🌐 Read this issue on the web

At a glance

1. For the first time, the gap between NVIDIA and AMD can be measured in public, by anyone willing to re-run the numbers. The dividing line is not raw speed; it is one software interface AMD spent a generation without.
2. Anthropic's revenue figure and its flagship model's adoption rate, set side by side, produce a false conclusion. We are declining the most tempting reading of them today, because the denominators do not match.
3. Silicon wafers, the cheapest material at the very top of the chain, are reportedly going up in price for the first time in more than three years — while ASE, the packaging house spending the most money, says what it lacks is construction time, not capital (rumour-grade).

This issue draws on the research digest our system produced on August 24. The events fall between August 21 and August 24, the most recent of them in the small hours of this morning. The overnight routine swept in 192 pieces; 16 clickable external 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. [Today] Above 20 concurrent users, every one of NVIDIA's cheapest configurations moves the conversation cache off the GPU. On the AMD side, not one of them does

SemiAnalysis is an independent research firm covering semiconductors and AI infrastructure, founded by Dylan Patel and known for taking apart chip supply chains and data-centre cost models. In the small hours of this morning it launched AgentX 1.0 on its own measurement platform, InferenceX: an open benchmark for agentic coding inference. Agentic coding means letting a model work like an assistant across dozens or hundreds of turns, reading files, editing them and running tests on its own. Terminal coding tools such as Claude Code are the archetype (SemiAnalysis, 08-24).

Start with why the ruler had to change. Chip comparisons for AI have always fixed one input length and one output length. Agentic traffic looks nothing like that: turn n's input is turn n-1's output with more appended to it, so nearly everything in it was already computed a turn ago. When a model reads a passage it produces intermediate state that can be kept rather than recomputed, and the industry calls that stored state the KV cache: it is the physical form of "the work already done." Keep it and costs stay low; lose it and the whole passage runs again. The share you keep is the cache hit rate.

The cleanest dividing line in the results is this: above 20 concurrent users, every one of NVIDIA's cheapest configurations pushes cache out of GPU memory and into main memory. GPU memory is expensive and small, main memory cheap and large, so when it no longer fits it moves outward; the industry calls this KV offload. On the AMD side, the number of cheapest configurations using offload is zero. The original is blunt about it: "All of Nvidia's Pareto optimal points include KV offload above concurrency 20, but for AMD none of the Pareto optimal points use KV offload to DRAM." ("Cheapest configurations" renders the article's Pareto optimal points: the best-value setups at a given speed requirement. It is a set of configurations, not a single number.)

The cause is not that AMD chose to skip it. The platform was missing a part. The interface that bundles many small transfers into one large one (hipMemcpyBatchAsync) was absent from AMD's ROCm software platform until version 7.14. Before that, transfers went one at a time, offloading ran slower than not offloading, and so every best-value answer fell back to keeping everything on the GPU. What today's readings capture is the state before that part arrived. Whether the fix changes the outcome depends on how far 7.14 spreads across deployments. The moat is being lost on an absent interface, not on arithmetic speed.

Verification: this is a single source, and it is a measurer reporting on its own measurements, with no third party having re-run them. The conflicts of interest have to be named in the same breath as the finding: SemiAnalysis describes years of consulting-style contact with AMD, and it trades named acknowledgements with the NVIDIA, vLLM and SGLang teams; it also cares openly about whether this benchmark can move industry opinion. The counter-evidence is that it criticises both sides by name. We split "what it open-sourced" from "what it measured" and checked them separately today. We checked the first of the two the same day, and the open-source side does deliver: results pages are public, the code carries 1,463 stars on GitHub, the public interface is documented, every measurement run executes on GitHub Actions with logs fully open, and the real Claude Code conversation traces used for replay are already published on HuggingFace, the public platform for models and datasets. The licence is Apache 2.0: outsiders really can take it apart, and that no longer rests on the firm's own word (the InferenceX public platform). Three claims remain unconfirmed: that it cost more than US$3M to build, that it runs on about 2MW and more than 1,000 chips, and that several frontier labs already use it. Those three rest on the firm's own statement alone. Our internal confidence score on this record therefore rises from 0.7 to 0.8 (out of 1), meaning the deliverable is solid while the claims about scale remain unchecked by anyone (how we score).

Judgment update: our July 25 issue recorded a judgment: NVIDIA's software moat has not fallen, it has moved — out of code quality on a single chip and into how hundreds of cards divide work in real time. We wrote down what would overturn it at the same time: if AMD could get several key optimisations to assemble reliably together inside the default open-source software, the judgment falls. That condition sat only in our own records; readers never saw it that day, so we are printing it now. The judgment also came from a single source that day, and we recorded it at low confidence. A month later, today's measurement gives the first public reading on the KV offload part of that condition: on the AMD side, even the cheapest configurations cannot get there. The overturning condition has not fired. We are not writing "judgment unchanged," though — a judgment written that way is still the same sentence a year later, and never gets any smarter. We are narrowing it instead: the original spoke of a whole basket of optimisations, and today's numbers show the weights inside that basket are wildly uneven. What actually separates the two companies is the handful of items tied to the cache lifecycle. Narrowing costs range and buys a test: instead of "can your system be assembled reliably," a question that invites a well-phrased answer, you get one that is binary, checkable and hard to talk around. Look at whether a vendor's cheapest configuration above 20 concurrency uses DRAM KV offload. You can put that question to a supplier today. If you run your own inference, or you are choosing a cloud inference backend, you do not need to wait for anyone's quarterly roadmap — ask that one binary question. What would prove this wrong: AMD's cheapest configurations start using offload once ROCm 7.14 spreads, and the per-dollar performance gap between the two does not narrow along with it. That would mean we narrowed in the wrong direction. The verdict date is mid-September: the publisher says an update article is three to four weeks out and will include AMD performance optimisations. The other route carries no date and matters more — anyone at all taking the Apache 2.0 release and re-running it.

Investor note: the prevailing story reads AMD's catch-up as a continuous curve you check quarter by quarter. This evidence says the place where catching up happens has moved, and moved somewhere with a named missing part, a version number and a checkable answer. For the assumption that AMD is steadily closing the gap, that is a weakening; for the reading that the cause of the gap can be verified from outside, a strengthening. The strengthening is bounded by the fact that this remains a single source that nobody has re-run.

2. [This week] (event dated August 23) Anthropic's revenue figure and the adoption rate of its flagship model, set side by side, produce a false conclusion. We are declining it today

Two numbers are on the table. The first: the Financial Times, citing people familiar with the matter, reports that frontier AI lab Anthropic reached US$65B in annualised revenue in July, against US$47B in May — 38% growth in two months. The company also told investors it has 6,000 customers spending more than US$100,000 a year, and expects to be profitable in the third quarter. We read this through independent developer Simon Willison, co-creator of Django, author of Datasette, and a long-time public tracker of the revenue series each lab reports (simonwillison.net, 08-23); the FT original sits behind a paywall and we did not obtain the full text. The second: Ramp is an American corporate card and spend-management company, and it uses the actual card statements of its 70,000 customers to estimate how July spending on Anthropic split across models — the expensive flagship, Fable 5, holds just 8.0%, less than a third of the share held by the year-older Opus 4.8 (28.0%) (the Ramp AI Index).

Put the two together and a very pleasing judgment appears: the revenue is being carried by the mid-tier models, and the flagship will not sell. That is precisely the judgment in the FT headline.

We are declining it, because the denominators do not match. Ramp's population is American small and mid-sized businesses and startups paying by card. Flagship usage most plausibly lands on three other channels: direct enterprise contracts, cloud-marketplace resale and large annual API commitments — that population excludes the flagship's home ground almost exactly. The revenue figure's denominator is the whole company; the card index's denominator is the long tail. Both can be true at once, and very likely are. Dividing one by the other takes a share of the long tail and uses it to explain revenue across the whole market. That is not an insight. It is a mismatched denominator.

Verification: three reservations travel with this one. The first is that the evidence chain runs two removes deep: we read a relay, and the FT's numbers come from people familiar with the matter. The next one matters more: "annualised revenue" is a company's own figure, some month multiplied by twelve, not audited revenue, and for a fast-growing company it amplifies whatever the current period is doing. On that same curve, the US$47B for May comes from Anthropic's own press release and is primary; the US$65B for July arrives through an anonymous chain. Plotted on one chart, the two points do not carry equal credibility. One more thing needs saying up front: the FT attached its own qualifier to "expects to be profitable in Q3", using the same accounting approach it used when declaring Q2 profitability. That is a reminder that profit here depends on how training costs are amortised. It is not an endorsement. The card table has a pothole to fill in as well: Opus 5's 3.5% means nothing — it shipped on July 24 and had one week of life, a point the relayer made before we could. What is worth reading is that Fable 5, out for far longer, sits at 8.0%.

Judgment update: this one deserves a couple of extra sentences, because it is the reason our whole rulebook exists. Both numbers are real, and side by side they yield a false conclusion that reads more like an insight than either true number does. The defence is not scepticism about numbers. It is asking, every time, whether these two numbers share a denominator. You can take that action away today: any time you see "adoption is low but revenue is high" or "few users but lots of income" placed side by side, ask about each denominator separately first. We have also written down what would rescue the judgment: a model-by-model usage split that includes the enterprise channels — a vendor disclosure, or a second spending index built on a different population. We will reopen it then.

Investor note: the story currently treats a third-party card index as a proxy for model adoption, then treats adoption as a proxy for revenue composition. This evidence says the second proxy is broken: using long-tail adoption to infer a frontier lab's revenue structure does not stand up today, and our confidence in that practice goes down. In the opposite direction, this episode makes the case for checking each proxy's population first, and that practice comes out stronger for it. Whether the flagship sells is a question today's material cannot answer, and we are not pretending otherwise.

3. [Today] SemiAnalysis says that even if AMD's chips were given away, NVIDIA still works out cheaper. That collides head-on with a record we took in July, and we are not reconciling them

Read alongside main line item 1. The same measurement produces a threshold reading: NVIDIA buys up to 5x the performance per dollar that AMD does on GLM 5.3 (the frontier open-weight model from China's Zhipu, whose weights are published for download so both platforms can run the identical thing) when holding to 150 tokens per second per user, a common interactivity requirement. The article goes on to say the gap is wide enough that "even if the competitor chip hardware was sold for free (but with providers still of course paying for datacenter hosting and power and other operating costs), cost per token would still be cheaper when using Nvidia" (SemiAnalysis, 08-24).

The value of that sentence is that it changes the comparison: from "how much cheaper" to "who wins once the hardware is free." If it holds, then within that range the chip's price has stopped being a decision variable, and what decides cost is the hosting and power standing behind each chip. Not one condition can be stripped from it, though: one model, one speed point, and both sides running open-source software. Inside the same article, AMD's own closed software beats NVIDIA's flagship rack in other ranges. Anyone using that sentence to argue AMD is finished has removed the conditions.

Verification: single source, single range, and it runs against a record we took in July. Our July 25 issue carried a different reading from the same firm: AMD chasing deals with equity rebates as high as 105% (a rebate worth more than the sale price itself), which, on that firm's own cost model, put the rack's cost per million tokens at "practically negative cost" (SemiAnalysis, 07-25). Inside a month, one measurer working from one cost model reached two opposite conclusions.

Judgment update: we have marked the two as contradicting each other, and we are not reconciling them on anyone's behalf. They may be two faces of one coin: July explains why AMD has to buy share with subsidy, and today explains that the subsidy has a limited reach — what it cannot buy back is hosting and power. They may also be genuinely exclusive: if the rebates run large enough to cover hosting and power, today's reading falls. Forcing them into agreement is easy, and the cost of doing it is that the next round of measurement can no longer tell you which one was right. So we are leaving it as an unresolved tension and waiting for a re-measurement after ROCm 7.14. The action you can copy is the same one: with any pair of opposing readings in hand, establish whether the two are even measuring on the same basis before deciding whether to reconcile them.

Investor note: bulls and bears each quote one of these as directional evidence. This collision says the two do not overlap in what they measure at all — one prices the deal structure around a specific large customer, the other the per-token cost at a specific model and speed point. That is a weakening for the assumption that subsidy can offset an efficiency gap indefinitely. It is a strengthening for the reading that the efficiency gap finally lands on hosting and power — on costs no rebate reaches. Neither is far enough along to conclude anything.

Also happened

1. [This week] (August 23) TSMC is reported to be planning to buy two panel plants from AUO, a Taiwanese display-panel maker, for NT$30B (about US$940M) and convert them into advanced packaging facilities, and Largan Precision, Taiwan's leading phone-camera lens maker, is reported to have finished developing a co-packaged optics module and to be joining NVIDIA's optical-communications supply chain — the poster flagged both as rumours himself, and the TSMC one is still in on-site due diligence, not closed (@dnystedt on TSMC, on Largan).
2. [This week] (August 23) An opinion piece in the data-centre trade press argues that the sector's reliance on lithium-ion battery storage is changing its risk profile while fire codes lag behind, citing the battery explosion and fire at South Korea's national data centre last September that interrupted more than 600 online government services — this is an opinion piece rather than reporting, and we have not checked the incidents it cites one by one (Data Center Dynamics).
3. [This week] (event dated August 21) NVIDIA's B200 overtook AMD's MI355X within three days, and no chip changed (SemiAnalysis, 08-24). ⚠️ Only one line fits today; the linked original has the detail.

Chips & semiconductors

[This week] (posted August 23) The cheapest slab of material at the top of the chain is reportedly going up in price for the first time in more than three years — and the packaging house spending the most money says what it lacks is construction time, not capital.

Dan Nystedt is an Asian semiconductor supply-chain reporter and analyst at TriOrient, and he relays Taiwanese Chinese-language business coverage of the supply chain on X. He posted several relays that day. First: silicon wafers (the starting point of chip manufacturing, a disc of high-purity single-crystal silicon on which every circuit is built, and the most generic base material at the very top of the chain) are said to be going up across 6-inch, 8-inch and 12-inch sizes, the first substantial increase in more than three years, on the grounds that AI demand has finally absorbed a long-standing supply overhang (@dnystedt, 08-23). ⚠️ Read the basis closely: the original says individual product prices still vary, and the expected effect on company revenue is a 10% rise in average selling price — not 10% on every wafer. Second: Taiwan's five largest packaging and test houses will spend more than NT$460B, about US$14.4B, in capital this year, with ASE (the largest packaging and test house in the world) accounting for NT$335B, about US$10.5B, on its own; but ASE says its biggest problem now is how to speed up construction, tool installation and capacity ramp across 13 new builds and 8 conversions (@dnystedt, 08-23). Third: graphics card makers have already raised third-quarter prices and are working on a second round, with both dies and card memory short at once; total available card supply this quarter is estimated to shrink at least another 20%, and the steepest rises are on entry-level and older-generation cards — the capacity data centres took away is being paid for at the cheapest end of the consumer market (@dnystedt, 08-23).

How to use it: read together, the three point one way — the binding constraint upstream is spreading out from leading-edge process in both directions, up into the most generic base material and down into consumer products, while the middle of the chain is no longer constrained by money. That last admission breaks a chain of reasoning the industry has leaned on for two years: a company announces a capital figure, and capacity is inferred from it. The company spending the most says its bottleneck is not the money. ⚠️ But these three are relays from one reporter on one day, not three independent sources, and he labels them as media reports; Taiwanese supply-chain news is also frequently a vehicle for vendors negotiating through the press. So the use we are recommending is a discounted one: put these on the question list for third-quarter earnings calls — how many builds have broken ground, how far the tool-installation schedule runs, what the ramp curve assumes — rather than revising capacity assumptions today.

Named commentary

[This week] (posted August 23) A practitioner-writer says the free lunch is over — and we are not treating it as evidence for the judgment in main line item 2.

Drew Breunig is an independent writer on data strategy and applied AI who has long covered the trade-offs between model capability and practical engineering. In "Fable & The End of the Free Lunch" he writes that before Fable, working hard on your own coding harness and context strategy did not pay, because a new model always arrived at the same price or cheaper and covered most of your problems for you. After Fable landed, it was strong but expensive enough that his team began thinking about which work should go where, since a cheaper tier is already good enough for most of their code (dbreunig.com, 08-23; we reached it through Simon Willison's citation page).

Why we took it, and why it is not evidence. It supplies a mechanism for the 8.0% in main line item 2 — not that the flagship will not sell, but that it is expensive enough to be worth tiering work across price points. That is a far more precise reading than "the flagship failed." But it is one person's and one team's first-hand experience, with no measurement, and the population is a team rather than a market, so it can neither prove nor disprove the judgment we declined. We take it as an account, not as evidence — and that distinction is the same thing item 2 is about. If you build coding-assistant products, his approach is itself something you can test: tier the work by difficulty, and cost out which tier goes to the flagship and which to a cheaper model. That is one team's practice, though, not our recommendation, and not a conclusion we have measured.

Model watch

[Today] (published August 24) A long-standing standard optimisation drops the cache hit rate from a theoretical 96.0% to a measured 28.8% under today's kind of traffic.

The same measurement holds one section you can act on directly. DP-attention splits the attention computation (the step that works out which words in a sentence relate to which) along data-parallel lines, and in single-turn fixed-length settings it usually pays for itself. Under long-context multi-turn traffic it turns negative. On MiniMax M3, a model from the Chinese AI lab MiniMax, at concurrency 40, throughput falls to 0.60x of not using it, and time to first token (the wait between pressing send and seeing the first word) degrades by more than 3x. At concurrency 32 the measured cache hit rate is 28.8% against a theoretical 96.0%. ⚠️ Those two figures come from different concurrency points (40 and 32), not from two columns of one measurement. The mechanism is stated plainly: each compute unit owns only a private quarter of the cache pool, so a 300,000-token conversation that lands on a different unit next time is recomputed from scratch (SemiAnalysis, 08-24).

Two things you can lift straight out: one, measure the cache hit rate before you look at throughput — once the hit rate collapses, every throughput number after it is wasted arithmetic. Two, if you want main-memory offload, main memory has to be substantially larger than the GPU's cache capacity, because the design copies on write: every entry written into GPU memory is written into main memory at the same time. Too small and you have effectively not enabled it. The measurer suggests 1.5 to 3 times, though the original does not explain how either end of that range was derived. ⚠️ Both come from today's single-source measurement, with no third-party re-run. Separately, the academic papers we swept this week held nothing new worth its own item, so this column has nothing new for you this issue — we do not dress an evergreen concept up as news.

Sources & accounting

The past 24 hours. The overnight routine swept in 192 pieces: 109 academic papers, same-day posts from 77 X accounts, 3 company and personal blog posts, 3 industry newsletters. One number needs explaining before you read further: our automated count came to 6 today. It misclassified last night's large intake as back-fetching of older material, so 186 pieces vanished from the day's figure and 6 were left. Those 186 are last night's normal output, not backfill — we are using the real figure of 192 here, we have put the misclassification on the list to fix, and we are not quietly adjusting a number. Of those 192, 4 were actually judged inside these 24 hours, none were excluded for breaching any rule, and the other 188 remain unread (all 186 are still there, still waiting to be judged). Of the 4 judged, 3 were taken into our records and 1 held no signal — an AI policy researcher's account, whose two posts that day covered British food and a baked potato, with nothing about the industry in either. Fetching something is not the same as taking it. The 3 newsletters came from 2 sources: 1 from SemiAnalysis and 2 from Gary Marcus's Marcus on AI (by the same logic, 3 is what we actually read last night, while the "46 newsletters" on the roster below is the long-term total we track; different populations). Among those 77 X accounts the heaviest posters were @bhorowitz with 26, @teortaxesTex with 23, @GaryMarcus with 15, @deanwball with 14 and @Miles_Brundage with 12. The only account we pulled anything from for this issue was @dnystedt with 9 — ninth by volume, not first. Overnight we pulled 290 original posts from 748 tracked accounts (reposts and replies were never on the list). 748, 77 and 290 measure three different things: 748 is accounts on the roster, 77 is accounts that actually posted yesterday, 290 is a count of posts. They neither add nor divide. This issue uses 16 clickable external receipts (our own methodology page, back issues and the account links above are not counted).

What you are not getting today. Seven things, said plainly. One, every judgment today that touches that measurement rests on its free public section — the paid section after "Read more" we did not read, and the methodology deep-dive, the replay framework details and the historical performance curves for each vendor all sit back there. Two, the revenue figure reached us only as a second-hand relay: the FT original is behind a paywall and we did not obtain the full text. Three, no macro section this issue: no new official macroeconomic data reached us overnight, and we do not pad a section with old figures. Four, no trend-lineage section this issue: the lineage material we hold has gone stale, today's new material does not connect to that line, and forcing the join would be a join made for the layout. Five, no product moves column this issue. No product news this issue. No product company published anything new in the past 48 hours, and we do not write news from an existing product's feature page. 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 piece reached us as an introduction and a single image only: Gary Marcus's "more bad news" post, whose body sits after "Read more" and which we did not read, so we have not even described what it argues — this is our fetch step failing, not a judgment that the piece does not matter, and it is on the list to fix. Seven, no new deep dive today: the most recent one is dated August 21, and we do not re-run an old one to fill space.

Backfilled material. No sources were added by hand as one-offs this issue. The long-running backfill that started July 1 (not counted inside the 24 hours above) remains a set of break points: academic papers and X posts reach August 23; blogs and industry newsletters reach August 22; company filings reach August 16; industry analysis and podcast transcripts both stop on July 7; macroeconomic 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 14 items taken into our records today, 12 came from two external sources — 7 from that research firm and 5 from that supply-chain reporter. Today was not a day of "signs from several directions." It was a day of two thick pieces of material, which is why that phrasing appears nowhere in this issue. The two carry different discounts. The research firm's items are different facets of one article and do not corroborate each other; the discount comes from a measurer reporting on its own measurements, plus its dealings with both companies being compared — offset by one thing nobody else has: what it published is open-source and can be re-run by a third party. The reporter's items are relays from one reporter on one day, labelled by him as media reports and rumours; the discount comes from the double relay plus the negotiate-through-the-press ecology of Taiwanese supply-chain coverage. Three lines taken on one day, three different grades of evidence, three different uses — that grading is itself the thing we are handing you.

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. 645 and 529 measure different things: one person may 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. Beyond those there are 77 institutional sources (official blogs of companies and institutions) — ⚠️ this 77 and the "77 accounts posted last night" above are a coincidence and unrelated: this one counts institutions on the roster, that one counts X accounts that actually produced something last night. The institutional category counts institutions rather than people, and does not fold into the 645. Representative names: on X, Lucas Beyer, Sergey Levine, Arvind Narayanan; in newsletters, Zvi Mowshowitz, Dean Ball, Ian Cutress; among paper authors, Ion Stoica, John Jumper, Yann LeCun; on podcasts, Satya Nadella, Sam Altman, Demis Hassabis; among institutional blogs, SemiAnalysis, Data Center Dynamics, More Than Moore.

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 · 16 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 →