Daily Brief Nov 15, 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 13 to early November 15. Last night's collection took in 391 new pieces, 90 of them read in full, and this issue uses 27 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] Taiwanese media report that TSMC received NT$167.6 billion in government subsidies in the first nine months, almost all of it in the third quarter
Not proven yet: Arizona's profit depends on the wafer prices it charges the TSMC parent, so it doesn't show that US costs are competitive
How we checked
All the figures come from the same TSMC financial report; the three outlets relay one document, so they don't count as independent checks. The Arizona figure is the subsidiary's own profit and loss, not US revenue in TSMC's consolidated statements. First-half subsidies were only NT$590 million. After TSMC's large third-quarter subsidies, the Arizona subsidiary's third-quarter profit was about NT$17.92 billion, still below the first quarter's NT$18.808 billion. Our inference: the quarterly figures show no sign that this year's subsidies went into Arizona's profit in one go. The reports don't say how the subsidies are booked. Last year's nine-month subsidies were about NT$71.9 billion; that figure comes from Nystedt's post, and the Central News Agency and Liberty Times reports don't carry it (Dan Nystedt, 2026-11-14).
2. [Today] Ethan Mollick's count: on 63% of 173 AI benchmarks the best reported score reaches 85%, against 24% when only third-party scores count
Not proven yet: Mollick chose both the 85% bar and the rule for which benchmarks count
How we checked
The above is based on SatBench's documentation on GitHub and Mollick's post. SatBench is Mollick's personal compilation and hasn't been peer reviewed. Mollick chose both the 85% bar and the inclusion rule, and a different line would change the shares; a benchmark is included once lab release posts have mentioned it 10 times. The documentation lists two biases that point in opposite directions: frequently cited benchmarks are more likely to be ones where labs can show good results, while recent benchmarks are undercounted because they haven't yet been mentioned enough. The underlying scores come from the Benchmark Almanac maintained by Brian Zhou (Benchmark Almanac), with data through November 14. The code and underlying data are public, so a recalculation with a different bar is possible; no one else's recalculation has turned up yet.
[This week] (earnings call, November 10) CoreWeave, a cloud provider that rents out GPU capacity, says third-quarter rental prices for NVIDIA's A100 and H100, two older GPU generations, were 30% to 50% higher than in the second quarter and above the original contract prices.
How we checked
Amazon's accounting estimate points the other way. The two measure different things: one is a rental price, the other an accounting estimate, and they took effect more than a year apart.
(CoreWeave earnings call transcript, 2026-11-10) (Amazon annual report, 2026-02-06)
[Today] Taiwanese media report that AI server maker Gigabyte posted record quarterly net profit of NT$8.57 billion in the third quarter and is raising capital spending from between NT$2.8 billion and NT$3 billion this year to NT$4 billion in 2027.
How we checked
Both reports come from the same earnings call, and the figures are as relayed by the media.
(Economic Daily News, 2026-11-14; Liberty Times, 2026-11-13)
[Today] Elon Musk says many of the compute deals between Musk and Jensen Huang were verbal agreements, with no written contract.
How we checked
[Today] Dylan Patel, founder of the semiconductor research firm SemiAnalysis, argues that the US should not let China buy AI chips, and should instead let US cloud providers rent compute to Chinese companies while keeping the power to switch it off at any time.
How we checked
[This week] A conference paper from the Indian voice-AI company Gnani.AI reports that scores on the MATH math benchmark jumped after reinforcement learning, and the authors say about 95% of the gain came from the model more often putting its answer in the box the grader requires.
How we checked
Both papers come from the same group of authors, and no one else has repeated them (arXiv 2611.15717, 2026-11-11; arXiv 2611.16228, 2026-11-12).
[This week] A JPMorgan Chase preprint reports that the Large Payments Model, a foundation model trained from scratch on payment transaction data, is serving in production at under US$0.0001 per inference.
How we checked
[This week] A Microsoft Research preprint reports that a backdoor can be planted in the weights of a model used by AI agents without retraining, by rewriting fewer than ten neurons that normally stay inactive.
How we checked
Still missing today. The gap that most affects judgment is TSMC's original financial report: the make-up of the subsidiaries' profit and loss, the subsidy amounts by country and how the subsidies are booked all rest on Taiwanese media reports so far. CoreWeave gave neither the original contract prices for its older GPUs nor a breakdown by model. Which companies the Nikkei Asia report interviewed rests on Nystedt's relay. Gigabyte's figures haven't been checked against the earnings presentation. The day's three newsletters from Interconnects, Exponential View and Latent Space haven't been read yet. Of the 148 pieces not yet read, 49 are academic papers, 50 are social-platform posts and 41 are SEC filings.
Sources & accounting
The past 24 hours. 391 new pieces came in from November 14 to November 15: 146 academic papers, 105 social-platform posts, 72 linked articles, 47 SEC filings, 14 from a second academic-paper feed (counted separately from the academic papers), 3 subscription newsletters, 2 podcast transcripts, 1 company or personal blog post and 1 filing-diff comparison. Of those 391, 90 were read in full, 150 were screened out as off-topic (11 of them before fetching), 3 were only indexed without a full read, and 148 haven't been read yet. Academic papers, 146: 23 read, 74 screened out, 49 not yet read. Social-platform posts, 105: 10 read, 42 screened out, 3 only indexed, 50 not yet read. Linked articles, 72: 56 read, 16 screened out (1 of them before fetching). SEC filings, 47: 6 screened out, 41 not yet read, mostly older annual and quarterly reports collected late. Second paper feed, 14: 10 screened out, all before fetching, and 4 not yet read. Newsletters, 3: none read yet. Podcast transcripts, 2: 1 read, 1 screened out. Blogs, 1: 1 screened out. Filing-diff comparisons, 1: not yet read. Last night's scan covered arXiv, X and SEC filings; several blog and newsletter sources couldn't be collected.
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. Amazon's 2025 annual report, cited in the Chips & semiconductors section, is an older document filed in February this year, and the CoreWeave earnings call transcript was retrieved separately; neither is among the 391 pieces above.
Source concentration. Main-line item 1, the Also happened item and the Gigabyte item in Chips & semiconductors all reached SecondSource through Dan Nystedt's posts on the same night; upstream they come from the Economic Daily News, the Liberty Times, the Central News Agency and Nikkei Asia. Main-line item 2 is one person's compilation, Mollick's. The CoreWeave item in Chips & semiconductors is the seller's own account. All three Research & models items are arXiv preprints, with results reported by their authors.
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. Representative names include Dan Nystedt, Ethan Mollick, Elon Musk and Gary Marcus on social platforms; Zvi Mowshowitz and Dylan Patel among newsletters; and Lisa Su and Satya Nadella on the earnings roster. This issue uses 27 outside sources in the body, counting only outside links cited in the body, a different list from the 529 tracked sources. Sources from Taiwan total 266, with 44 added in October; Taiwanese media, Taiwanese companies' official channels and local reporters in Taiwan count, while foreign outlets relaying Taiwanese media do not.