🌐 Read this issue on the web

At a glance

1. We got our own lead story backwards yesterday. The warrant Google holds is a discount given by the seller, and it runs 5 to 6.5 percent of what Google purchases.
2. Nebius borrowed US$5B at a 0.50% coupon and has to repay 110% of principal at maturity. The real cost is 3.27%, so coupons cannot be used to rank one borrower against another.
3. Applied Materials' quarter: not a dollar of this quarter's 25% growth came from China, where revenue fell 2% year over year.

This issue draws on the research digest our system produced today and on the deep dive it finished at 03:59 last night. The events fall on August 19 and 20, plus two July records we went back and checked. The overnight routine read 47 long-form pieces; 19 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] Our own overnight deep dive overturned yesterday's lead: the warrant is not Google clawing back cash but Marvell discounting, and the discount has a price

Here is what our August 20 issue said. Marvell issued Google a warrant on August 18 — a certificate carrying the right to buy newly issued shares later at an agreed price — struck at US$206.58, capped at 58.97 million shares. Of those, 57.61 million unlock only as Google itself buys custom silicon, one tranche for every US$500M of product, 240 tranches in all (our August 20 issue; Marvell's filing, 08-19). The reading rested on one number: the strike sat just 4.4% below the closing price on the day of issue, right up against the market, which made it the mirror image of AMD handing OpenAI roughly a tenth of the company at a one-cent strike. That one was the seller paying up. This one, we said, was the buyer turning money already spent into the supplier's share-price upside. What is new today: our own deep dive dug into the accounting rules overnight, and that direction is wrong.

The call: US accounting standards define equity granted to a customer precisely. It is consideration paid to the customer, measured at grant-date fair value, charged straight against the seller's own revenue, and not remeasured afterwards. If that holds, Marvell and AMD are doing the same thing in the same direction — both are sellers giving ground — and only the price differs. And the price can be worked out: roughly 5 to 6.5 percent of what Google buys.

Verification: three layers, of very different strength, and they need to be kept apart. The hardest layer is the share count, the strike and the unlock rules, all from a statutory filing where misstatement carries securities-law liability. The softest layer is precisely the ground this refutation stands on: for the accounting rule we read several large audit firms' interpretations of the May 2025 pronouncement from the US financial accounting standard setter, and we did not obtain the text of the standard itself, so the whole refutation rests on secondary material. In between sits the estimate: the 5 to 6.5 percent comes from a standard option-pricing formula, with the share price on the issue date, the US$206.58 strike, a seven-year term, a 4% risk-free rate and a 0.11% dividend yield as inputs. The one input with no market reading, and therefore an assumption, is volatility, where we used 40% to 60%. The four transactions of this type are Marvell to Google, AMD to OpenAI, Cerebras to OpenAI, and Amazon to Plug Power, and only one has actually been reported in financial statements (Cerebras to OpenAI, a discount of at least 19.7% — and what we read was a secondhand relay of that figure, not the original). One more thing has to be put on the table against ourselves. That "4.4% in the money" is brutally sensitive to which day you pick as the reference. On July 29, the day the agreement was signed, the strike was 26.4% out of the money. On August 17 it was 11.8% in; on August 19, 12.9% in. Four days, a swing of 39 percentage points — and August 18 itself fell 7.8% from the previous day before rebounding 9.9% the next. Across those three weeks, the August 18 close was the single day nearest the strike. We picked the one day on which "right up against the market" could be true. (Closing prices come from a subscription market-data service, so there is no public link to point to.) Separately, some exhibits to the warrant were omitted as the rules permit, so we have not seen the anti-dilution terms, the acceleration terms, or whether the strike can be reset.

Judgment update: we are acting on it today and marking that judgment down. Its internal confidence score falls from 0.6 to 0.45 (out of 1, where anything below 0.5 means the evidence does not yet carry the claim; how we score), and its type changes from structural analysis to unsettled. Three observations survive untouched: the strike mechanism, which still differs from AMD's; the unlock, still tied to Google's own purchasing; and a dilution ceiling of about 7%. What fell is the next step only — "therefore this is a buyer recovering cash." We did not cut to 0.4 as the deep dive recommended, and the reason belongs here. The accounting characterisation doing the refuting is itself draft material in that same report, carrying the same 0.5 confidence, sourced from six firms whose readings are not independent of one another. Using a 0.5 to push another claim down to 0.4 gives the refuter better evidentiary treatment than it gives itself. Both sides stopping at "waiting for a primary document" is the honest place to stand.

Investor note: the prevailing story reads a chip vendor handing equity to a large customer as a tell that the seller is negotiating from weakness. This evidence says all four cases are sellers giving ground, all four charged against seller revenue, and only the price differs. Where a comparable rate can be computed, Marvell's 5 to 6.5 percent is the cheaper of the two, against Cerebras at 19.7% or more. AMD's one-cent strike puts it at the top of the range, but its terms do not convert cleanly onto the same denominator, which is the fair value of the warrant divided by the revenue that unlocks it. For Amazon to Plug Power we never pulled the original filing, so there is no rate for it at all. For the assumption that equity to a customer signals a weak seller, that leaves it unchanged: the direction is right, but the strength has to be graded by discount rate rather than inferred across the whole category. What it strengthens is the reading that these suppliers' future revenue growth will be held down by a discount nobody can see.

Why we dug this up today: because the judgment carried its own watch condition — "a third instance is needed before calling this an industry template." The deep dive found that the count had the wrong reference class. The real class is "using equity to anchor a customer," and it has been running for nine years in logistics and hydrogen (Amazon to Plug Power, April 2017, with the same unlock condition of the buyer purchasing US$600M of product; we read that one secondhand and did not pull the original filing). In the right reference class, Marvell's warrant means only that a mature practice has arrived in custom silicon for the first time.

What would prove this wrong: if Marvell's August 27 earnings call discloses a fair value for this warrant well below US$4B, the six percent has to be rewritten. That US$4B is a reference line we backed out of our own valuation, not a figure Marvell published. The other single point is the full text of the warrant exhibits — anti-dilution, acceleration, whether the strike can be reset. That document can void both sides at once, because the refutation and the thing refuted both rest on the premise that the strike is fixed.

Verdict date: August 27, 2026. And the rules say this charge is recognised only once the customer buying that volume becomes probable — which means that how much Marvell recognises amounts to Marvell publicly declaring how much it thinks Google will buy. It is a forward-looking disclosure forced out by an accounting standard: a supplier has to convert its internal forecast for a single customer into a number on the income statement. Nobody had to do that before.

Who this is for: cloud buyers now have one more thing they can ask for, and it comes with a price tag — 5% to 20% of the purchase value, depending on how badly the supplier needs your order. Watch how you get it: the strike is set against an average over a period, so which stretch of the market your negotiating window falls in matters more than how well you negotiate. For custom-silicon suppliers, the mirror image: those six points land on the revenue line, not the expense line, so every future quarter's revenue growth gets held down a little, and how much is decided by the auditors' judgment of how much the customer will buy.

2. [Today] (event date 08-19) The bond with the 0.50% coupon really costs 3.27% — and the same company's money, at the same tenor, got 211 basis points more expensive in five months

Nebius — headquartered in Amsterdam, listed on Nasdaq, an AI cloud operator whose business is renting out GPU capacity — priced US$5B of convertible notes on August 19, above the US$4.5B it first announced. A convertible note is a form of debt: you borrow, you pay interest, and the holder can swap the debt for stock when the conditions are met. It came in two tranches: US$3B at a 0.50% coupon due February 2030, and US$2B at 4.50% due February 2034. The clause that matters sits in the middle of the release. What comes due at maturity is not the principal but the principal after accretion. The first tranche repays 110%, the second 125%, and interest accrues on the original amount borrowed rather than on the accreted figure (Nebius's filing, 08-20). The conversion hurdle has two layers as well: nominally the stock has to rise 40.0%/45.0% to pay off, but the company works it out for you, and with accretion counted the real break-even is 54.0%/81.3%. The use of proceeds is stated plainly: data centres and GPUs. Net proceeds are about US$4.94B, settling August 24.

Verification: every figure above comes from a statutory filing, and the issuer carries legal liability for it. The "true annualised" column below is ours, though, and no party to the deal supplied it: we assemble the coupons and the maturity multiple into a stream of cash flows and back out the internal rate of return — what this money actually costs per year.

(A basis point is 0.01 of a percentage point. The March terms come from the same company's March 17 pricing release.)

Judgment update: the coupons range ninefold, from 0.50% to 4.50%. The true costs range by a factor of 2.2. The lowest coupon is not the cheapest money — it is the shortest paper. And the gap is not a constant (+366/+250/+277/+273), so you cannot add one back in your head: any ranking of funding costs by coupon comes out in the wrong order. Matching tenors is the fair comparison. March's seven-year money cost 5.12%; August's seven-and-a-half-year money cost 7.23%. That is +211 basis points in five months, with six extra months of tenor nowhere near enough to explain it. Over the same stretch the company's stock ran from US$116.33 to US$223.90, close to a double, while the company's own stated effective conversion hurdle at the long end actually fell (86.0% → 81.3%). Investors are asking for more interest on the debt leg and accepting a cheaper option on the equity leg. Equity marks it up, credit marks it down, both on one press release. There is a hole here we have not filled, and it goes first: if benchmark rates or credit spreads themselves rose roughly 200 basis points between March and August, the 211 could be entirely a market-wide move with nothing to say about this company's credit. We have not pulled the government-yield and credit curves for that window, so this cut has not been made. Readers do not have to wait for us. The next time this issuer prints the same tenor, set its true annualised next to 7.23%. Higher again means the credit market is repricing this class of borrower; back down means August was down to market conditions or a one-off. That is also why our internal confidence score here is only 0.65 (out of 1, meaning the direction holds up but one competing explanation is still standing; how we score). And one item is ours to own. The March record we filed in July described the raise in plain language as "cheap convertible financing." Going back to the primary document today to build the comparison, we found we had missed the same 120% accretion clause — we understated the cost by 250 to 366 basis points, and the word "cheap" was simply backwards. It has been corrected, with the correction left visible. The failure this item describes is one we ran for five weeks ourselves.

Investor note: the prevailing story assumes the financing window for compute landlords is still open, on the evidence that each round is larger than the last. This evidence says "can borrow" and "can borrow cheaply" are two different things, and the second is deteriorating: the same issuer, the same tenor, 211 basis points more in five months, while the stock nearly doubled. That weakens the assumption that capital markets are still backing data-centre construction indiscriminately. It strengthens the reading that the number to watch is the true annualised cost at matched tenor rather than how much was raised, provided the missing cut leaves the direction intact. For buyers assessing this class of compute supplier, the use is not to predict a price but to add a diligence field: at what price the money funding that expansion was raised. These operators have no large existing cash flow, so growth runs almost entirely on capital-market funding, and every increment on the true cost at matched tenor takes away some of the cushion.

There is a second half to the same day, and reading one without the other misleads. In the same release, Nebius also negotiated privately with a handful of existing noteholders to swap US$800M of face value in old notes for roughly 15.8 million common shares. Multiply by the day's close of US$223.90 and those shares are worth about US$3.54B, or 4.4 times the face amount retired (that multiplication is ours; the company gave no such figure, though both numbers going into it are printed on the same release). We cannot compute what percentage those 15.8 million shares represent, and we are not going to guess: turning it into a percentage needs shares outstanding as the denominator, and this document does not give one. The company itself discloses that recipients may sell, or unwind hedges they had put on, and that this could push its own share price down.

3. [Today] (event date 08-20) Applied Materials' quarter: not a dollar of this quarter's 25% growth came from China

Applied Materials is one of the largest semiconductor equipment makers in the world. Fabs — TSMC, Samsung, SK hynix, Intel and the rest — buy its tools to manufacture chips. Its customers are the chipmakers; its products are the tools they build chips with, which turns its quarterly report into an upstream read on how much capacity fabs intend to add this year. For the quarter ended July 26, revenue was US$9.115B, up 25% year over year. But China fell 2%, from US$2.548B to US$2.506B, and its share slid from 35% to 28% (Applied Materials' quarterly report, 08-20). The growth is everywhere else, ranked by size: Taiwan US$2.025B (+10%), Korea US$1.521B (+31%), the United States US$1.367B (a clean double), Japan US$839M (+18%), Europe US$483M (up more than 200%), Southeast Asia US$374M (+92%). The seven regions add to exactly US$9.115B.

Verification: this is the statutory quarterly report of a US-listed company, bound by audit process and regulatory rules, and among the strongest evidence in this issue. Two definitions have to be clear first, or the numbers get read wrong. First, geography is recognised by the location of the customer's plant, which is not where the chips end up and not where the customer is headquartered. Second, this quarter closes on July 26, not at a calendar quarter end, so it cannot be lined up directly against calendar-quarter figures from other companies. Note also that Japan is the one line in the table where the money rose and the share fell (from US$713M to US$839M in dollars, but from 10% to 9% in share) — a demonstration that a falling share is not a revenue decline, while China's line is the opposite combination. The same document discloses something that belongs beside this and cannot be used to explain it: on February 11 the company settled with the US Department of Commerce over shipments to Chinese customers and an export-control compliance investigation, paying US$253M, with a further penalty suspended and waived after three years. We draw no causal link between that settlement and China's flat revenue. We are noting that it sits there when you read the China line.

Judgment update: the common assumption in equipment-cycle discussion for several years has been that Chinese mature-node expansion keeps buying tools as a floor, whatever leading-edge demand does. This quarter that floor is flat (−2%), while the world is still at +25%. Any cycle model that treats Chinese restocking as a cushion needs re-estimating. The next test is the same equipment maker's China share next quarter: flat a second time and the cushion assumption stops being a one-quarter story. What the tools are being sold to build is shifting too. Inside the semiconductor systems segment, memory (DRAM) rose from 22% to 26%, logic and foundry fell from 69% to 67%, and flash fell from 9% to 7%. Management splits the logic line into two opposing halves: leading-edge demand is strengthening while mature-node demand is falling (the latter stated over the nine-month period, where the offsetting relationship appears in the original). So "logic demand is growing" as a headline works the same way "+25% revenue" does: both net two opposing signals into noise. One guardrail belongs here: yesterday we relayed a secondhand claim that memory will reach 48% of large cloud operators' capital spending by 2027, and today's 22% to 26% cannot be set against that number. One denominator is one segment of one equipment maker's revenue; the other is a year of global cloud capital spending. The dimensions do not even match, so you can compare directions and not values.

Investor note: the prevailing story reads equipment makers' revenue growth as one block and calls it evidence that AI capital spending is still accelerating. This evidence says the composition has shifted: China's line is flattening, and the United States and Europe absorbed almost the entire share shift (Asia-Pacific combined fell from 89% to 80%). That weakens the assumption that Chinese mature-node expansion is the floor under equipment demand. It strengthens the reading that memory's share of equipment spending really is rising, but what it strengthens is the direction, not any particular percentage.

Also happened

1. [Today] (event date 08-20) A venture-capital reporter says he obtained an investor letter stating that NVIDIA is paying the code-model startup Poolside US$6B for a non-exclusive licence, and separately investing US$1B at a US$12B pre-money valuation (Newcomer, 08-20). We did not take any of it as fact: a single origin, resting on an anonymous letter we have never seen, and the report sits behind a paywall where we read only the opening line. Deal structure, timing, whether it carries a purchase commitment — all unknown. What makes it worth writing down has nothing to do with the amount. It is the direction. The past six months kept producing the same shape: a chip vendor handing equity to a large customer in exchange for orders. This one runs the other way, with a chip vendor paying cash to a downstream model company. The accounting consequences and the competitive signal are completely different, so we are deliberately keeping it out of the instance count for that pattern. Counting unconfirmed cases is manufacturing confirmation out of volume.
2. [Evidence update] Stripe's acquisition of OpenRouter has now been confirmed by the buyer itself, but the price has not — the announcement carries no figure at all, so we are raising only half of it (our July 27 issue). (Room for one line only today; the issue linked above carries the detail.)

Model watch

[This week] (event date 08-20) Z.ai chief executive Tang Jie: what really blocks post-training is no longer the model, it is the environment — and he says environments and graders can both be synthesised.

Start with who this is. Z.ai (Zhipu) is a Chinese frontier model team, developer of the open-weight GLM series, and its chief executive Tang Jie was a professor in Tsinghua University's computer science department before this. He made two points about the new GLM-5.3 (Latent Space, 08-20). First, parameter count no longer describes a model — it has to be read alongside how much data went in, where the compute went, and who runs the thing under what conditions. The second point is the one that matters. A model reads vast amounts of text in pre-training, then goes through post-training with methods such as reinforcement learning, and most of the capability jump of the past two years came from the second stage. Scaling post-training is no longer hard because of the model. It is hard because of the environment: a simulated workplace the model can actually operate in, where success takes many steps to establish (hand it a whole machine-learning stack, ask it to diagnose the bottleneck in a training run, implement an optimisation, run the experiments, and deliver a measurable speed-up without breaking correctness). Environments like that used to be built by hand, one at a time. Tang says Z.ai has a production line that synthesises environments end to end, and for some tasks synthesises the reward signal too. More striking still: the grader is synthesised without sight of the reference answer, and it has to clear three checks — feed it the correct answer and it must pass, feed it "did nothing" and it must fail, feed it a half-finished state and it must fail as well.

What this supports or contradicts in our own judgments: we have tracked one judgment for a long time — that the real constraint on AI capability is not compute but human verification bandwidth: the faster machines produce, the more the people who can tell whether a result is right become the bottleneck. If Tang's account holds, machines have routed around a stretch of that bottleneck. We are keeping that tension and leaving the judgment's score where it is, and the reasons deserve stating: (1) this is a vendor describing its own unpublished method; (2) there is no third-party replication, no control condition and no performance number at all — the piece gives no benchmark scores for GLM-5.3 and no size of improvement; (3) the piece says the capability jump comes "solely" from long-horizon reinforcement-learning environments, and "solely" is its word, which we do not read as "the only cause." Those three checks also remain a proxy, and a model can still find ways to score without really solving the task — which is exactly the failure mode that judgment describes. We also turned down an inference derived from this today ("environments can be mass-produced, so their scarcity value gets competed away"): its only foundation is one vendor's account of itself, and promoting it would turn a marketing line into a mechanism claim of ours.

How to use it: if you are drawing up a budget for whether to build training environments in-house, this material gives you not an answer but the test between two opposite worlds. If environment supply really can be automated, an advantage built by hand-crafting task sets gets compressed; if it only works in domains that are easy to check, environments stay a scarce asset and the value concentrates with whoever holds real workflow data. Two things settle it: whether anyone else can reproduce that size of improvement on their own long-horizon tests, and whether the automatic grading really does block the shortcuts.

Sources & accounting

The past 24 hours. Overnight the routine read 47 long-form pieces: 37 company and personal blog posts, 8 industry newsletters, 2 company filings. Alongside them we pulled 277 original posts from 161 accounts, all of them extracted. What we actually finished reading and judged today: 5 pieces (4 taken, 1 filtered out). This issue uses 19 clickable receipts, every one of them in the brackets above. By category, company filings: Nebius's 6-K, Applied Materials' 10-Q and Marvell's 8-K; main line items 1, 2 and 3 all come from here. Newsletters: Latent Space, Newcomer, Gary Marcus and The Zvi. Company announcements: Stripe's newsroom. Posts: the largest accounts were @teortaxesTex with 46, @elonmusk with 32, @pstAsiatech with 30 and @Miles_Brundage with 18. This round added 5 sources we did not have before: Nebius's filing, Applied Materials' quarterly report, Stripe's official announcement, Newcomer's scoop and Latent Space's roundup.

What you are not getting today. Five things, said plainly. One, of the 47 pieces only 5 were finished today, and the largest block by far — the 37 company and personal blog posts, 79% of the total — went entirely unread; not one of the 277 posts reached the judgment stage. Together with the previous day's 612, that is 889 original posts across two days that we have not used at all. Two, three columns are missing from this issue. No chips & semiconductors item this issue. No product news this issue. No named commentary this issue. The chip material all went into main line items 1 and 3, and opening a separate column would be saying one thing twice. On products, 11 new articles arrived, but we have titles and nothing else, and we do not write an item from a title. Among named commentators, three Gary Marcus pieces and one from The Zvi went unfinished. Three, no archive pick this issue. That column's stock of retrospective material has been empty for 22 days, and we would rather leave it blank than replay an item we have already published. Four, four parts of last night's fetch failed, and three of them came back clean on a re-run this morning. The remaining one was a 503 from X's servers, a temporary outage on their side, so last night reached only 161 of the 302 accounts on the roster, and that line was running at half strength all day. Five, there is no macro section and no trend-lineage section this issue: the most recent official data is dated August 16 and the most recent lineage material August 18, both outside our own 48-hour window.

Backfilled material. No newly backfilled older material this issue. The long-running backfill that started July 1 remains a set of break points rather than a continuum: arXiv papers, X posts and a second paper channel keep up to August 20; blogs, newsletters and company filings reach only August 16; industry analysis and podcasts stop in the first week of July; macro data has a single day, July 22, and supply-chain intelligence a single day, July 4. Those six break dates are identical to the previous day's, not one has moved, which is not random. Those lines are not running.

Source-concentration warning. Close to 60% of the material behind today's judgments comes through one channel: statutory filings with the US Securities and Exchange Commission. That is not one source — Nebius, Applied Materials and Marvell are three unrelated filers, each carrying its own legal liability — but one channel dominating brings its own bias. Statutory filings tell you what has been signed, never why it was signed or what the other side was thinking, so this issue leans toward the contract and financial-terms layer and is thin on industry dynamics. There is a narrower concentration inside it: Nebius alone accounts for a third of today's material, and all four aspects of main line item 2 come from one document from one company. If the reading of that document is systematically wrong, the whole passage falls with it. And one of the load-bearing materials under main line item 1 is our own deep dive, marked as such in the body. That is not third-party endorsement.

The sources we track. 529 named voices on the roster; channels are counted separately: 302 X accounts (Mark Zuckerberg, Lucas Beyer, Arvind Narayanan and others); 90 podcasts (Sam Altman, Dario Amodei, Demis Hassabis and others); 51 media outlets and press rooms (Stephanie Palazzolo, Ilya Sutskever and others); 48 paper authors (Yann LeCun, Noam Shazeer, Boaz Barak and others); 48 blogs (Lilian Weng, Terence Tao, Armin Ronacher and others); 46 newsletters (Zvi Mowshowitz, Dean Ball, Ian Cutress and others); 26 results and earnings calls (Jensen Huang, Lisa Su, Matt Murphy and others); 23 keynotes (Bill Dally, Werner Vogels and others); plus 4 YouTube channels, 2 online courses, 2 books, 1 open letter and 1 government document. Company and institutional blogs add another 76 (NVIDIA Technical Blog, SemiAnalysis, More Than Moore, Data Center Dynamics and others); those are institutions rather than people, so they are not counted in the 529. Channel counts and head counts are two separate ledgers and do not add together.

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 · 17 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 →