Daily Brief Nov 3, 2026

Oracle files US$20B of stock and bonds, the size of its cap

At a glance

1. Oracle filed to sell US$10 billion of stock and US$10 billion of bonds, US$20 billion in all, the same size as the fundraising cap through fiscal 2030 that Oracle announced on October 28.

2. Marvell will buy optical-component maker Lumentum in an all-stock deal that values Lumentum at US$115 billion, about 23.9% above Lumentum's October 28 share price.

3. OpenAI self-reports ChatGPT's search volume at about a fifth of Google's; SEO tool maker Ahrefs, counting only queries traditionally searched on Google, estimated ChatGPT's volume at about 12% of Google's in February.

Today's main line

1. Oracle files for US$10 billion of common stock and US$10 billion of bonds at once, matching the funding cap it announced five days earlier

Concept illustration: at a long counter, a person in black, seen from behind at the left, rests both hands on two red books side by side, lettered STOCK and BONDS. Across the counter, a seated person in a black top with a lanyard badge holds out an open palm toward the books, below a black sign lettered SEC on a pink wall at upper right; to the right of the books, an open document lies on the counter, lettered ORACLE in red and ORACLE in black at the top, with a large blank rectangular frame on each of its two pages. Printed on the image: US$10B each Common stock and bonds Oracle filed to sell. — Oracle filed preliminary prospectuses with the US Securities and Exchange Commission on November 2 to sell common stock and bonds, matching in total and split the cap on outside fundraising through fiscal 2030 that Oracle committed to on October 28. Both prospectuses are preliminary, so the share price, the bond coupons and the final amounts are still blank. Whether the offering counts toward the cap is still unstated, and the prospectuses do not say if any of the offering is refinancing, which the cap excludes.

Our inference: if these two offerings count toward the US$20 billion cap Oracle announced on October 28, any further data-center buildout has to be funded from cash flow, customer prepayments and leases. That matters most to Oracle's bondholders and to the developers that lease data centers to Oracle.

Oracle's October 28 cap matches these filings exactly, in total and in split.

Oracle's February 1 financing-plan press release said that, apart from the bond sale early in the year, Oracle "does not expect to issue additional bonds" in 2026 (Oracle financing plan press release, 2026-02-01).

Outside the cap, Oracle has two routes that don't count as outside fundraising.

Read the full item →

2. Marvell agrees to buy optical-component maker Lumentum in an all-stock deal that values Lumentum at US$115 billion

Concept illustration: on the left, a low red storefront has a white sign lettered LUMENTUM beside a fan of coloured lines, and its window shows a pale box-shaped device sending a red beam from a round port on its right toward three small modules lined up to its right; on the right, a taller black storefront has a white sign lettered MARVELL beside a blue M-shaped mark, with three dark square chips with gold pins in its window. On the black building's roof stands a billboard on two posts showing a low red building section joined to a taller black one, and a person wearing a lanyard stands on the roof to its right, holding a paintbrush against the billboard's right edge. Printed on the image: US$115B Lumentum enterprise value, Marvell as acquirer. — Marvell agreed on November 2 to buy Lumentum, which makes lasers and optical transceiver modules, in an all-stock deal the two companies announced jointly. The valuation and the exchange ratio are terms from the joint press release, and the projected cost savings and addressable market are estimates by Marvell and Lumentum. The deal is expected to close in the first half of 2028 and still needs approval from both sets of shareholders and from regulators.

Read the full item →

3. OpenAI tells a reporter ChatGPT's search volume is about a fifth of Google's; SEO tool maker Ahrefs estimated about 12% in February

Concept illustration: in the upper right, a black interlocking-loop emblem sits beside large black lettering reading OPENAI, above a long counter with a beige top standing on a tiled floor. Along the counter's white front, from left to right, are a red door panel with a horizontal handle set against a faint pink diagonal patch, a laptop-like screen lettered CHATGPT, five white door panels each with a horizontal handle, and a laptop-like screen lettered GOOGLE with a blue, red, yellow and green letter G in front of the word. — OpenAI told tech reporter Alex Kantrowitz how ChatGPT's search volume compares with Google's, citing OpenAI's internal data, and said ChatGPT's share is growing fast. The claim is self-reported by OpenAI and relayed in Kantrowitz's November 2 post on X, and OpenAI hasn't said how search volume is counted. Ahrefs, which makes website traffic and keyword analysis tools, estimated in February that ChatGPT's volume was about 12% of Google's, lower than OpenAI's number, counting only topics people have traditionally searched on Google.

Read the full item →

Also happened — not verified by us yet

No new items in this section this issue.

Deep dive: Who gets new AI models first: labs nominate, and US government letters list who is cleared (November 3, 2026)

The evidence says: AI labs give a few trusted partners a new model for several weeks before public release. From late June to mid-September, the White House Office of the National Cyber Director wrote to AI labs model by model, listing the trusted partners cleared for early access. Those partners were drawn from confidential customer lists the labs gave the government. Our inference: an early-access seat has to pass two gates outsiders can't see, a lab's nomination and government clearance. Seats will therefore tend to go to customers the labs value that can also pass US national-security review.

The dates and descriptions of these letters come from the withholding list the government attached when it held back documents in a records lawsuit; the list gives each document's date, subject and reason for withholding, and SecondSource covered it on November 2. This deep dive adds the Commerce Department's June export-control letter restricting foreign persons' use of new models, the text of the president's June executive order, and the opposing account a White House official gave CNBC in July, that the government doesn't approve companies' releases.

What would prove this wrong: by March 31, 2027, court orders or documents the government releases itself show that the office's letters only acknowledged the labs' lists and never cut a name from any of them; "government clearance" would then become "government receipt".

What this means for you: If this inference holds, security firms that rely on new models to find vulnerabilities but aren't on a trusted-partner list will get the new models only after public release.

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

Chips & semiconductors

No chips & semiconductors item this issue. The Marvell–Lumentum deal is main-line item 2.

Named commentary

No named commentary this issue. Only one show transcript came in last night; see Sources & accounting at the end.

Research & models

1. AWS's November 2 weekly roundup says OpenAI's GPT-6.1 Terra is generally available on Amazon Bedrock, priced at least 50% below GPT-6.1 Sol.

Concept illustration: a large framed window is set in a pale grey wall. Inside the window on the left, a black screen lettered SOL stands above a plain beige counter with a big pile of coins on a tray in front of it. In the middle, a red screen lettered TERRA has a much smaller pile of coins on a tray in front of it, and a speech bubble filled with grey lines sits above each screen. To their right are a standing sign with an interlaced symbol and the word OPENAI, and the edge of a grey screen. A black band lettered BEDROCK runs beneath the TERRA screen, the sign and the grey screen. On the wall right of the window are the large letters AWS, smaller lowercase letters aws with an orange curved line under them, and a grey door. Printed on the image: at least 50% Terra priced below GPT-6.1 Sol, per AWS roundup. — AWS's November 2 weekly roundup says OpenAI's GPT-6.1 Terra, the model meant for everyday work, is generally available on Amazon Bedrock, AWS's platform that lets companies use models from several providers through one interface. Terra matches the flagship GPT-6 Astra on a test of multi-step tasks, a result that rests only on OpenAI's own evaluation, which no one else has repeated.

Read the full item →

From the archive

1. [Looking back] (SecondSource call, July 14) On whether synthetic data has a ceiling, SecondSource split fields in two by whether answers can be checked automatically.

Concept illustration in two side-by-side panels: on the left, a table lettered MATH holds a laptop with a speech bubble on its screen, and arrows carry two lined sheets in an arc rightward to a red laptop on a tall stand. From there one arrow points down to a sheet showing a bug above a wastebasket on the table, while more arrows carry two sheets on to a larger dark-screened laptop at the right end; on the right, a table lettered WRITING holds a similar speech-bubble laptop, and arrows carry three lined sheets directly rightward to a smaller dark-screened laptop. Printed on the image: Only checkable fields broke ceiling, our call. — On July 14, SecondSource put math and code in one group and writing and general capability in another, to judge whether a new model trained only on an older model's synthetic data can outgrow the older model. SecondSource based that judgment on OpenAI's process-supervision paper and DeepSeek's paper on reinforcement learning for reasoning.

Read the full item →

Academic papers, 285: 93 read, 60 screened out (1 of them before fetching), 132 not yet read. Social-platform posts, 115: 47 read, 54 screened out, 4 logged only, 10 not yet read. Blogs, 34: 24 read, 10 screened out. Company filings, 13: 12 read, 1 not yet read. Second paper feed, 11: none read yet. Newsletters, 8: 5 read, 3 screened out. Linked articles (3) and the show transcript (1): all read.

There is only one show transcript because last night's transcript fetch failed, and a fallback attempt through YouTube returned nothing either. Oracle's February press release in main-line item 1 and the Ahrefs estimate in main-line item 3 were looked up separately to check the main line, and aren't among the 470. Last night's collection pulled 456 original X posts from 374 accounts. These 456 are a separate count from the 115 above and can't be compared with them directly.

One-time catch-up. The database also holds several one-time catch-up batches dated between July 1 and October 30 this year, counted separately from the past 24 hours above. None of it was used in this issue.

Still missing today. The gap that most affects judgment is the pricing version of Oracle's prospectuses: the preliminary versions leave the share price, the coupon on each tranche and the final amounts blank, and the prospectuses don't say whether this offering counts toward the cap. Kantrowitz's post doesn't say how ChatGPT search volume is counted, and the AWS roundup doesn't give GPT-6.1 Terra's absolute prices. 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; Terence Tao and Lilian Weng among blogs; and Sam Altman, Dario Amodei and Demis Hassabis among show interviewees. This issue uses 12 outside sources in the body. Only outside links actually cited in the body are counted; these 12 are a separate count from the 529 tracked sources. Taiwanese-side sources total 150, with 43 added in October and 9 so far in November; Taiwanese media, Taiwanese companies' official channels and local reporters in Taiwan count, while foreign outlets relaying Taiwanese media do not.

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.