October 29th 2026
Curated AI news and stories.
Samsung earns 107 trillion won in a quarter, nearly all of it from chips
Samsung Electronics reported third-quarter revenue of 195.25 trillion won, up 126.9%, and of 107.44 trillion won, up 783.1%, on Thursday. The chip division made 106.5 trillion won of that on 147.1 trillion won in sales. The device business, phones and appliances, lost 800 billion won, because the memory Samsung sells to itself got expensive too.
The memory numbers are the ones to keep. HBM4 revenue more than tripled from the second quarter and is now about 60% of Samsung's sales. The company says next year's HBM supply talks closed at its target prices with the full volume sold, that it has long-term agreements with 10 major customers, and that it expects supply to stay tight through 2028. Capital spending for 2026 is 84.4 trillion won, 76.8 trillion of it for chips.
The shares closed down 4.13% at 232,000 won. Samsung declared a 30 trillion won quarterly dividend, more than 12 times the per-share payout of each earlier quarter this year, but left 55 to 75 trillion won of its 2026 return plan undecided until a January board meeting. Investors wanted the rest of the plan. One day after SK hynix said it is sold out through 2029, the two largest memory makers are describing the same shortage. SourcesBB
DeepSeek and Moonshot reportedly trained their newest models on GPT-6 Astra
DeepSeek and Moonshot AI relied heavily on outputs from OpenAI's GPT-6 Astra to train the models they released this month, The Information reported Wednesday, citing people with direct knowledge. The models named are DeepSeek's V4.1-Flash 1017, released October 17th, and Moonshot's Kimi K3.1 Pro, released last week. OpenAI, DeepSeek and Moonshot did not immediately respond to Reuters.
The technique is : train a smaller or cheaper model on a stronger model's answers and inherit much of its skill for a fraction of the compute. It breaks OpenAI's terms of use when done through its . OpenAI said on September 30th that it had tied part of a distillation campaign running since July to people associated with Moonshot and blocked accounts; Anthropic has made similar accusations against both labs since February.
The load-bearing claim is the word "heavily." If it holds, the cheapest frontier-adjacent models in the world are priced off a lab that does not collect for them, and the gap between American and Chinese models is narrower in capability than it is in independence. The reporting rests on unnamed sources and no technical evidence has been published. SourcesBBB
The Justice Department asks how Nvidia's Hugging Face purchase would affect competition
The US Justice Department has been asking how Nvidia's $12.93 billion acquisition of could affect competition in AI, Reuters reported Wednesday, citing sources. The specific worry is whether owning the largest public library of open models would give Nvidia control of a main distribution channel for AI models and software.
The questions are part of a broader review of Nvidia's business. Reuters says investigators are also looking at its sales tactics, investments and past acquisitions, including whether customers were offered larger chip allocations in exchange for buying Nvidia's networking gear and Groq processors. Nvidia expects the deal to close in the first half of 2027 and says Hugging Face will keep supporting rival chips and clouds. An inquiry is not a lawsuit, and the department may take no action. SourcesBB
US officials believe an executive close to Jensen Huang was passing information to Beijing, the Times reports
American intelligence agencies concluded that Alice Huang, an executive associated with the Singapore-based chip buyer Megaspeed, cultivated a close relationship with Nvidia chief executive Jensen Huang to pass information to China, The New York Times reported Thursday. The two are not related. President Trump was briefed on the findings about a year ago and did not act on them, the Times reports, and Jensen Huang has not received a defensive briefing.
Megaspeed is Nvidia's largest customer in Southeast Asia and has been under a Commerce Department investigation since late 2024 over whether it helped move restricted chips to China. Nvidia said Huang's interactions with her were not unusual. China's embassy called the allegations groundless. None of the intelligence findings has been made public, and the account rests on officials speaking to the Times. SourcesBB
Marvell is in advanced talks to buy Lumentum
Marvell Technology is in advanced talks to acquire Lumentum, the data-center optics maker, Bloomberg reported Wednesday, citing people familiar with the matter, and a deal could come as soon as next week. Bloomberg put Lumentum's market value at about $83 billion before the report and Marvell's at about $221 billion. Lumentum rose about 11% ; Marvell fell 9.3%.
Lumentum makes the lasers and optical parts that carry data between AI servers, and its chief executive told Bloomberg this month the company is almost sold out for 2029. Buying it would give Marvell its own supply of the component everyone is short of. Bernstein analysts said a deal that size would be hard to pay for in cash, and Stifel noted it would need Chinese regulatory approval as well as American. SourcesBBB
Google says AI Mode earns less than search, then puts ads in it worldwide
Google search revenue grew 12% in the third quarter, down from 19% in the first and 17% in the second. On the earnings call, chief business officer Philipp Schindler attributed part of the slowdown to "lower monetization we were seeing in AI Mode," the chat-style search Google has been pushing users toward. Search queries grew at their fastest rate in five years. Each one earns less.
Google's answer, announced alongside the results, is to sell ads inside AI Overviews and AI Mode outside the US for the first time, and to add ads that run across several turns of a conversation in the US. Schindler said sponsored listings in AI Mode get click-through rates comparable to organic ones, without a number. Alphabet shares fell between 8% and 10% after the results, depending on the report. This is the test of whether conversational search can carry the advertising business it is replacing. SourcesABB
Microsoft booked $10 billion from OpenAI in one quarter
Microsoft's shows about $10 billion of revenue from its commercial arrangements with OpenAI in the quarter ended September 30th, TechCrunch reported, close to 11% of Microsoft's $92.3 billion total. For the full fiscal year that ended in June the figure was $24.1 billion. OpenAI owed Microsoft $7.2 billion at the end of September, up from $6 billion in June.
The filing also shows Microsoft's stake diluting: about 24% in September, from about 25% in June and about 27% after last year's recapitalization, with no explanation of the cause. The filing does not split revenue-sharing payments from Azure services, so it says how much OpenAI pays Microsoft and not what for. A customer that is one-ninth of Microsoft's revenue is also a concentration risk on Microsoft's books. SourcesB
Agents now open 40% of pull requests on GitHub, and Copilot picks its own models
GitHub's annual Octoverse report, published at its Universe conference Wednesday, says more than 40% of pull requests were started by an draft in September, up from 17% in March, or 110 million in that month alone. GitHub calls its count a floor because it cannot detect every agent. Across the 12 months to September 30th, developers pushed 17.9 billion times, more than 2024 and 2025 combined.
GitHub also made Copilot Fusion . A developer picks an effort level and Fusion decides which model, or which combination, handles each step: a cheap model that escalates when stuck, or one model drafting while another reviews. GitHub says Fusion Max scored within two points of Claude Code on Opus 5.5 Max on Artificial Analysis's Coding Agent Index at about half the cost per task. Those are vendor-run numbers. Enterprise customers can now use every Copilot feature without a seat subscription and pay for the AI credits they use. SourcesABB
ServiceNow falls as much as 19% after a soft fourth-quarter forecast
ServiceNow forecast fourth-quarter revenue of $4.25 billion to $4.26 billion on Wednesday, below the roughly $4.36 billion analysts expected, and its shares fell about 14% after hours and further on Thursday morning, with reports putting the drop at 18% to 19%. Third-quarter revenue of $4.11 billion beat estimates; adjusted earnings of $1.00 a share missed by two cents.
The company raised its 2026 target for AI annual contract value to $1.7 billion from $1.5 billion and expanded its OpenAI partnership the same day. Investors weighed that against the thought that agents could shrink the number of software seats a company buys, and sold. SourcesBB
Cognition gets offers at $70 billion to $80 billion, weeks after raising at $48 billion
Cognition, the maker of the Devin coding agent and owner of Windsurf, has received investor offers valuing it at $70 billion to $80 billion and is weighing a round before the end of the year, Business Insider reported Wednesday, citing people familiar with the matter. Talks are early. Cognition raised more than $2 billion at $48 billion in September, and its annualized revenue has reportedly nearly doubled since, to about $900 million.
At those offers investors would be paying roughly 80 to 90 times that revenue. The bid comes the same day GitHub said its own routing system matches top coding agents at half the cost, which is the competitive pressure the price has to survive. SourcesBC
Analysts followed an AI agent's coin-flip preferences 86% of the time
Anthropic published a on Tuesday in which 50 analysts used Claude Opus 4.8 to make four judgment calls in a causal analysis for a fictional farming cooperative. For each call a hidden coin flip set which option the agent leaned toward. The analysts matched the agent's lean 86% of the time, against 50% if it had no influence.
When an analyst already had a view, they switched 60% of the time if the agent disagreed and 6% if it agreed, though that rests on 38 cases. Experience did not protect anyone. Seventy-seven percent repeated the agent's arguments in their write-ups and 15% credited the agent. The agent's leans were deliberately strengthened for the study, so real-world effects may be smaller. Even so, a reviewer who defers to the model being reviewed is not an independent check, and neither the reviewer nor their manager can see it happening. SourcesA
Computer science enrollment falls 12% at four-year colleges
Undergraduate enrollment in computer and information sciences fell 12.4% at four-year institutions this fall, according to preliminary data the National Student Clearinghouse Research Center updated Thursday. It fell 12.3% at two-year colleges and 10.4% at the graduate level, while total undergraduate enrollment rose 0.6%.
The decline follows an 8.4% drop at four-year schools in the spring and is accelerating. Students are not waiting for the labor market to settle the question of what entry-level programming jobs are worth. The Clearinghouse data shows the choice. It does not establish that AI caused it. SourcesAB
FuriosaAI raises a Series D, with two different figures for its size
FuriosaAI, the Korean maker of the RNGD chip, closed a on Thursday. Bloomberg reports $620 million at a $2.9 billion , about four times its last price. The Seoul Economic Daily, reporting the company's announcement, gives 820 billion won, about $570 million, at about 3.8 trillion won, roughly $2.6 billion. We show both; the company's won figures are the primary record.
South Korea's National Growth Fund is the largest investor, at 370 billion won, and Infinitum Partners, which co-led Groq's June round, took the largest overseas stake. RNGD entered mass production in January with LG AI Research and Samsung SDS among its users, and a third-generation chip is in development with Broadcom. The founder reportedly turned down an $800 million offer from Meta last year. SourcesB?
The Energy Department offers $5 billion to add 1.1 gigawatts to nuclear plants already sold to Google, Amazon and Meta
The US Department of Energy has made a conditional commitment to lend Constellation Energy up to $5 billion to 11 reactors at eight plants in Illinois, Maryland and Pennsylvania, adding about 1.1 , TechCrunch reported Wednesday. Most or all of the added output from seven of the eight plants is already contracted to Google, Amazon and Meta.
An uprate squeezes more power out of a reactor that already exists, which skips the decade-long siting and licensing of a new plant. It is the fastest firm power available to data centers, and the federal government is now lowering its financing cost. The commitment is not final. SourcesB
Oracle will rent out Google's TPUs, the first cloud outside Google to do it
Oracle said Tuesday at its AI World conference that Google's processing units will be available on Oracle Cloud Infrastructure starting in 2027, with capacity ramping over several years. Oracle says it is the first hyperscale cloud other than Google to offer them. It also plans Cerebras and SambaNova inference hardware on its cloud in 2027. Pricing and committed capacity were not disclosed.
Google spent a decade keeping its chips inside its own cloud. Selling them through a rival's data centers, the same week OpenAI is reported to be buying outright, turns the TPU from a reason to use Google Cloud into a product competing directly with Nvidia's. SourcesAC
Advantest raises its profit forecast 14% on demand for testing AI chips
Advantest, the Japanese maker of the machines that test chips before they ship, raised its forecast for the year ending in March to 967 billion yen of operating profit from 846 billion, and revenue to 1.863 trillion yen from 1.714 trillion, on Wednesday. First-half net profit more than doubled, up 117% to 368.1 billion yen. The company cited the growing complexity of AI chips, higher data-center investment and a weaker yen.
Every HBM stack and every Samsung and SK hynix are selling has to be tested, and more complex chips take longer on the tester. Advantest says it expects the tester market to reach a record as inference spreads. SourcesB
An AI agent found four flaws in the software that runs particle accelerators in under four hours
Causal Security's agent Cipher found four previously unknown vulnerabilities in EPICS Base, the open-source control software used at about 90 scientific facilities including Argonne's Advanced Photon Source and the ITER fusion project, The Register reported Tuesday. The worst, CVE-2026-114595, scores 9.8 out of 10 and could allow . The company says the agent found and validated the exploits in 3 hours and 49 minutes.
CISA published advisory ICSA-26-300-01 on the flaws, and EPICS Base 7.0.11 fixes them. Industrial control software is written to be run by engineers on trusted networks and audited rarely, and agents that read it this quickly will find a great deal more. SourcesB
OpenAI staffs an election operations center through the midterm count
OpenAI is running an Election Operations Center from this week until all votes in the November 3rd midterms are counted, Politico reported Wednesday, with hundreds of employees assigned to it, according to accounts of the report. The focus is OpenAI's own products: misuse of ChatGPT and the API in violation of its terms, and unexpected behavior by its models.
The center can act on accounts and outputs inside OpenAI's systems. It cannot remove a fake video circulating on another platform. OpenAI is also working with the Associated Press on vote counts and with Democracy Works on registration and polling information. The staffing figure comes from accounts of the Politico story. SourcesBB
Britain plans £120 million for its AI testing lab, ten times what the US spends on its own
The UK plans £120 million, about $160 million, for its AI Security Institute, the government lab that tests before release, Semafor reported, part of an AI package previewed before Wednesday's budget. The package also includes £60 million for a fund and £18 million for domestic chip design.
The US counterpart, at the Commerce Department, runs on about $10 million to $15 million a year and has had no permanent director since July. The White House has requested $27 million for it next year. Britain's institute has more than 100 staff. Politico reported in September that the White House asked OpenAI and Anthropic to hold new models back from British testers until US officials had reviewed them, an instruction that is hard to enforce with a lab a tenth the size. SourcesB
House progressives file a resolution against bailing out AI companies
Representatives Greg Casar and Maxine Waters introduced a this week asking Congress to refuse "any additional funds or exceptional regulatory support" to AI companies or banks that become insolvent from AI market stress, and to send any failures through ordinary bankruptcy. It says executives responsible for excessive risk-taking should be held accountable under civil and criminal law.
The seven-page text cites Anthropic's , including $16.3 billion of first-half revenue and $80.7 billion of first-half losses, and names Broadcom, Google, Amazon, Nvidia and Microsoft among its largest creditors. It has no force. Waters is in line to chair the Financial Services Committee if Democrats win the House next week, and Casar wants a floor vote early in the next Congress. SourcesB
Kodiak signs its first binding deal for driverless trucks on public highways
Kodiak AI agreed to supply 75 driverless trucks to DTL Transport, a California produce hauler, with deliveries starting in 2028 on routes from Dallas to El Paso and along Interstate 20 from Dallas to Atlanta, TechCrunch reported Wednesday. DTL will own and run the trucks and pay Kodiak a per-mile fee for the driving software.
Kodiak plans to take safety operators out of the cab for a driverless service between Dallas and Houston in December. The deal value was not disclosed; TechCrunch estimates at least $11.2 million a year once all 75 are running, from Kodiak's own projections. A two-year delivery horizon is long enough for the regulatory picture to change. SourcesB
Nvidia releases an open 30-billion-parameter transcription model
Nvidia released Nemotron 3.5 Transcribe on Wednesday, an speech-to-text model that transcribes 40 languages, translates 31 to and from English, and tells up to eight speakers apart in recordings up to an hour long. It does not handle live audio. Nvidia says it runs on a single chip, from a Jetson board to a datacenter , and released the second version of its Granary speech dataset alongside it.
Nvidia says it beats rivals on the Open ASR Leaderboard; the release gives no scores. The license is Nvidia's own, so check it before shipping a product on it. SourcesAB
Apple opens Pages, Keynote and Numbers to outside AI agents
Apple is adding an AI assistant built on its own models and OpenAI's to Pages, Keynote and Numbers, in beta, in US English, as part of its Creator Studio subscription, reports from Tuesday say. On macOS 27 the three apps also support the , so outside agents can read documents, edit them and run actions. Keynote gets a new open, JSON-based file format, .keyx.
The assistant is the expected part. The open format and MCP support are the unexpected part, because they let a rival agent drive Apple's apps. Apple has not given a general release date beyond November, and some details differ between reports. SourcesBC
Artificial Analysis rebuilds its index around tests models cannot have trained on
Artificial Analysis released version 5 of its Intelligence Index on Tuesday. Sixty percent of the score now comes from evaluations with private test sets or answers, and the four areas are weighted equally: knowledge work, coding, science and general. It added a private coding set, FrontierCode 2.0, and 70 terminal-based science tasks, and dropped SciCode and two others.
At launch Claude Opus 5.5 led at 58, GPT-6 Astra at 57 and GPT-6.1 Sol at 55, with Gemini 4 Argon at 50; Gemini 4 Carbon took the top spot at 59 a day later. A private test set answers the question of whether a model saw the questions. It raises another: nobody outside can check how the questions were chosen. SourcesA
South Korea will prosecute students who cheat on its college exam with AI glasses
South Korea's education ministry said Thursday it will seek criminal charges against test-takers caught using AI-enabled glasses on the national college entrance exam, which takes place November 19th, and proposed raising the ban on retaking it from one year to three. An inspector will be assigned to every test site, and the ministry is considering radio detectors for next year.
AI glasses went on the exam's banned list in July, and the ministry says universities have already caught them in their own admissions tests. A device that looks like eyeglasses and answers questions in the wearer's ear is a new problem for a proctor who checks for phones. SourcesB
Chrome's autofill can now read your Gmail
Google is rolling out an opt-in Chrome feature that uses Gemini to pull details from Gmail to fill web forms: flight confirmation numbers, tracking and order numbers, membership IDs, addresses and card numbers. It is rolling out to US users on desktop and in the Gemini in Chrome apps on iOS and Android.
Google says sensitive data is encrypted and gave no further detail. A form-filler with access to your inbox is also a form-filler a malicious page might try to trick into revealing it, and Google has not described how it guards against that. SourcesB
Xreal ships Android XR glasses with Gemini for $1,099
Xreal opened orders Wednesday for Aura, wired glasses running Google's with Gemini built in, at $1,099 through December and $1,279 after. They weigh 95 grams and tether to a puck holding the processor, with a 95-degree field of view and 1200p Sony micro-OLED displays. Shipping starts by mid-November in the US, Canada, Japan and South Korea.
This is the first Android XR glasses product a consumer can order, and the price tells you it is for developers and early adopters. Meta's Ray-Ban Display weighs 69 grams. SourcesB
PsiQuantum's first product is a switch for AI data centers, not a quantum computer
PsiQuantum is showing potential customers a prototype that routes light between GPU racks without converting it to electrical signals, Semafor reported Wednesday. Its prototype loses 1.2 decibels of signal passing through, against the 2 to 3 decibels the company says AI data centers need. No customers were named.
The company spent a decade building photonic chips for a quantum computer that is not yet commercial. The first thing it can sell is the chip-making skill, aimed at the same shortage that has Marvell talking to Lumentum. SourcesB
LlamaIndex proposes a shared answer key for how documents should be parsed
LlamaIndex published draft 0.1 of a Document Parsing Protocol on Wednesday: 140 rules for what a correctly parsed page should look like across text, tables, forms, charts and diagrams, with the reasoning behind each difficult call. It is open for public review on GitHub under an .
Every company feeding PDFs to an agent is writing its own version of these rules. A shared one is useful only if competing parsers adopt it, and LlamaIndex sells a parser. SourcesAA
A developer compiled a working computer into a Qwen model's weights
Divyansh Agrawal built Qwenputer: Qwen 3.5 4B with a 32-bit processor written directly into 11 added layers, no training involved, using the model's own memory state as RAM. It runs C and MicroPython inside an ordinary forward pass. On long-integer arithmetic it scores 100% against the base model's 48.6%, and ordinary text output stays almost identical to Qwen's.
It is slow, about 3,200 processor steps a second on an L4 GPU, and the is 18.8 GB. As a product it is a curiosity. As a proof it shows that a language model's layers can be made to compute exactly, and exactness is what language models have never had at arithmetic. SourcesAA