Daily Brief SecondSource Morning Brief · August 28, 2026 · Aug 28, 2026
1. Delay the frontier labs' best models and the first casualty is not safety — it is what they can afford to bid for 2028 compute.
2. One watt of power carries four different prices along the AI chain, and a six-year lease signed this week gives the lowest of them its first outside check.
3. A column arguing that AI revenue is accelerating has footnotes that say it is slowing. Check the ruler before you compare the numbers.
This issue draws on the research digest our system produced on August 28, plus six company announcements and industry reports caught overnight that we read in the original. The events run from August 25 to 27, and each item carries its original date at the head of the sentence. The overnight sweep took in 381 pieces, and 14 clickable receipts made it into this issue. This is the email edition; the full edition of this issue is the archive of record.
Dylan Patel spent more than two hours on the compute supply chain on August 25, on the long-form interview show Dwarkesh Podcast (the episode, 08-25). He founded SemiAnalysis, a research firm covering semiconductors and data centres. Declare the interest first: SemiAnalysis sells compute-market research for a living, so it has a direct commercial stake in the story that compute is scarce and capital spending is enormous. Most of his figures are his own firm's estimates rather than audited disclosure. He laid out a chain with four links and never joined them up. We joined them.
Link one: concentration has to be bought. OpenAI and Anthropic, the two frontier model developers, go from roughly a third of this year's new compute to seventy or eighty percent of it by 2028. The population is compute that comes online in a given year, attributed to the end customer: capacity somebody else builds and leases to them still counts as theirs.
Link two: today's price will not buy that. Data centres are sized in electricity because power is the hardest physical ceiling, and revenue per MW per year is simply how much money one unit of power capacity earns in a year. He was blunt about it. At US$10M to US$15M per MW anyone can make money — buy a rack, download publicly available model files, stand up software to serve them, list on a third-party resale platform, and the arbitrage is there. Winning instead means pushing the clearing price up, and the winning levels he gave are three separate points, US$25M, US$30M and US$50M per MW, not a continuous range.
Link three is the hinge. The only way to afford that price is to monetise better than everyone else, and monetising better requires holding a model nobody else has. Right now they are withholding exactly that. In his words: "They're clearly not releasing their best models, in which case their revenue per megawatt stalls or can even start to decline again because other models are competitive again." He named three instances: OpenAI not shipping Astra, a codename for an unreleased model; OpenAI halting training for two weeks; Anthropic not shipping the next version its own safety evaluation refers to. Link four: once bidding power shrinks, concentration decelerates on its own.
Verification: all four links come from one man on one episode, and we did not independently confirm any of those three release instances today. The independent second angle sits somewhere else entirely. AI policy analyst Dean Ball reached the same causal starting point from the regulatory side in What Should Be Done, published June 26 (the essay, 06-26). Frontier model profit, he argues, is a narrow-window business: training cost has to be earned back within months of release, so every week of delay removes a week from the window. It is not a fixed cost you pay once and forget. The two have nothing to do with each other and used different methods, one a regulatory analysis and one a supply-chain measurement. ⚠️ Ball had already announced he was joining OpenAI when he wrote it, and had not yet started; the argument runs directly in favour of his incoming employer, and we logged that in our August 3 issue. Ball stops at the labs' income statement. Walking from there down to the ownership structure of compute is ours. ⚠️ The boundary that matters most: not one link in this chain carries a coefficient. There is no such thing as "revenue per MW stalling by X percent costs them Y points of share" — this is a mechanism argument, not a measurement.
Judgment update: a regulatory story that reads on its face as "launch delayed by six months" actually cuts the frontier labs' bidding power in the compute auction, and that is the variable deciding who owns the world's compute in 2028. The policy lands asymmetrically. If your goal includes "do not let two companies control most of the world's compute," this is an unintended bonus. If your goal is "the West stays ahead," it is an unintended penalty. Those two goals collide head-on along this chain, and public debate treats them as pointing the same way. One question to put to your own team: does our compute demand forecast have a column for when the major labs ship their flagship models? What would prove this wrong: if open-weight and second-tier models keep failing to close the gap on what a model actually knows, then withholding releases will not loosen frontier revenue per MW at all, and the chain breaks at link three. Honestly, the current evidence leans that way. Third-party evaluator Artificial Analysis publishes AA-Omniscience, which measures factual correctness, and it reads 28% for GLM-5.3-Flash, a Chinese open-weight model, against 47% for GPT-5.6 Terra; ⚠️ we have those figures via a community summary and never reached the original results page.
Investor note: the prevailing story treats AI regulation as a question of compliance cost and launch schedules. This evidence ties it instead to the price of buying compute. It weakens the claim that frontier lab compute spending grows in a straight line, and it strengthens the relative standing of the second tier and the open-weight camp.
In that same interview, Patel laid the compute value chain out in four layers (the episode, 08-25). Layer one, what the compute buyer pays: US$10M to US$15M per MW. Layer two, what a lab monetises: as high as US$50M per MW in Anthropic's case, which is a ceiling rather than an average, and he never said which quarter. Layer three, the sublease: US$25M to US$40M per MW, his example being SpaceX reselling data-centre-grade GPUs it held to Google. Layer four, what the end customer captures: US$200M to US$500M per MW, his example being Jane Street, the American quantitative proprietary trading firm — an extreme sample, not a typical enterprise.
Here is what is genuinely new today. Bloomberg reported on August 26 that Anthropic has signed a US$45B AI cloud capacity agreement with Nscale, leasing 460MW over six years at Nscale's West Virginia campus (Bloomberg, 08-26; relayed by data centre trade outlet DCD, 08-27). Nscale is one of the new clouds built specifically to rent AI compute: it builds the halls, buys the GPUs and leases capacity to labs. Microsoft has signed for 1.35GW on the same campus. We did the arithmetic ourselves: US$45B ÷ 460MW ÷ 6 years ≈ US$16.3M per MW per year.
Verification: that US$16.3M is our division, not anyone's claim, and three definitional uncertainties come with it. Whether the contract includes GPUs is unstated. Whether 460MW is IT load or total facility capacity is unstated, and the two typically differ by twenty to forty percent, enough to put the figure anywhere between US$12M and US$20M per MW. Whether the payments are level across six years is unstated too. Even with all three laid open, the direction still matches. It sits a little above the top of Patel's "anyone can make money" band, which he says explicitly is now rising, and it comes from a completely independent reporting chain published a day after the interview. ⚠️ DCD says it approached Nscale and has not obtained confirmation, and we did not independently verify it either. ⚠️ The point to be blunt about: layers two, three and four still rest on this one speaker today, with not a single independent reading between them. Only the bottom layer got an outside check. NVIDIA's CFO gave a per-GW figure in the same week, pointing the same way but measuring something else; we handle that below in Chips & semiconductors.
Judgment update: "compute is expensive" is not one number, it is four. Before any per-MW figure goes into a cost model, ask four things: who pays it, who books it, does it include GPUs, and is it an annual revenue measure. And the roughly fourfold gap between layer one and layer two has a low enough barrier that anyone can arbitrage it, which makes it unstable. What is filling it in is not new technology, it is people with spare inventory reselling it. What would prove this wrong: if the next few comparable contracts work out to a per-MW annual price well below US$16M, then our division here mistook GPU-free rent for a GPU-inclusive price.
Investor note: the market treats revenue per MW as a single cost curve you can extrapolate. This contract says it is four independent price lines. It strengthens the claim that compute clearing prices are rising, and it weakens any attempt to estimate frontier lab gross margin from a single per-MW number.
Epoch AI is a non-profit that tracks AI compute, capability and economic indicators, publishing its methods and datasets, and on August 27 it ran a column on AI revenue (Epoch AI, 08-27). Two things about its status: first, it appears in what Epoch itself labels its more opinionated column and does not represent an institutional position; second, the piece is a second-hand compilation whose period revenue figures are relayed from other outlets.
The headline numbers are bright. OpenAI's annualised revenue went from US$13B to more than US$40B inside a year; annualised revenue multiplies a recent period by twelve, so when growth is fast it runs far above the money actually booked that year. Anthropic went from US$1B to US$9B across 2025, reaching US$65B by the end of July. Together the two went from US$30B to US$105B year to date, a 3.5-fold rise, and the author adds that it is only August.
But its own footnote 2 says the two are not measured with the same ruler. For services sold through a cloud platform, OpenAI books only its own revenue share while Anthropic books the gross. In the original: "which inflates Anthropic's reported revenue relative to OpenAI's." Footnote 4 then says "The growth rate within 2026 may be slowing down": the pair stood at roughly US$55B at the end of the first quarter, and the first quarter itself nearly doubled, but the next doubling took about four months.
Verification: we went looking for a second source and came back with an intermediate point Epoch does not give — Anthropic at US$47B in May 2026 (CNBC, 08-17). With that you can do the arithmetic yourself. May's US$47B to July's US$65B is 38% over two months, a doubling time of about 4.3 months; US$9B at the end of 2025 to US$47B in May is more than fivefold over five months, a doubling time of about 2.1 months. The doubling time really is lengthening, and that comes from a different set of numbers reached by a different calculation. The same report independently carries the accounting caveat as well — "the two companies may not measure revenue the same way" — and that is a separate line from the reporting chain Epoch cites. ⚠️ The honest boundary in full: the doubling arithmetic above is Anthropic alone, while Epoch's footnote 4 is about the two combined, and with different accounting definitions their rhythms need not move together. We claim only that the slowdown reading has independent support; we do not claim to have reproduced Epoch's calculation. ⚠️ The US$105B combined figure, and the piece's separate note that the whole generative AI market is running at close to US$200B annualised, still rest on this one house.
Judgment update: both readings are correct, and the difference is purely the ruler. Year over year shows acceleration because the 2025 base is low. Within-year rhythm shows deceleration. Only the second is about now. Two things you can do today. First, confirm the accounting definitions before you put the two companies' revenues side by side. Second, look at the within-year doubling time rather than a year-over-year multiple, which last year's low base inflates. Read alongside item 2 of today's main line: that one measures how much money a unit of compute generates, this one measures the ruler those revenues are reported with — and when the rulers do not line up, the per-MW monetisation figures distort with them. What would prove this wrong: either company breaking AI revenue into line items by accounting definition, or a clear reacceleration in the third quarter.
Investor note: the table where you allocate capital by AI revenue growth rate needs one more column today, for the definition. On "AI demand is still growing fast" this is unchanged; on "the growth rate itself is still accelerating" it is weakened.
The Information, a subscription tech outlet, broke the story yesterday that NVIDIA has agreed to acquire Hugging Face for US$12.9B, and CNBC, TechCrunch and Bloomberg each relayed it (CNBC, 08-27). Hugging Face is the world's principal platform for publishing and distributing open model weights, often described as GitHub for AI models, and a great many companies and open-source projects treat it as neutral infrastructure. NVIDIA was already a shareholder from the 2023 round, when the valuation was US$4.5B, so US$12.9B is about 2.9 times the price three years ago. But as of today there is no confirmation at an official level: no NVIDIA announcement, and no 8-K filed with the US Securities and Exchange Commission to report a material event. The Information itself notes the agreement is unsigned and could still fall apart. The three relays word it as "agreed to acquire," "close to completing" and "in discussions" — three different degrees of certainty.
Our own correction belongs here. The source that first brought this to us was Latent Space AINews, an automatically compiled digest of community discussion, whose headline said US$13B and "roughly 80x their US$150M ARR" (the digest, 08-27). On checking, the amount is US$12.9B; and if that US$150M of annual revenue holds, the multiple is 86x, not 80x. Dividing its own US$13B by US$150M does not produce 80x either — it lands around 86x as well. The gap between 80 and 86, in other words, is not something we created by revising the price down — the source's own arithmetic does not add up. ⚠️ We did not independently verify that US$150M annual revenue figure today either. The same digest also says NVIDIA's initial offer in January this year was US$7B; that number appears in none of the mainstream relays, and we are not accepting it.
Verification: four relays sharing one exclusive are not two independent sources, so we treat this by our single-source standard. "Reported" is verified. "Completed" is not. We are not ruling on where the deal goes today, only recording the facts.
Judgment update: this is the classic shape of a rounded number being cited downstream as a precise one, and today it is happening to a figure that will be pasted into countless slides. One thing you can do today: reassess any plan built on Hugging Face as neutral infrastructure — it now carries single-vendor risk — rather than waiting for a closing announcement. What would prove this wrong: an official announcement or filing from either party, or The Information retracting.
Investor note: the story treats the open-source ecosystem as a commons in the competition at the model layer. This evidence turns ownership of the publishing and distribution channel into a variable. It weakens the claim that open source keeps compressing frontier pricing, and it strengthens the claim that the infrastructure layer keeps consolidating upward.
1. [This week] (08-25 to 26) Independent evaluator Artificial Analysis has taken apart the "intelligence per dollar" advantage of the Chinese open model GLM-5.3-Flash: it burned 149 million output characters getting through the evaluation, more than the Chinese open models Kimi K3 (133 million) or Qwen3.8 (136 million) at the same score, so the cheapness comes from extremely low per-character pricing, not from using fewer characters. Efficiency is an asset and pricing is a policy — do not extrapolate its measured US$0.09 per task into a cost curve; ⚠️ relayed via a community summary, with no access to the original results page (relayed by Latent Space AINews, 08-27).
2. [This week] (08-27) AWS has agreed to deploy another 2 million NVIDIA GPUs across 2027 and 2028, against what it said in March about bringing more than 1 million online during 2026; the set of buyers is widening, and it is not just two companies (DCD, 08-27).
3. [This week] (08-27) Georgia regulators have approved Georgia Power building 3.2GW of new generation for OpenAI, with OpenAI bearing the full cost; ⚠️ delivery runs from 2028 to 2032, so this is not near-term supply (DCD, 08-27).
4. [This week] (08-27) A Salesforce survey of 2,025 decision-makers on AI agents — systems that carry a task through on their own — across 20 countries found that companies with lighter governance average 7.2 months to positive return, against 9.3 months for heavier governance. But among those with below-average governance, 32% discovered an agent had exceeded its intended scope only after something went wrong, against 18% of those above average; ⚠️ a vendor's own survey, with all outcomes self-reported (Salesforce, 08-27).
Deep dive: there is no new one this week. We publish a deep dive only when the evidence is strong enough, and never to a schedule. The last one — That warrant was Marvell's discount to Google, and it prices out at about 6% of the purchase commitment, August 21, 2026 — exists in the Traditional Chinese edition only; no English edition of it has been published, so there is nothing here to send you to.
[This week] (earnings reported August 27) NVIDIA's CFO gave two numbers, one supporting the case that compute is concentrating into two labs, one arguing against it.
NVIDIA's latest quarter came in at US$96.2B of revenue, up 106% year over year, with US$89B from the data centre segment and a company-wide gross margin of 75% (data centre trade outlet DCD, 08-27). CFO Colette Kress gave two figures that map directly onto today's main line. Supporting concentration: "we expect demand from the AI labs for which we expect to leverage our balance sheet to contribute toward roughly a quarter of our business next year." Arguing against it: hyperscaler revenue was US$49B, up 13%, while the separately reported "AI clouds, industrial and enterprise" segment was US$40B, up 25%, growing faster than the hyperscalers. CEO Jensen Huang makes the same point: "This time last year, one lab alone was driving the buildout; today, we have a golden age of new AI labs and startups, multiple frontier labs scaling in parallel, a thriving open-model ecosystem and physical AI coming online."
⚠️ Three definitions to pin down first. One, that 75% is the company-wide gross margin, not the data centre segment's. Two, her separate line about revenue from each GW of data centre rising from US$30B to US$60B over five years measures NVIDIA's full-stack take from one build-out, and it is not the same ruler as the revenue-per-MW-per-year figure in item 2 of today's main line; the two cannot be set side by side and subtracted. Three, the "next year" she refers to is NVIDIA's own fiscal year, which does not line up with the calendar year, and we have not worked out the mapping. ⚠️ There is an interest here too: a seller financing its buyers decouples the demand reading from real end demand, and that is precisely what "leverage our balance sheet" means.
What this means for you: concentration and dispersion now have evidence pointing both ways from a single statutory disclosure, for the first time, and the dispersion side is the only material in this issue that puts reverse pressure on the main line. An anchor you can watch yourself: next quarter, check whether "AI clouds, industrial and enterprise" keeps growing faster than the hyperscalers. If it does, the concentration story in items 1 and 2 has to be pulled back.
No named commentary this issue. The only named speaker we judged today was Dylan Patel, whose whole argument is worked through in items 1 and 2 of the main line and is not replayed here. None of the other named views the overnight sweep turned up carried new evidence.
[This week] (published August 27) For the first time, someone has completed an outside evaluation of a commercial frontier model without handing over the questions or the model weights.
Google DeepMind announced yesterday that it has completed a double-blind evaluation with outside parties (Google DeepMind, 08-27). The old dilemma was a choice of two: in a high-stakes external evaluation, either the evaluator hands over the questions and the model vendor may see the exam in advance, or the vendor hands over the weights and its intellectual property may leak. This time each side's asset was locked into a confidential computing environment on Google Cloud. The hardware first proves which code is about to run. Both parties review and approve before releasing their asset into it, and only the agreed output comes back out. The evaluator never sees the model weights, and Google never sees the questions. Cryptography's job here is proving the isolation actually held, not implementing the double blind itself. Also taking part were Singapore's official AI safety institute, the open-source community OpenMined, frontier AI audit organisation AVERI, and benchmark standards body MLCommons. AVERI's own announcement adds that the model under test was Gemini 2.5 Flash-Lite, a version number DeepMind's post omits, and that the questions came from previously unused prompts in MLCommons' safety benchmark (AVERI, 08-27). ⚠️ Four honest markers: what was announced is a method, with no evaluation results of any kind; AVERI describes it as a small-scale quantitative assessment; the two sides publish inconsistent partner lists; and "world's first" is a first-party claim.
What this means for you: anyone writing evaluation requirements now has a workable practice to cite. AVERI's announcement notes that current rules "rightly demand that security and privacy be respected in third-party assessments, but provide little guidance on how to achieve this," and this pilot is an attempt to supply the how. An anchor you can watch yourself: when the technical report goes public, and whether it contains actual evaluation scores. Until it does, this is a demonstration of method and nothing more.
No product news this issue. The only qualifying change overnight was Cursor dropping the requirement to connect GitHub first for its cloud agents and starting to host code itself. That is a real change of direction, but we cannot say whose decision it changes, or how — a vendor's own announcement, no numbers, no third-party reading — so by our own rules we are leaving it out until adoption figures exist.
No archive pick this issue. We have used up the older material worth reusing. We would rather leave the column empty than replay an item we have already run.
The past 24 hours. Last night's sweep took in 381 pieces: 141 X posts, 133 academic papers, 50 company and personal blog posts, 49 unclassified papers, 3 industry newsletters, 3 company filings and 2 podcast transcripts. X has a separate set of readings of its own: 374 accounts reached, 497 native posts pulled back, zero reposts and zero replies. 497 and 141 count different things: the first is what was pulled back, the second is what was newly added to the database. Where the 356 in between went is something we have not worked out today, so we are printing both numbers as they stand. Against that population of 497, the heaviest posters were @TheStalwart with 27, @geoffreyirving with 26, @bhorowitz with 26, @teortaxesTex with 17 and @TheZvi with 14. Across newsletters and industry analysis we read 8 pieces from 6 houses (Newcomer 2, Zvi 2, Epoch Gradient 1, Gary Marcus 1, Latent Space 1, Stratechery 1), of which 3 were classified as industry newsletters and the rest came through their own channels; 2 podcasts (Dwarkesh, No Priors); and filings from 3 companies (Applied Materials, Lam Research, Marvell). Five pieces were read all the way through and judged by hand today, yielding 7 new facts and 2 new judgments, plus 4 decisions not to take something in and 2 where no quotable original passage could be extracted. ⚠️ Another 47 pieces had their body text pulled out automatically at the moment they arrived, 45 of them X posts, but nobody has looked at them or judged whether they are worth taking in; we count only the 5 read by a person.
What you are not getting today. The first and most important: today's four main-line items are not the best four out of 381, they grew out of the 5 pieces a person actually read — the other 333 never went through our judgment about whether to take them in. The six further pieces of material used in this issue — the Nscale contract, NVIDIA's earnings, the AWS GPUs, the Georgia power case, DeepMind's double-blind evaluation, the Salesforce survey — we read in the original while writing this, and the limits of each are written into its own item. The second: two company announcements that looked most relevant overnight came back from their URLs as a headline with no body text, so not a word of them is written here. They were Anthropic's "Model Hardware Standard" research preview and NVIDIA's "Vera CPU begins shipping." If the first is real, it is a frontier lab proposing a hardware standard, which may be the single most relevant thing to today's main line. The third: Ben Thompson's Stratechery column on Apple's and OpenAI's hardware launches, of which we obtained only the opening paragraph, the rest being behind the paywall. The fourth: nobody read a single one of the 133 academic papers or 49 unclassified papers from last night's sweep in the original today. Older material added by hand: none today. The database separately holds material backfilled from July through mid-August, which is not from last night's sweep, and this issue cites none of it.
Source concentration. Three things we have to say about ourselves. First, the single largest source in this issue is one man on one episode. Dylan Patel accounts for six of the ten items entering our long-term tracking today, far past our one-third warning line, so our own rule says to state it plainly: this issue leans on him. The mitigation is finding independent readings to set beside him, and today there are three. The Nscale contract gives the bottom of his price ladder an outside check (main line item 2). NVIDIA's statutory disclosure gives his concentration argument one supporting and one contradicting figure (Chips & semiconductors). And Dean Ball's June essay reaches the same causal starting point by a different method (main line item 1). Second, each speaker's interest is marked inside the item itself: Patel sells compute-market research, Dean Ball had already been hired by OpenAI when he wrote that piece, and NVIDIA is talking about its own quarter. Third, this issue contains no sentence of the form "several parties are saying it."
The sources we track. After de-duplication the roster covers 529 sources; a separate ledger organised by channel holds 722 records: 302 X accounts, 90 podcasts, 77 institutional and company blogs, 51 outlets and press rooms, 48 personal blogs, 48 paper authors, 46 newsletters, and 60 others covering earnings calls, keynotes, books and government documents. One person may hold an X account, have appeared on a podcast and have published a paper, and gets counted three times: 722 counts records, 529 counts sources after de-duplication, and the two do not add together. The same applies to last night's 374 X accounts reached, which is not the 302 X sources on the roster: the first is how many accounts actually returned data that night, the second is what we track over time. Newsletters work the same way — 3 read last night is a daily quantity, 46 is the full roster. Four readings, four different populations: this issue uses 14 clickable receipts, last night's sweep took in 381 pieces, the roster holds 529 sources after de-duplication, and the channel ledger holds 722 source records.
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.