Daily Brief Oct 31, 2026

Broadcom backs a second rack lease at 72.5%, down from 85%

Source: Broadcom 8-K, 2026-10-27

Today's front-page cartoon: one single picture that brings together today's stories.

At a glance

1. Broadcom filed a second AI rack-lease backstop: if the AI lab renting the racks stops paying, Broadcom pays the gap between 72.5% of the unpaid rent and what the racks sell for, capped at about US$43 billion. The first backstop, from June, covered up to 85%.

2. Oracle and Google announced that Google's TPUs will be offered on Oracle's cloud from 2027; neither scale nor pricing has been announced.

Today's main line

1. [Today] Broadcom's second AI rack-lease backstop covers up to 72.5% of unpaid rent, capped at about US$43 billion, down from 85% in the first

Concept illustration: two platforms on posts, side by side, each carry a full row of black server racks, the right row longer, and at the right end of each row a thin pole holds up a black sign lettered BROADCOM. Under each platform a red net hangs from the leftmost post to an inner post, leaving the right end of the platform with no net beneath it, and a short line joins each net's lowest point to a label below the ground line, FIRST GUARANTEE on the left and SECOND GUARANTEE on the right, while a person in dark clothes with a badge on the chest stands between the platforms, both hands gripping the upper left corner of the right net. — On October 27, Broadcom filed with the US Securities and Exchange Commission a second AI rack-lease backstop that covers up to 72.5% of unpaid rent, against up to 85% in the first. The second ceiling is about US$43 billion, against about US$29 billion in the first, and Broadcom would pay the full US$43 billion only if the customer stopped paying entirely and the racks sold for nothing. Our inference: if the AI lab renting the racks stops paying, the lenders who financed the racks carry whatever Broadcom's 72.5% does not cover.

Why it matters: Our inference: if the AI lab renting the racks stops paying, the lenders who financed the racks carry whatever Broadcom's 72.5% does not cover, and their rates and terms on the next rack deal will reflect that gap.

Read the full item → How we checked →

2. [Today] Oracle and Google say Google's TPUs come to Oracle's cloud from 2027, which the two call the first large cloud other than Google's to offer them

Concept illustration of a sheet pinned to a drafting table, with a ruler and a pencil on the ledge below: on the left a solid black equipment rack is drawn with a red square component at its centre, labelled TPU by a line, and on the right a dashed outline of a tall rectangle of similar size holds a smaller red dashed square at its centre. Above the sheet, a multicoloured G mark and the lettering GOOGLE sit over the rack, the red lettering ORACLE sits over the dashed outline, and the lettering PLAN sits between them, each with a short line pointing down to the top of the sheet. — Oracle and Google announced that TPUs, the AI chips Google designs, come to Oracle's cloud from 2027, which the two call the first large cloud outside Google to offer TPUs. From 2027, companies renting Google TPUs will have a large-cloud option other than Google Cloud. Scale, pricing and first customers haven't been announced.

Why it matters: From 2027, companies renting Google TPUs will have a large-cloud option other than Google Cloud.

Read the full item → How we checked →

Also happened

1. Not verified by us yet: [Today] Cohere announced a C$4 billion commitment to AI in Canada (Cohere, 2026-10-30).

Deep dive: AI video's training costs stop rising once viewers can't tell clips from real footage; whether apps earn like software depends on who pays for each second generated (October 31, 2026)

Our inference: Training costs for AI video stop rising once viewers can no longer tell generated clips from real footage, so app companies can afford them. Whether AI video app companies make money depends on whether the compute for each clip is paid by paying customers or absorbed by the company. HeyGen, which makes AI avatar videos, charges paying customers and reports breaking even on cash flow so far this year. Sora, OpenAI's video app, was burning about US$1 million a day according to a Wall Street Journal investigation, and it shut down in April. Neither company discloses gross margin, so these two cases show a direction, not a measured margin.

What would prove this wrong: by June 30, 2027, HeyGen or Kuaishou, the short-video platform that sells the Kling video generator, reports video gross margin below 40% while still losing money; or Google or OpenAI offers video generation free in its main products while revenue from HeyGen or from Kling falls for two straight quarters. The deadline and thresholds are SecondSource's own.

The next check is Kuaishou's third-quarter results in November: whether Kling's revenue tops the second quarter's RMB 850 million.

What this means for you: Our inference: a video product that absorbs its users' compute costs pays more with every clip they make. For investors in and operators of video apps, the first question is whether the product bills customers by usage.

The full deep dive is published in Chinese and Japanese only; there is no English edition.

Nothing new this issue in: Chips & semiconductors, Named commentary.

Research & models

1. [This week] (preprint, arXiv 2610.38799) When several coding agents work together, a compromised testing or deployment component that hides malicious instructions in its status reports got Codex, OpenAI's coding agent, to act on them 61.3% of the time in the authors' tests.

Concept illustration: a large sheet hangs from a thin line by two binder clips, with a black header band lettered "ARXIV"; on the left is a page headed "COMPONENT" with grey bars, and on the right a black sign lettered "CODEX" sits above a laptop whose screen shows a grey speech bubble and a black beetle. A red sheet lettered "REPORT" stretches from the lower right of the left page, across a black rope strung between two square posts, to the edge of the laptop screen; its left end is lifted to reveal another black beetle underneath, and three grey bars run along the bottom of the large sheet. — A preprint's authors found that a compromised testing or deployment component hiding instructions in status reports got Codex to act on those instructions 61.3% of the time, against 58.7% for Claude Code, in self-reported tests no other team has repeated yet. In the same tests, coding agents that treat other components' reports as trusted debugging guidance carry out actions nobody authorized, even though the compromised component has no permission to touch the victim agent's files.

Read the full item → How we checked →

2. [This week] (preprint, arXiv 2610.38521) Researchers put malicious instructions into CLAUDE.md, the project memory file Claude Code reads when it starts work; against runs without them, the share of secure code the agents wrote moved between -8 and +6 percentage points, with no consistent direction.

Concept illustration: two black-framed panels hang side by side on a pale wall above a gray-brown floor, with a black sign lettered "PAPER AUTHORS" in white at the upper right. Both panels show the same scene, with a document headed "CLAUDE.MD" and filled with gray lines on the left, a black laptop lettered "CLAUDE CODE" in front of it with a speech bubble rising from its top, and on the right a tray holding a stack of paper with a black bug on the top sheet; the document in the right panel also has a red band lettered "MALICIOUS INSTRUCTION". — The preprint's authors put malicious instructions into CLAUDE.md, the project memory file Claude Code reads on starting work, and the agents' share of secure code moved between -8 and +6 percentage points against runs without those instructions, with no consistent direction. The clearer difference the paper measured is whether the agent tells the developer that CLAUDE.md influenced the agent, a result from the authors' self-reported small sample.

Read the full item → How we checked →

Product moves

1. [Today] OpenAI committed to honoring the preferences websites state in robots.txt on whether their content may be used to train AI; the labels come from a draft by the AI Preferences working group of the IETF, the body that writes internet protocol standards, and OpenAI expects to implement it within a few months.

Concept illustration: left of centre, a monitor displays the large heading ROBOTS.TXT above grey lines and a red slide switch with its white knob at the right end, beside a label reading AI TRAINING, while a hand reaching in from the lower right presses the knob with its index finger; in front of the monitor's lower left stands a sheet of paper lettered IETF, with grey lines and a dashed outline of a switch. A dotted line runs from the monitor's right side up and to the right toward a distant laptop whose screen shows a speech bubble with grey lines and a dashed oval, and behind the laptop to the right stands a white plaque lettered OPENAI with a black interlaced emblem. — OpenAI committed to honoring the AI training preferences websites state in robots.txt; so far that is only a commitment, with implementation expected within a few months. Websites can say whether content may train AI with labels from a draft by the AI Preferences working group of the IETF, the body that writes internet protocol standards.

Read the full item → How we checked →

2. [Today] Google Cloud spend caps are now generally available: spending on AI services is counted from real-time cost estimates, and new requests are blocked within minutes of reaching the cap; requests on capacity reserved in advance are not blocked.

Concept illustration of two laptops, one above the other, with the same screen layout: beside each is a sign lettered GOOGLE CLOUD with a four-colour cloud icon, and each screen has a grey bar at the top with a red vertical line labelled SPEND CAP, plus a bottom row where a pale blue RESERVED label leads through rightward arrows and three white boxes toward a white panel holding a speech-bubble icon on the right. On the upper laptop the grey bar is short, with a dashed line running on to the red line, and a middle row of three grey boxes leads by rightward arrows to the panel; on the lower laptop the grey bar reaches the red line, and the middle row of three grey boxes ends at a thick red vertical bar, with no arrow reaching the panel. — Google Cloud made spend caps generally available, blocking new AI service requests within minutes of reaching the cap. The cap counts spending from Google Cloud's real-time cost estimates, and requests on capacity reserved in advance are not blocked.

Read the full item → How we checked →

From the archive

1. [Looking back] (SecondSource deep dive, October 23) If AI data centers are overbuilt, long-term data-center leases and the debt of companies that rent out compute break before the big cloud companies' balance sheets do.

Concept illustration: on the left wall, two signs hang one above the other, the upper lettered MICROSOFT beside a four-colour grid of squares and the lower lettered BORROWINGS, with a beige open-top holder containing one black folder beneath them. On the right, a huge stepped pile of brick-red bundles tied with pale bands rises toward the upper right and runs off the frame, with a large sign lettered LEASES set on its front, a pale pink rectangle behind it and a tiled floor below. — SecondSource's October 23 deep dive found that if AI data centers are overbuilt, long-term leases and compute lessors' debt break first. Part of the AI capital cycle's risk sits in those places, which the borrowing line alone misses: Microsoft's annual report puts borrowing principal at US$46.1 billion and separately lists US$329.1 billion of leases signed but not yet started, a multi-year payment total that is not the same kind of number.

Read the full item → How we checked →

Academic papers, 207: 47 read, 47 screened out, 113 not yet read. Social-platform posts, 126: 49 read, 76 screened out, 1 to be fetched again. Second paper feed, 55: none read yet. Blogs, 31: 14 read, 14 screened out, 3 logged only. Newsletters, 9: 3 read, 6 screened out. Company filings, 4: 2 read, 2 screened out; the four are two from Digital Realty and one each from Amazon and Intel.

The Broadcom 8-K behind main-line item 1 is not among those four filings. It was found through the October 30 weekly roundup from SemiEngineering, a semiconductor-industry publication, and then read in the original on the SEC's site; that roundup is counted among the 31 blog posts. Main-line item 2 was found through the same roundup, and then both companies' press releases were read.

Last night's collection pulled 486 original X posts from 374 accounts. These 486 are a separate count and can't be divided by the 126. @teortaxesTex alone accounted for 79 of the 486, followed by @elonmusk and @TheStalwart. Newsletters included Stratechery, AI Supremacy, Dwarkesh, Epoch Gradient, Interconnects, Latent Space, Newcomer and The Zvi; none was used in this issue.

One-time catch-up. The show-transcript catch-up totals 770 pieces, 3 more than yesterday, now running through October 30; the linked-article catch-up has 61 pieces, dated October 30. None of it was used in this issue.

Still missing today. The gap that most affects judgment is the second Broadcom deal's rack volume, lease term and amount for each debt tier, which neither the 8-K nor 24/7 Wall St.'s summary gives. The scale, pricing and first customers for TPUs on Oracle's cloud haven't been announced in the press release. In last night's routine collection, one show episode's official page couldn't be found, and the Federal Reserve's rate table failed to parse and wasn't updated, so this issue has no macro section.

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 and can't be added to the piece counts above; the 374 accounts actually pulled last night are also not the same population as the 302 social-platform sources on the roster. Representative names include Daniel Kokotajlo, Sergey Levine, Arvind Narayanan and Chris Lattner on social platforms; Dean Ball, Zvi Mowshowitz, Miles Brundage and Jamin Ball among newsletters; Terence Tao, Lilian Weng and Nicholas Carlini among blogs; Tri Dao and Song Han among paper authors; and Sam Altman, Dario Amodei and Demis Hassabis among show interviewees. This issue uses 11 outside sources in the body. Only outside links actually cited in the body are counted, and links to SecondSource's own site are left out; these 11 are a separate count from the 529 tracked sources. Taiwanese-side sources total 140; Taiwanese media, Taiwanese companies' official channels and local reporters in Taiwan count, while foreign outlets relaying Taiwanese media do not.

Correction: the October 29 issue's Sources & accounting said academic papers were "52 read, 63 screened out and 4 skipped before fetching". Those 4 papers were already counted among the 63 screened out, not a separate category, so the items add up to 115, not 119.

SecondSource isn't a news digest: AI reads primary sources across fields, every judgment comes with its original source, and mistakes are corrected in public. The focus is what has changed and which claims still lack evidence, not what happened today.

Written from the same research and judgments as the Traditional Chinese edition. Sources are linked; we distinguish original documents from reporting and mark what we could not verify.