Authority charges new data-centre developments an up-front capacity commitment of about US$1.5M per MW. Together they are two faces of one structure: power is now priced as an up-front financial commitment rather than rationed as a physical quantity, and what a land parcel sells is the interconnection already negotiated on it. One action here copies straight across: add a column to the siting checklist for how far the power deal has got, and whether the locality has a withdrawal on record.

Investor note: the story reads the data-centre bottleneck as whether you can get power at all. This evidence says the bottleneck has moved forward into land acquisition and local consent, and that even the chip supplier has begun putting capital at that end. That undercuts the assumption that power is a question of quantity, a weakening; it firms up the reading that what to watch is who holds land with power already arranged, a strengthening. But this is a single transaction with an undisclosed price, so the strengthening reaches "worth putting on the radar" and stops well short of "ready for a model".

4. [Evidence update] (the claimed events fall in May and June, the figures relayed in July; our August 22 issue put it in the tiered price table, and today we went and checked) The number we asked you to quote yesterday: we checked it ourselves, and there is no primary source

Yesterday we put four conflicting compute-price readings into a table showing that different layers carry different prices. One of them (single-card spot at US$4.22 an hour, down 31% in three weeks) we flagged at the time as a July relay of unknown provenance. Today we went back for it. The original relay of those figures turns out to be a news blurb on a cryptocurrency exchange (not a compute-market measurement outfit), and the trail stops one step above that. The result: we cannot find a primary source for these figures.

Verification: the checking did turn something up, and we are not printing it, for a reason worth stating without hedging. The best first-hand basis available is a published hourly rental index for the B200, and we obtained its current value, but the history sits behind a paywall, and we have no public page to hand you. That number would be no receipt for you, only our assertion. Our rule is that every figure comes with a source you can open yourself, and where it cannot, we do not print it. Today that rule applies to us. A separate cross-vendor median rental series points the other way, rising over the past year, but its basket and method differ and it likewise has no linkable public page, so it neither refutes anything nor gets its numbers quoted here today.

Judgment update: what the checking changed is not the number but its use. Quoting "single-card spot down 30% in three weeks" on its own as evidence of a compute glut does not hold from today: the primary source cannot be found, and the claim drawn from it, that compute supply is now growing faster than new workloads ramp, drops back to being an assertion. Our internal confidence score here stays at 0.5 (out of 1) and we leave it on the record, unresolved. That score is our own read of how well a piece of evidence stands up; 0.5 is the ceiling for a single source and means the evidence cannot yet carry either side (how we score). What would turn it from unresolved into confirmed or refuted is very specific: the actual values of that public index on May 30 and June 21. Get those two points and yesterday's tiered price table firms up with it.

Investor note: both camps quote single compute-price readings as directional evidence. This evidence says the most widely quoted of them runs back to a blurb with no stated measurement method. For the assumption that compute prices are already softening, a weakening; for the practice of asking where the primary source is before quoting any compute price at all, a strengthening.

Also happened (not yet checked by us)

1. [This week] (posted August 21) Google used several AI agents to work out which biomarkers from wearable-sensor signals are worth prioritising; we read only the opening section, and can answer neither the sample size nor the results (Google Research).

Chips & semiconductors

[This week] (reported August 21) What does Alibaba tell analysts when quarterly net profit falls 75% and capital spending rises 75%? Switch to chips we design ourselves.

On Alibaba's Q2 2026 earnings call, chief executive Eddie Wu told analysts the company will lean harder on chips designed in-house by its subsidiary T-Head and run in its own data centres; the report does not say whether any will be sold externally. He expects profit and gross margin to improve as those chips "account for an increasing proportion of total chips and replace commercially procured chips" (Data Center Dynamics, 08-21). T-Head is Alibaba's chip-design subsidiary, building RISC-V processors and AI accelerators. Three results figures from the same call: quarterly capital spending of US$10.07B, up 75% year on year; quarterly net profit down 75% year on year; AI and cloud product revenue of US$7.139B, up 45%, of which AI-specific items were US$1.82B — a subset of that revenue, not a separate line beside it. The two 75% figures are a coincidence, not two sides of one number.

Verification and how to use it: the report does not carry the date of the call, so we can confirm only the date of publication. All of it comes through one trade publication, and we have read neither the results statement nor the call transcript, so our internal confidence score here is capped at 0.5 — the ceiling for a single outlet relaying results (how we score). Separate the results from the targets. All three of these are the company's own targets: US$15B of external cloud revenue by 2030, a 20% gross margin, and AI capital spending with a payback the company hopes to shorten to 2.5 years. "In-house chips will improve gross margin" is likewise an assertion rather than a fact: the cost advantage of designing your own has to net out design, tape-out and software-ecosystem spending, and those rarely surface in a per-chip price comparison. The company gave a direction and no arithmetic. Where it earns its place is as a sample point — the discussion in our records about capital spending racing ahead of revenue is at present entirely American, and this is the first non-US comparison, under a different governance structure. If gross margin fails to improve next quarter as the in-house chips roll out, the cost-reduction story takes its first discount.

Named commentary

[This week] (posted August 15) A standard-bearer for open research says the published literature has stopped surprising her. And we decided today not to elevate it into a judgment.

Sara Hooker formerly led Cohere For AI — the non-profit research lab inside the model company Cohere, long associated with open-weight models and cross-lingual work — and has stood on the open-research side for years. She writes that she used to spend every Saturday reading papers and now forces herself through the ones already bookmarked, and she splits the reason in two: frontier work is no longer published, and "progress is narrowing so most papers echo each other in boring ways" (@sarahookr, 08-15). A minute later she asked her own followers for counter-examples: "What was the last genuinely surprising paper you read?" (the follow-up) We took it because a sentence like that carries different weight coming from someone whose own side it cuts against: from the standard-bearer of open research it lands far harder than from anyone inside a closed lab. The two reasons are two different diseases, and reading them together breaks them: one is work leaving the public channel, the other is homogenisation among the work still being published. Merge them into a single sentence and two separately refutable claims become one impression nobody can refute.

And we decided today not to elevate it into a judgment. Two other records point the same way — one on homogenisation in model outputs, one on the compute ceiling in academia, which our August 15 issue covered. Three arrows point the same direction with entirely different mechanisms behind them: converging research agendas, evaluation feedback loops, resource constraints. Not one of the three is quantified, and three weak signals stacked together only agree on a direction; they do not become evidence. We have cross-annotated the three records and are waiting for something measurable. We cannot say how much it has narrowed, and when we cannot say, we do not say.

Model watch

[This week] (posted August 21) One vendor ran the same method on two head-to-heads and reached opposite conclusions. What you can actually take away is the test for whether to chain two models together, not who won.

Together AI serves inference on open-weight models, and it published two comparisons on the same day using one method and one problem set — 113 software-engineering tasks, four runs each — changing only the opponent. GLM-5.3 is an open-weight model from Zhipu in China. The first (GLM-5.3 against Claude Fable 5): 69.0% against 69.7% on first-attempt pass rate, effectively level, at US$3.99 against US$21.63 per run — a 5.4x price gap. The second (GLM-5.3 against GPT-5.6 Sol): 69.0% against 72.7%, at US$3.99 against US$8.37. The two produce opposite operational advice — one says do not run both, the other says chain them. The two results come from one method pointed at two different opponents, so the opposite advice is what the test found, not a contradiction in it: when two models fail on the same problems (correlation 0.65), combining them only costs more; when they fail on different ones (0.43), the combination buys real complementarity. The cascade in the second piece — run the cheap model first, escalate when the tests reject the answer — reaches 85.9% at US$6.61 per task: 13.2 points better than the expensive model alone, and 21% cheaper.

Verification: the evaluator is an interested party, and the direction is obvious — "a cheap open-weight model plus orchestration beats expensive closed-source" is its commercial story, unmediated. There is also a scoring question large enough to flip the ranking: in the first comparison Claude Fable 5 recorded 16 infrastructure errors caused by model routing, and counting them all as failures puts it at 67.3% against the 69.7% officially reported — a 2.4-point spread between two scoring conventions, against a first-attempt gap of only 0.7 points, so changing the convention turns the conclusion over. Costs on the GLM side are given at index level with no per-run detail, so the two sides are not equally auditable. What you can take away, then, is the test — measure how far two models' failures overlap before deciding whether to chain them — and not the conclusion about which one is stronger.

Sources & accounting

The past 24 hours. Overnight the routine pulled in 75 long-form pieces: 36 company filings, 16 podcast transcripts, 14 X posts, 5 academic papers, 4 macroeconomic releases. Alongside them, 354 original posts from 748 accounts (reposts and replies were never on the list), and all 354 went through machine extraction. What we actually judged inside those 24 hours was 4 X posts, none of them excluded by any rule, with the other 71 pieces unread. The day's judgment work went entirely on older material from outside that window: the 6 remaining X posts from August 15, including the one we named yesterday as first in line for today. Four were taken, and 2 judged to hold nothing extractable. The reasons for not taking them belong here: one carried two things our records already held, the other was a product complaint about a social platform plus some thoughts on follower counts, with no figures, no mechanism and nothing that could be proved wrong. Fetching something is not the same as taking it. This issue uses 14 clickable receipts. Of the 79 accounts that produced anything, the largest were @teortaxesTex with 73 and @bhorowitz with 26, then @Miles_Brundage, @GaryMarcus and @emollick on 12 each. On podcasts: Dwarkesh 12 episodes, Colossus 3 and Gooaye 1. Among company filings the most came from AMD, Nebius and TSMC, 4 each. This round added 5 source records we did not have before, all of them used in today's extraction and checking.

What you are not getting today. Seven things, said plainly. One, another 356 blog posts landed overnight and we judged none of them, and one detail here would throw your arithmetic off: only 39 of those 356 were published in the past two days, the rest being back-fetches of older pieces from June through early August. We sampled 9; all 9 were old, the oldest by 24 days. So they count as backfill rather than yesterday's news, and we have not folded them into the 75 above. Two, one piece we owe you has come back as a summary only, two days running: the NVIDIA technical blog post headlined as reaching 100% on ARC-AGI-3, of which we hold 1,842 bytes of summary and no body. We said so yesterday, the state is unchanged today, and we have now established that this piece was not re-fetched in full alongside the others swept up with it. That is our fetch pipeline failing, not a judgment that the piece does not matter, and it is on the list to fix. Three, no macro section this issue: four official data series entered the store overnight, but the other half of the numbers needed to connect them to the AI capital-spending chain went unverified today, and we would rather write nothing than pad a section out of half a chain. Four, no trend-lineage section this issue: the most recent lineage material points at the model-memory line, nothing in today's material touches it, and forcing the join would be a join manufactured for the layout rather than for the substance. Five, no product moves column and no archive column this issue. No product news this issue. The new articles from product companies over the past two days went unfinished today, and we do not write an item from a title. No archive pick this issue. The older material we can use has run out, and we would rather leave the column empty than replay an item already published. Six, one item we finished reading and judged good enough still did not get written today: Hugging Face quantified optimising-for-the-benchmark with three probes — erase the digits from an audio file, then ask the model what it heard, and the strongest speech models "restored" roughly 30% to 40% of digits that were never in the audio (Hugging Face Blog, 08-21). Cramming it into today's issue would dilute both; it is first in line tomorrow. The link is yours now. One more was held over as well: former OpenAI head of policy research Miles Brundage predicts that Chinese model companies may invest more in content labelling than Elon Musk's xAI does, because they care about EU compliance, about being sued and about their reputations (@Miles_Brundage, 08-15) — we took it, but its key referent is one we inferred from context, and that step has to be nailed down before we write it for you. Seven, no new deep dive today: the most recent was finished on the evening of August 21 and the past two issues already carried it, and we do not re-run an old one to fill space.

Backfilled material. Beyond those 356 blog posts, no newly backfilled older material this issue, and no sources added by hand as one-offs. The long-running backfill that started July 1 remains a set of break points: academic papers, X posts, blogs and industry newsletters reach August 22; company filings reach August 16; industry analysis and podcast transcripts stop on July 7; macro data has a single day, July 22, and supply-chain intelligence a single day, July 4. To head off a misreading: a backfill stopping in July does not mean those lines fetched nothing last night. Routine and backfill are two separate pipes, and the break points sit only on the backfill one.

Source-concentration warning. Of the evidence added today, X posts account for more than four-tenths, and 5 of them came out of a single fetch on August 15, with 4 items resting on the same set of posts — main line items 1 and 2 and the named-commentary item all sit inside that concentration. This is why we marked "event dated August 15" so prominently today: those things are newly known to us, not newly happened in the world. Two more belong here. No material at all today came from the macro or supply-chain lines (the first did not run overnight, the second has had nothing new for a month and a half), and that is a gap in what reaches us rather than a preference in what we pick. And of the four speakers behind main line items 1 and 2, three are interested parties: two are principals in the exchange, one a researcher who has criticised one of them for years. Not a sentence from any of them can be quoted as neutral description.

The sources we track. After de-duplication the roster covers 529 named voices. Channels are a separate ledger, 645 source records in total: 302 X accounts, 90 podcasts, 51 media outlets and press rooms, 48 paper authors, 48 blogs, 46 newsletters, 26 results and earnings calls, 23 keynotes, plus 4 YouTube channels, 2 online courses, 2 books, 1 open letter, 1 internal document and 1 government document. Representative names: on X, Lucas Beyer, Sergey Levine, Arvind Narayanan; in newsletters, Zvi Mowshowitz, Dean Ball, Eric Topol; among paper authors, Ion Stoica, John Jumper, Yann LeCun; on podcasts, Satya Nadella, Sam Altman, Demis Hassabis. 645 and 529 measure different things: one person might hold an X account, have written a book and have appeared on a podcast, and gets counted three times. 645 counts records, 529 counts people, and the two do not add together. By the same logic, the 748 accounts actually pulled last night are not the same as the 302 X sources on the roster; the populations differ.

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 →