Daily Brief Nov 11, 2026 · ← Back to this issue
How we checked the items below
For readers who want to dig in. For what happened, in brief, go back to this issue.
This issue draws on material from November 10 to early November 11. Last night's collection took in 795 pieces, 129 of them read in full, and this issue uses 29 outside sources with checkable links; the later sections also draw on material from the past week. The full edition of this issue is the archive of record.
1. [Today] TSMC will buy Anthropic shares at the IPO price and OpenAI shares at an expected price of about US$992 each, capped at US$500 million per company
Not proven yet： the OpenAI purchase agreement isn't public, so terms the filing omits can't be ruled out
How we checked
The main sources are four material-information filings TSMC made with the Taiwan Stock Exchange: the board resolution, the Anthropic investment, the OpenAI investment and the capital budget. The Anthropic side comes from the cover page, offering summary, Concurrent Private Placement section and AMD Agreement section of the amended IPO registration Anthropic filed with the US Securities and Exchange Commission. The two companies filed separately, and their filings match on TSMC Partners buying about US$500 million at the IPO price. The OpenAI purchase rests on TSMC's filing alone. What was approved is a cap, not money already paid. Both filings list the "relationship with the company" as "none" and the purpose as "equity investment." Our calculation: 504,000 shares at US$992 come to about US$500 million, which matches the cap.
2. [Today] Anthropic launches its IPO roadshow, and seven chip and cloud suppliers will separately buy US$26 billion of shares at the IPO price
Not proven yet： the price is still an estimated range, and the registration isn't effective yet
How we checked
The main sources are Anthropic's press release and the second amended IPO registration Anthropic filed with the SEC on November 10, with exhibit EX-10.15. For the seven companies' terms, the amendment's exhibit EX-10.16 is only a blank template with no investors or amounts filled in. The price is still an estimated range and the registration isn't effective, so the figures may change. The seven amounts come from the offering summary and are shown at the US$74 midpoint. Of the 825 million shares in the public offering, 23,648,646 are sold by existing shareholders, and Anthropic receives none of that money. Our calculation: 801,351,354 new company shares at US$74 is about US$59.3 billion. Adding the seven companies' US$26 billion gives US$85.3 billion, of which US$26 billion is about 30%. The amendment also states that at the midpoint, Anthropic expects net proceeds of about US$84.6 billion from the offering plus these purchases, after underwriting discounts and expenses.
3. [Today] Google opens Gemini 4 Carbon at Argon's introductory API price, doubling after November 30, and Carbon ranks second on a third-party-run knowledge-work benchmark
Not proven yet： the release post doesn't say whether the cost comparison uses the introductory or doubled price
How we checked
The main sources are Google DeepMind's release post and Artificial Analysis's MKA page. MKA was designed by Google DeepMind and is run by Artificial Analysis, with 200 tasks across 10 industries. Google's own red team tested in a simulated cyberattack environment, and the attacker's per-attempt breach rate was 0.8% for Carbon against 3.1% for Argon. According to the release post, those security results measure the Carbon version offered first to cyber defenders through the Fairwind program, not the general API version. Google also says Carbon ranks first on benchmarks at Cursor, Perplexity and Zapier, at under half the cost per task of the next-best outside model. The post names neither that model nor the scores. The release post doesn't say whether the cost comparison uses the introductory price or the price after November 30. If it uses the introductory price, the cost advantage shrinks once the price doubles. The release post doesn't give each benchmark's scores, the names of the compared models or the context length. Google published a model card and a safety report alongside it.
4. [Today] CNBC and The Information report on the same day that OpenAI's third-quarter revenue reached US$10 billion, with no OpenAI filing to check it against
Not proven yet： both outlets call their reports exclusive, and whether the sources overlap is unknown
How we checked
The main sources are Bosa's post on X and the headline of The Information's report, both media reports. Both outlets call their reports exclusive and neither says whether the sources are the same people, so they can't count as two independent sources. OpenAI isn't public and has no filings to check against. Neither the post nor the headline gives the loss amount. Anthropic's US$17 billion and 47% come from Bosa's account.
[Today] TSMC's board approved a capital budget of about US$40.9 billion, and the filing groups advanced packaging with mature and specialty processes in one line.
How we checked
(shares are our calculation; Taiwan Stock Exchange, 2026-11-10) (Taiwan Stock Exchange, 2026-11-10)
[Today] AMD's sixth-generation EPYC server processors are shipping, and AMD says integer throughput reaches up to 3.7 times that of NVIDIA's Vera server processor, a comparison using AMD's internal estimate for Vera.
How we checked
In AMD's comparison only the AMD and Intel scores are results published on spec.org, while the scores for NVIDIA's Vera and for AGI, an Arm-designed server processor, are AMD's estimates and can't serve as a third-party benchmark. The metric is SPECrate2017_int_base, which measures integer throughput, on two-socket systems throughout: EPYC 9996 scores 5,470, Intel's Xeon 6980P 2,510, and AMD estimates Vera at 1,459. AMD's other figure, "agents per watt," is estimated from processor thread counts according to AMD's footnote, not measured by running agents.
[Today] Sayash Kapoor, author of the newsletter Normal Tech, announced STABL, a nonprofit that evaluates whether agents can do open-ended research, and says papers written by frontier agents were clearly rejected by expert reviewers.
How we checked
"The performance gap between automatically verifiable tasks and open-ended tasks is widening" is the judgment of Kapoor and coauthors, based on CRUX, an evaluation the authors run themselves, and the post doesn't give the sample or the models tested.
[Today] Nathan Lambert, author of the newsletter Interconnects, says publicly that he doesn't expect an "intelligence explosion" by early 2028.
How we checked
This is Lambert's personal judgment with no data attached, and Lambert also writes that he isn't sure how an intelligence explosion should be measured.
(Nathan Lambert, 2026-11-10) (Tamay Besiroglu, 2026-11-10) (Daniel Kokotajlo, 2026-11-05)
[Today] A preprint recomputes public scores on eight agent benchmarks and says the crowding at the top of the leaderboards comes from how they are scored.
How we checked
No other team has adopted this scoring yet, and the abstract doesn't say how the test error was estimated or which quantity Terminal-Bench was compared on (arXiv 2611.13541, 2026-11-10).
[This week] A preprint holds model, tasks and tools fixed and changes only the agent architecture; multi-agent setups use 1.4 to 4.7 times the tokens of a single agent, mostly with no significant difference in success rate.
How we checked
According to the paper's abstract, the only setup that significantly beat a single agent was a redesigned multi-agent implementation, a result the authors report and no one else has repeated. The abstract doesn't say which model was used (arXiv 2611.12953, 2026-11-07).
[Today] NVIDIA open-sources Snapshot, which saves an initialized GPU inference replica as a checkpoint and restores it, and design partner Crusoe's own test cut time-to-ready from about 8 minutes to 31 seconds.
How we checked
[Looking back] (SecondSource reading of July 14) AI agents' memory currently sits outside the model: stored as text and placed in the input window each time, not written directly into the model's weights, the model's own parameters.
How we checked
Still missing today. The gaps that most affect judgment are two sets of contracts: the OpenAI–TSMC share purchase agreement isn't public, and for the six companies other than Google, Anthropic's amended registration files only a blank agreement template, exhibit EX-10.16, with no investors or amounts filled in. The amended registration doesn't state Anthropic's market value after listing directly, and the post-offering share count has two possible bases that haven't been settled, so this issue doesn't print a conversion. Neither CNBC's post nor The Information's headline gives OpenAI's loss amount or revenue recognition method. Gardizy's post doesn't give the number of turbines Son reserved, the suppliers or what happened to them afterward. TSMC's capital budget has no separate figure for advanced packaging and no line-item breakdown of past budgets to compare. The quarterly revenue and backlog figures in CoreWeave's open letter haven't been checked against its formal financial statements, so they aren't printed.
Sources & accounting
The past 24 hours. 795 new pieces came in from November 10 to November 11: 489 academic papers, 151 social-platform posts, 73 company and personal blog posts (including Taiwan Stock Exchange material-information filings), 55 from a second academic-paper feed (counted separately from the academic papers), 12 company filings, 6 subscription newsletters, 6 linked articles, 2 macro data releases and 1 show transcript. Of those 795, 129 were read in full, 161 were screened out as off-topic (14 of them before fetching), 12 were only indexed without a full read, 2 were held for another try, and 491 haven't been read yet. Academic papers, 489: 54 read, 98 screened out, 337 not yet read; last night arXiv rate-limited the keyword fetch, 5 of 37 keywords weren't fetched, and paper coverage is incomplete. Social-platform posts, 151: 41 read, 38 screened out, 5 only indexed, 67 not yet read. Blogs, 73: 21 read, 4 screened out, 3 only indexed, 1 held for another try, 44 not yet read. Second paper feed, 55: 14 screened out (all 14 before fetching), 41 not yet read. Company filings, 12: 2 read, 3 screened out, 4 only indexed, 1 held for another try, 2 not yet read. Newsletters, 6: 5 read, 1 screened out. Linked articles, 6: 3 read, 3 screened out. The 2 macro data releases and 1 show transcript were all read.
One-time catch-up. No new one-time catch-up material came in this window. The catch-up material already collected is dated between July 1 and November 9 this year, counted separately from the past 24 hours above, and none of it was used in this issue.
Source concentration. Main-line item 1 and Chips & semiconductors item 1 come from filings on the same TSMC board meeting, all disclosed by TSMC alone; where main-line items 1 and 2 match each other, the match comes from separate filings by two issuers, TSMC and Anthropic. Main-line item 3 comes from Google's own release, with only MKA run by a third party. The two outlets in main-line item 4 both call their reports exclusive and don't count as two independent sources. Also happened item 3 comes through reporter Dan Nystedt, who also made one of the two transcripts of the July remark in main-line item 1. Both Product moves items come from the same NVIDIA blog post.
Tracked sources. The long-term roster currently has 529 named sources: 302 on social platforms, 90 shows, 51 news outlets, 48 blogs, 48 paper authors and 46 newsletters, with the rest spread across earnings reports, keynotes and other channels. The same source may be tracked on several channels, so the channel figures add up to more than 529. These are counts of tracked sources, a different list from last night's piece counts above, so the two can't be added or compared. Representative names include Daniel Kokotajlo, Sergey Levine, Arvind Narayanan and Chris Lattner on social platforms; Sam Altman, Dario Amodei and Demis Hassabis among show interviewees; and Lisa Su and Satya Nadella on the earnings roster. This issue uses 29 outside sources in the body. Only outside links actually cited in the body are counted. Sources from Taiwan total 210, with 43 added in October; Taiwanese media, Taiwanese companies' official channels and local reporters in Taiwan count, while foreign outlets relaying Taiwanese media do not.