Afternoon Brief · November 3rd 2026

Afternoon Brief, November 3rd 2026

AMD forecast about $17 billion for the quarter and its shares jumped. Binance says the hackers moved through its internal AI tools. ChatGPT has 1.2 billion weekly users. And Apple told the that may defy their instructions.

28 min read·Editorial by Elias Marchetti

EMElias Marchetti
Capital & Consequence

AMD forecasts about $17 billion for the fourth quarter as data center sales nearly double

AMD reported third-quarter revenue of $13.38 billion after the close on Tuesday, up 45% and above the $13.09 billion analysts expected, with adjusted earnings of $2.09 a share against $1.95. Data center revenue rose 97% to about $8.5 billion, or 64% of the total, while gaming fell 40%. The fourth-quarter forecast of about $17 billion, plus or minus 3%, compares with a $16.2 billion and would be 27% growth in a single quarter. The shares rose about 7% , following three straight down days in a sell-off over AI spending.

Lisa Su tied the step up to scaling production of Helios, the AMD sells against Nvidia's, and "Venice" server chips. Meta, Microsoft, Anthropic, OpenAI and Oracle are beginning to use Helios, and Su plans to add a large amount of supply in 2027. Two of those five customers also take AMD's money: it agreed to put $1 billion into Meta's and has committed up to $5 billion to Anthropic. The forecast of 55% is a point below the quarter just reported, and slightly under the 55.3% estimate. SourcesA B

Updated

Binance says the North Korean hackers behind its $6.3 billion loss moved through its internal AI tools

Binance's account of the October 29th theft puts a company AI tool in the attack path. In June an engineer searching for a popular AI coding tool clicked an ad for a tampered copy of ; Binance caught it, but the attackers had learned who managed company wallets. In September, fake recruiters on LinkedIn took a second engineer through two rounds of interviews and handed over a coding exercise carrying malware. From that laptop the attackers used Binance's internal AI tool to take over the accounts of a wallet DevOps engineer and a wallet software engineer. Co-CEO Richard Teng told CNBC they "moved violently through our internal AI tools" to reach the system that authorized wallet transactions.

Binance has rebuilt the tool and has not explained how it was exploited. It says it has replenished 98% of the affected assets. Sources differ on what came back: Fortune reports about $2.5 billion recovered after Tether and Circle froze , and CNBC cites reports of about $3 billion frozen. An internal assistant that can act across staff accounts has an administrator's reach, and on Binance's account the attackers used it as one. SourcesB B

ChatGPT reaches 1.2 billion weekly users, and OpenAI puts its search volume at a fifth of Google's

Adam Fry, who leads ChatGPT's consumer product, told Alex Kantrowitz that the service has 1.2 billion weekly active users, up from 1 billion about 60 days ago. "ChatGPT's volume of search-like queries is now roughly one-fifth the volume of queries of Google," he said. OpenAI counts every question that goes to the web as a search, and one conversation can contain many, so the ratio is the company's own and nobody outside has checked it. Fry also gave 250 million weekly users across dictation and voice, and about 1.5 billion images made a week. SourcesB

Apple adds AI agents to its risk factors and tells the SEC they may defy instructions

Apple's annual report, filed Tuesday, says AI agents "may behave in unintended or unpredictable ways, including by taking actions that were not directed, instructed or authorized, exceeding permissions, defying directions or resisting instructions." It adds that some consequences "may be difficult or impossible to reverse," and that its "ability to oversee and control third-party artificial intelligence agents and other technologies may be limited." The filing covers agents "developed or deployed by the Company or by third parties." Apple sells no agent of its own yet, though it changed Mac full disk access last month to limit what installed agents can reach, and it is opening Siri AI to outside developers this week. SourcesA B

Super Micro's revenue quadruples to $19.9 billion, and Arista beats on AI networking

Super Micro reported fiscal first-quarter sales of $19.9 billion, against $5.0 billion a year earlier and a $16.39 billion estimate, and net income of $1.59 billion. It forecast $19.5 billion to $21.5 billion for the December quarter and $72 billion to $82 billion for the fiscal year, and the shares rose more than 7% after hours. The midpoint of that quarterly forecast is about 3% above the quarter just ended. Gross margin was 12.7%, up from 9.3%.

Arista reported $3.66 billion, up 58.5%, against of about $3.3 billion, and forecast $3.9 billion to $4.0 billion for the fourth quarter. Its gross margin was 63.6%. The two sit in the same AI racks, and the server assembler keeps about a fifth of what the switch maker keeps from each dollar of sales. SourcesA A B

Revenue growth from a year earlier, latest quarter

AI hardware results reported Tuesday

Company releases, November 3rd 2026. Super Micro from Reuters; its release gives $19.9B against $5.0B

Fal doubles its valuation to $16 billion in seven months

Fal, which hosts and runs AI video, image and audio models, raised a $1.3 billion Series F led by Bessemer at a $16 billion , up from $8 billion in April. Nvidia, Sound Ventures and T. Rowe Price joined, and with a Sequoia-led announced alongside it the two rounds total $1.65 billion. Bloomberg reports annualized revenue passed $800 million earlier this year, which puts the price at about 20 times that figure. Gross margins remain thin across companies because compute is expensive to buy or lease, according to people familiar with their finances, and demand has run strongest among customers using open models. SourcesB B

CrowdStrike pays about $495 million for two AI security startups and ships a kill switch for agents

CrowdStrike agreed on Tuesday to buy Fiddler AI, which monitors what agents do, and Dreadnode, which tests AI systems for security holes before deployment. The Wall Street Journal puts the prices at about $400 million and about $95 million, mostly in cash; CrowdStrike's releases give no figures. Fiddler becomes a product called Falcon Observe, and that deal is expected to close in CrowdStrike's fiscal fourth quarter.

The same day it launched an AI inside Falcon Guardian. A person decides when to pull it. The tool then blocks the agent from relaunching, cuts its network access and uses connected AI gateways to cut it off from the models that drive its decisions. It leaves the agent's code in place. SourcesA A B B

Google will turn AI off by default for K-12 students, and schools that refuse it lose Search

Google said Tuesday that Gemini, Gemini Notebook, Gemini in Classroom and Google Search will stay off for K-12 student accounts until administrators turn them on, a change due next year. Search is on that list because Google will not offer it without AI: students at schools that leave AI off would see a message that Search is unavailable. "Search and AI are deeply integrated," said Chris Phillips, Google's vice president for education. The same post launched Guided Learning, a tutor-style mode, in Classroom worldwide, with a Search pilot in some US schools later this month. The opt-out therefore costs a school its search engine, and few districts will pay that. SourcesA B

Meta goes back to the bond market for up to $25 billion, with its first sale in euros

Meta filed a preliminary on Tuesday for a five-part dollar offering, with amounts left blank, and CNBC reports the total at $20 billion to $25 billion across dollars and euros, about two-thirds and one-third. Bloomberg says it is Meta's debut in euros. S&P moved its AA- rating to a on Friday, citing uncertainty about capital spending and negative , which was minus $2.12 billion last quarter. The sale comes a week after Meta said it would raise $95 billion in equity. Final pricing was expected Tuesday afternoon and had not been reported at publication. The shares rose 2.7% after falling in 12 of the prior 14 sessions. SourcesA B B

Pimco's chief says AI financing will become "too much," as Nikkei counts $3.2 trillion off the balance sheet

"At some point in time, it's going to be too much," Pimco chief executive Emmanuel Roman said at a Hong Kong Monetary Authority summit, setting about $550 billion of AI debt issued this year against a global corporate bond market of roughly $20 trillion. On the same panel, Blackstone president Jon Gray said lenders are becoming concentrated in the same few names and "there's going to be more equity needed." Goldman Sachs estimated this week that data center demand needs about $8.5 trillion of spending over five years; Roman cited General Atlantic chief Bill Ford's figure of about $10 trillion.

Separately, Nikkei reports that obligations from AI at US tech giants including Google and Meta reached about $3.2 trillion, up 90% in six months and equal to 80% of assets. SourcesB B

Updated

Apple Intelligence goes live for developers in China, running on Alibaba's Qwen

The first iOS 27.3 beta arrived Tuesday, and developers in mainland China can now switch on Live Translation, Writing Tools, Visual Intelligence, Genmoji, Clean Up, mail and notification summaries, and a typed Siri that relays requests to Qwen. Bloomberg reports an Alibaba-powered content filter sits in the system. Siri AI, the Siri app and the newer iOS 27 features are all excluded. A public release is expected in December. SourcesB B

Siri AI opens to every developer, and Apple says heavy use will cost extra

Apple released a Siri AI developer preview on Tuesday that lets any app bring its content and actions into the assistant through , chain them into "skills," and surface anything in the Spotlight semantic index. Siri AI reaches customers "in a few months," and App Store submissions open in the spring. It is unavailable in the EU on iOS, iPadOS and watchOS for now. A footnote says server-side Apple Intelligence features have daily limits and that "expanded access to such features will be available for a fee in the future." Apple has not said what it will charge.

The watchOS 27.3 beta adds Live Rewind, a transcript of the last 15 seconds of conversation, and Siri Recap, background summaries deleted after seven days unless saved. Both need a Series 12 or Ultra 4 watch. SourcesA B

DeepMind staff in Britain ask Google to drop a lifetime ban on criticizing the company

UK-based Google DeepMind employees are urging leaders to remove that bar public warnings about AI risks even after they leave Google, the Washington Post reports, citing sources. The ask matches the 2024 "Right to Warn" letter from current and former OpenAI and DeepMind staff, which urged labs not to bind people from airing risk concerns. This account rests on a summary of the Post's report, with no response from Google in it. SourcesB B

US officials hold up tools SK Hynix wants for China, complicating its New York plants

US officials have held up approval for American chipmaking equipment SK Hynix needs to expand in China, Reuters reports, citing sources, and that is complicating the company's plans for new plants in New York. Reuters reported last month that SK Hynix wanted the tools cleared as a condition of funding a factory in the US, with upstate New York under consideration. In a memory shortage, Washington is weighing a US against more capacity in China, and for now it is choosing to wait. SourcesB C

Anthropic releases FlashProve, a prototype that proves which model wrote a response

Anthropic's Frontier Red Team published FlashProve, dated Monday, under an license. A provider commits to a fingerprint of its , then produces a that those weights generated a given reply, and anyone can check it without seeing the weights. The use cases are catching a router that quietly swaps in a different model and checking that only approved models ran.

On Llama 4 Maverick, a 400-billion-parameter model at 4-bit, one H100 proves about 110 a second with proofs of about 18.78 MB, and a verifies one in about 17 seconds. That costs roughly 70 to 95 times the time of plain inference. It covers greedy decoding, integer weights and 1,024-token windows, so sampled production traffic at long context is out of scope. At that price it is an auditing tool for a slice of requests. SourcesA A

Anthropic tells IPO investors its 15-person chip team is designing with Claude

Anthropic has told investors in recent meetings that a chip design team of 15 people is using Claude for engineering and design work on its custom chips, The Information reports, citing sources. Anthropic has confirmed it is building silicon in-house but has given no timeline. A team that size would normally be far short of what a custom takes, which is why the pitch matters to a company carrying $895 billion of compute commitments. Investors have only Anthropic's word for the productivity. SourcesB C

Revolut says self-hosted GLM cut its AI coding bill by about 90%

Revolut made Z.ai's open GLM models its default for about 2,000 engineers and cut coding inference costs by roughly 90% in under a year, infrastructure lead David Abitbol told the AI Engineer conference in New York. He put self-hosted GLM-5.1 at $0.30 per million tokens, about a tenth of Codex or Claude Code. Revolut now serves GLM-5.3 on 56 Nvidia GPUs and routes vision work to Gemini or GPT. On the Chinese origin, Abitbol said the main concern is reaching the models through third-party APIs, and that on-premises or regional deployments carry little risk. The savings let Revolut stop rationing tokens, and he said the biggest gains came from debugging and ticket work. All figures are Revolut's. SourcesB

Stanford retested 50 well-known findings about LLMs, and 44% no longer hold

Owen Queen and James Zou reran 50 widely cited findings across 20 models released from July 2024 to September 2026. They class 22 as expired, 17 as weakened, 8 as persisting and 3 as stronger, and most expired or weakened within 15 months of publication. The expired ones are mostly capability failures: counting letters, finding facts in the middle of long contexts, abandoning correct answers when challenged. In 2023 ChatGPT fabricated 55% of its citations; recent models rarely do. Three grew stronger: judges favoring their own output, praise that follows a user's stated preference, and nonuniform random numbers. A claim about "LLMs" with no model and date attached tells a reader little. SourcesA

More than 400 apps and domains hid deepfake tools behind pet editors and language tutors

Wired reported Monday that a network of more than 400 connected apps and domains passed Apple and Google review posing as everyday services, then loaded a hidden interface for generating explicit AI images from uploaded photos under certain conditions. Researchers found it by following video ads on Facebook and Instagram, some of which contained AI child sexual abuse material. Apple removed 19 apps after Wired asked. Google had already removed most of the ones found. Meta says it acted on 11,600 ad accounts and 15,000 pages and blocked 209 apps from advertising. Twenty-seven apps live on Google Play in September had more than 350,000 combined downloads. The operation appears to run through Chinese developers and a Hong Kong entity, KeLing Technology Limited, which did not respond. SourcesB

Updated

Palantir recovers most of its premarket drop as analysts dig into contract value

Palantir shares were down 1.65% on Tuesday, CNBC reports, after trading down about 8% before the open. Analysts focused on US commercial of about $1.5 billion, up 11% from a year earlier and down from a record $2.1 billion in the second quarter. Enterprises are starting on shorter trials and are more willing to cancel, TechCrunch reports. Piper Sandler says demand still exceeds supply, and Goldman says the test is "converting shorter duration contracts into longer duration commitments." SourcesB B

OpenAI hires Uber president Jill Hazelbaker to run corporate affairs

Jill Hazelbaker will join OpenAI as chief corporate affairs officer, leading global communications and corporate positioning and, according to Axios, reporting to Greg Brockman. She is Uber's president and chief corporate affairs officer, and ran communications at Snap and Google before that. It is OpenAI's second senior outward-facing hire in two days, after former Army Secretary Dan Driscoll on Monday. SourcesA B C

Updated

AI political ads reach $187 million as polls open

Campaigns and political groups have now spent more than $187 million on nearly 300 AI-generated or AI-enhanced ads, the Wesleyan Media Project says, up from the $137 million it reported last month, and it calls that an undercount. More than 90% of the money came from Republican candidates, committees or aligned groups. Most of the ads carry no disclosure. No federal law covers them, about 30 state laws do, and CNBC found undisclosed examples including a fake video of New York Governor Kathy Hochul posted by her opponent. The New York Times reports Anthropic has its first-ever "war room" running. SourcesB B

Apple rates AI companions with romantic content 18+

Apple told developers Monday that apps offering dating or romantic relationship features must carry an 18+ rating, and that covers AI companions with sexual or romantic content, including apps such as Character.AI and Replika where that is one feature among many. New content descriptors become mandatory on November 16th. SourcesA B

Western startups are selling Chinese open models with the censorship edited out

Hirundo and Boston's Lazarus AI edit the weights of models from DeepSeek, Moonshot, Z.ai, Alibaba and Tencent to strip state-aligned content and reduce security risks, Rest of World reports. Lazarus's ReAligned Qwen models run at about a dozen US companies. Hirundo says its edits cut biased answers from 89.8% to 2.8% of 500 prompts on its own , which nobody has checked independently. The market exists because Airbnb, Uber and Harvey have built on Chinese models and top Chinese systems trail the best US closed ones by about 3% on benchmarks. Washington barred civilian agencies from DeepSeek products in October. SourcesB C

Updated

An open 27-billion-parameter model on a laptop repeats part of the Astra weight-reading test

Simon Willison gave the paper's raw-weight game test to Qwen 3.8 27B, running locally in about 17GB, and it named Pac-Man, Pong and Breakout from the numbers alone and failed on Freeway, taking 46 to 65 minutes each. Kimi K3.1 Pro got Breakout in under ten minutes. The paper presented reading weights as something GPT-6 Astra could do; an open model on a laptop now does most of it too, given an hour. Willison did not test the editing result, and says he is "not ready to declare that the age of source code is over." SourcesC

Netflix tests an AI Discover tab on US televisions

Netflix began rolling out an AI Discover tab on Tuesday for US users on supported TVs in English. People pick a theme such as "scare me" or "quick watches," then refine it, and a "Tune Your Recommendations" button saves or excludes titles. Netflix says it runs on a new "intelligence layer" and calls the tab a "test kitchen" for discovery features. It names no model or partner and gave no usage figures. SourcesB

Postman's desktop app installed plugins into Claude Code and Codex without asking, a user reports

A GitHub issue opened October 25th, which drew wide on Hacker News on Tuesday, says Postman 12.33.4 for macOS added plugins to Claude Code and Codex on launch, wrote a remote connection and a , and set a startup hook telling the agent to load Postman's API skill, approving that hook itself. The reporter does not claim any data left the machine. No Postman staff had answered in the thread as of Tuesday. Coding agents load whatever sits in their config, which makes those files a new place for a vendor to reach, with or without consent. SourcesC

Ballistic Ventures files for $490 million as AI reshapes cybersecurity investing

Ted Schlein's firm is raising a $360 million third flagship fund and a $130 million Fund B, SEC filings show, with no sales recorded yet. AI-enabled companies made up about 72% of US cybersecurity venture deals through May, up from 36% in 2019, according to JPMorgan citing PitchBook. One of its earlier companies, Pangea, was bought by CrowdStrike, which agreed to spend about $495 million on two more AI security startups on Tuesday. SourcesB

Tutor Intelligence raises $75 million from a16z to train warehouse robots

Tutor Intelligence, which builds AI warehouse robots and trains robot , raised a $75 million Series B led by a16z, Forbes reported Monday, taking its total to $117 million. Its last round was a $34 million Series A led by Union Square Ventures in December 2025. SourcesB C

Sources

111 citations · 29 primary · 68 · 14 weaker

Morning Brief, November 3rd 2026