Daily Brief Nov 1, 2026

OpenAI reportedly cuts raise to US$20B, per the FT

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

At a glance

1. The Financial Times reports that OpenAI cut this round's target from US$30 billion to US$20 billion; CNBC had already reported the US$1.25 trillion valuation, before new money, on October 22, ahead of this week's selloff.

2. Intel's data-center segment grew 78% year on year in the third quarter, yet fourth-quarter revenue guidance of US$15 billion to US$16 billion sits below the third quarter's US$16.5 billion; Intel's CFO expects the segment that includes PCs to fall more than 20% from the third quarter.

This week's column: Claims & Receipts

Every Sunday: one AI claim making the rounds, checked against the original documents.

No new claim was added for checking this week; this issue only returns to three older calls that came due.

OpenAI's system card, Anthropic's enterprise share, the US-China AI gap: can these three older calls be settled?

The short version: SecondSource rechecked three calls that came due, and none of the three can be judged, because the named sources published no new figure, or published a document that doesn't answer the question.

How to use it. OpenAI's October 26 system card gives no cause, so it can't explain why the model's abilities changed. This check covers three older calls, not a claim newly making the rounds this week.

Will OpenAI's next system card explain why models can reason without writing a chain of thought? Can't judge: a system card is the safety and capability document published with a model. The October 26 GPT-6.1 Terra system card says only that the model is less able than its predecessor to rewrite, on instruction, its chain of thought, the reasoning draft a model writes before answering, and gives no cause. Reasoning without writing a chain of thought is a separate matter, and the card never mentions it. SecondSource's October 22 call was narrower: that this card would give a cause for the rewriting ability alone. The card gave no cause at all. OpenAI's September GPT-6 Astra system card said the change in rewriting ability was not specifically caused by architecture changes, and this card doesn't repeat that sentence. Answer check →

Will Anthropic's enterprise market share fall to 35% or below? Can't judge: this question waits on the next enterprise report from Menlo Ventures, a venture firm that is also an Anthropic investor. The question originally ran to September 30 and got 30 more days, to October 30, and Menlo Ventures' article index still lists no new edition. The latest remains the December 2025 report, which estimated Anthropic's share of the enterprise large-model market at 40%, the baseline for this question. Answer check →

Will the US-China frontier AI capability gap land between 4 and 14 months the next time Epoch AI publishes it? Epoch AI, which calculates this gap in months, had published no new figure by October 31, so this can't be judged. The designated page still shows the January 2026 version. Converted from Epoch AI's capability index, that version has Chinese frontier models trailing US ones by 7 months on average, with a smallest gap of 4 months and a largest of 14. Answer check →

All of SecondSource's calls and their verdict dates → the full ledger

Today's main line

1. [Today] OpenAI's target reportedly falls from US$30 billion to US$20 billion, and the US$1.25 trillion valuation was in print eight days before that report

Concept illustration of a tablet screen with a dark header band lettered TECHCRUNCH and a green block mark at its right; below it, a black interlaced knot emblem is linked by a short line to a label reading OPENAI, above a horizontal rod linked by a short line to a label reading FUNDING TARGET. Three hooks hang from the rod: the left two each hold a beige sack with a round coin design, the rightmost is empty, and below and to the right of it stands a person in a black turtleneck, grey trousers and a lanyard badge, holding up in both hands a red sack with the same coin design, while four grey bars fill the lower left of the screen beneath the sacks. — OpenAI reportedly cut this round's fundraising target from US$30 billion to US$20 billion, according to the Financial Times. CNBC had already written of a US$1.25 trillion valuation before new money eight days before the cut was reported. Our inference: US$1.25 trillion is so far only an asking price, and late-stage funds weighing it as a benchmark have to wait for the first close, whose price and terms haven't been made public.

Why it matters: Our inference: this week's change is in OpenAI's fundraising target. OpenAI's US$1.25 trillion valuation is so far only an asking price, not a transaction price. Late-stage funds deciding whether to use it as the benchmark for the next round have to wait for the first close's price.

Read the full item → How we checked →

Chips & semiconductors

1. [This week] (October 29 earnings) Intel's data-center segment grew 78% year on year in the third quarter, and fourth-quarter total revenue guidance of US$15 billion to US$16 billion is still below the third quarter's US$16.5 billion.

Concept illustration: an open book lies on a dark ledge, with a sign above its spine lettered "INTEL" beside an oval mark lettered "intel"; on the left page, a large black laptop marked "PC" and a shorter black box marked "DATA CENTER" stand on a horizontal line, above a big solid-outlined brick-red block. On the right page the positions swap, with a taller pale box marked "DATA CENTER" on the left and a smaller pale laptop marked "PC" on the right, both in dashed outline, above a slightly smaller brick-red block that also has a dashed outline. — Intel's earnings report shows third-quarter data-center segment growth of 78% from a year earlier, while fourth-quarter total revenue guidance stays below the third quarter's US$16.5 billion. Our inference: AI lifting data centers doesn't guarantee total revenue growth at a company that also sells PC chips, and one Intel quarter can't be extrapolated.

Zinsner attributed PC weakness to memory shortages and limited IT budgets. The newsletter AINews wrote the cause as AI demand pushing up memory costs (AINews, 2026-10-30). In the remarks cited here, Zinsner doesn't directly make that link.

Next to watch: Intel's fourth-quarter report next January, whether the data-center segment reaches about US$8.4 billion (US$7.3 billion plus 15%) and whether the segment that includes PCs falls more than 20%. If that segment's drop is far smaller than 20% and total revenue tops the US$16 billion high end of guidance, this inference has to change. AMD also sells into both markets, and its next-quarter figures serve only as a comparison. How we checked →

Read the full item →

Named commentary

1. [Today] Azeem Azhar, author of Exponential View, rebuts McKinsey: data centers compete on how fast they can get power, not on the price of power.

Concept illustration in two rows: on top, two white buildings lettered CHINA (left) and LONDON (right) have a short stack of gold coins and a tall stack of gold coins in front of them respectively, with a sign lettered POWER BILL between them, a sign lettered MCKINSEY at right and giant black quotation marks in the top-right corner. Below, a building lettered DATA CENTER with server racks showing sits at left under a black quotation mark, a central pylon sends wires to it and to a dark building at right, and a dark coin stack stands in front of the pylon, while a cable from the left building passes behind the stack and ends in an orange plug just short of a wall socket on the dark building, with a silhouetted figure in a suit at lower right pointing at the socket. — Azeem Azhar, author of Exponential View, disputes McKinsey's line that the electricity bill decides the economics, arguing that data centers compete on how fast power arrives rather than on the power price. The dispute over data center economics is price versus speed of power, and Azhar's claim that power is a small share of costs is a personal judgment with no cost figures attached.

Read the full item → How we checked →

Research & models

1. [This week] (preprint, arXiv 2610.39641) Two authors at Samyama AI, which builds research knowledge graphs, hand-checked 41 model rulings and found 9 fully correct.

Concept illustration of a large open book with two people in black turtlenecks, one with short curly hair and one with long hair, standing at its center and reading the sheets of paper they hold; at the lower left page are a laptop showing a speech bubble and a stack of paper with a sheet lettered VERDICTS leaning against it. On the right page, a black banner at the top carries a blue-and-orange looped mark and the lettering SAMYAMA AI, and below it, from left to right, are a short red stack of paper, a tall white stack whose top sheet has a torn corner, and two cards each showing a black beetle. — Two authors at Samyama AI, which builds research knowledge graphs, hand-checked 41 model rulings that an open problem had been solved or refuted and found 9 fully correct. Having a model rule automatically that a paper solved an open problem isn't reliable yet, and status changes on open problems shouldn't be applied automatically, according to the authors' self-built audit, which no other team has repeated.

Read the full item → How we checked →

2. [This week] (preprint, arXiv 2610.39570) Leo Itoh and Junichiro Kadomoto self-report that an agent testing the same CPU design on an FPGA finished 8,000 CoreMark iterations, build time included, 9.6 times faster than single-threaded RTL software simulation.

Concept illustration of an open book whose spread has a black band of white capital lettering across the top and two rows below, each starting on the left page with the same person in a dark turtleneck seated at a desk, hand to chin, facing a laptop whose screen shows a speech bubble and a panel. In the upper row, under a black label reading SIMULATION, a black square plate with a chip outline sits near the left end of a long double-line track that crosses both pages to a stop at the far right; in the lower row, a cable runs from the laptop to a red circuit board under a black label reading FPGA, and a double-line track runs from the board across both pages to a matching black plate beside the stop at the far right. — In a preprint, Leo Itoh and Junichiro Kadomoto self-report that an agent testing the same CPU design on an FPGA finished 8,000 iterations of the CoreMark test, build time included, 9.6 times faster than single-threaded RTL software simulation, a comparison that is not production-grade parallel simulation. Chip design teams and design-software vendors have to decide whether to rebuild the evaluation step first so agents get fast feedback, though the result comes from one human-supervised single-core design.

Read the full item → How we checked →

3. [This week] According to the paper's abstract, Amazon's sandbox service Sandman, built for training tool-using agents, created 3.1 million isolated execution environments (sandboxes) across four training campaigns in October.

Concept illustration: on a four-panel folding screen, many small overlapping windows spread upward and rightward from behind a large grey laptop at lower left whose lid is lettered SANDMAN, and a larger red window at the centre is lettered SANDBOX. AMAZON is lettered at upper right above an orange curved arrow, and four small open laptops stand in a row at lower right, each screen showing a speech bubble with lines inside. — According to the paper's abstract, Amazon's sandbox service Sandman, built for training tool-using agents, created 3.1 million isolated execution environments across four training campaigns in October. The figure counts creations, including short-lived containers, not training samples.

Read the full item → How we checked →

Nothing new this issue in: Product moves.

From the archive

1. [Looking back] (SecondSource call, July 31) SecondSource put the odds at about 85% that Claude Sonnet 5 would return to its original price after the promotion ended, and the price didn't go back.

Concept illustration: on the left, a long-haired person in a long black coat stands facing right, with a thought bubble above their head showing three round discs on a tray; at the center, a grey board lettered "SONNET 5" stands on a counter with a tray holding two round discs in front of it. On the right, a person in black wearing a lanyard badge leans over and holds with both hands a red-framed black sign lettered "OFFICIAL PRICE" on the front of the counter, a string tag lettered "PROMO PRICE" lies on the floor in front of the counter, and the beige wall behind shows a black angular shape beside the lettering "ANTHROPIC". — SecondSource estimated on July 31 that Claude Sonnet 5 would return to the original price after the promotion, but Anthropic made the promotional price regular and the expiry date in the announcement was not enforced. Input stays at US$2 per million tokens instead of going back to US$3, and at heavy usage a unit-price gap that size means redoing a gross-margin model.

Read the full item → How we checked →

Also happened

1. Not verified by us yet: [Today] The X account Robocurve self-reports that two copies of the same AI, kept isolated, exchanged a secret word through robots and cameras. The post gives neither the number of attempts nor a success rate (Robocurve, 2026-11-01).

Academic papers, 112: 26 read, 33 screened out (8 of them before fetching), 53 not yet read. Social-platform posts, 80: 20 read, 55 screened out, 1 logged only, 4 not yet read. Second paper feed, 12: none read yet. Blogs, 5: 3 read, 1 screened out, 1 logged only. Linked articles (6), newsletters (3) and show transcripts (2): all read.

Intel's filing is not among those 220; it was looked up directly at the US Securities and Exchange Commission. Last night's collection pulled 237 original X posts from 374 accounts. These 237 are a separate count and can't be divided by the 80. @teortaxesTex alone accounted for 47 of the 237, followed by @elonmusk and @TheStalwart.

One-time catch-up. The show-transcript catch-up totals 770 pieces and the linked-article catch-up 61 pieces, dated through October 30. None of it was used in this issue.

Still missing today. The gap that most affects judgment is the amount, price and terms of OpenAI's first close, none of which has been published; OpenAI didn't respond to TechCrunch's request for comment. One show episode from last night has no official page that could be found. 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 20 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 20 are a separate count from the 529 tracked sources. Taiwanese-side sources total 143, 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.

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.