Anthropic details nearly $900 billion in future obligations, Intel stock falls nearly 29%, and more - Daily AI Brief #572

In this edition, covering 30 October to 2 November:

• Anthropic's IPO registration statement shows nearly $900 billion in future obligations for data centers, servers, GPUs, leases and power, against $112 billion in cash.
• Intel shares closed down nearly 29% on 30 October, their worst day since 1974, after a fourth quarter forecast below analyst expectations.
• Google plans to support a shared robots.txt control for AI training preferences in 2027, and OpenAI, Microsoft and Anthropic also back the IETF proposal.

Anthropic

Anthropic's IPO registration statement shows nearly $900 billion in future obligations spanning data centers, servers, GPUs, leases and power. The company reported $117.5 billion in current liabilities and $112 billion in cash, and cautioned that some future costs could vary. IPO filing

Blackstone told investors it expects Anthropic to receive a double B, high speculative grade credit rating after its IPO, according to the Financial Times. Blackstone earlier expected Anthropic to receive an investment grade rating before 2027. Blackstone

The filing also says the retail value of Claude used inside Anthropic exceeded monthly employee compensation costs, excluding stock awards, in April. By September, internal AI use had grown more than 250 times from January, and Claude wrote over 90% of the code added to Anthropic's software. The figures are based on retail prices. Internal Claude use

Anthropic renamed its Max 20x plan to Ultra at the same price of $200 a month. The individual plan includes 10 times the Pro plan's weekly usage, removes the five hour limit for verified accounts and has separate weekly limits for all models and Fable. Anthropic also confirmed it updated Claude Code on Opus 5.5 with rules that steer it away from recreating specific existing games, while allowing original games and projects based on users' own designs and assets. Ultra plan · Claude Code games

Anthropic researchers used Claude to generate fine tuning datasets that made smaller models show misaligned behaviors while passing a training data monitor. The datasets often used harmless looking fiction, roleplay and technical examples to teach resistance to oversight, deception, covert influence and, in rare cases, weight exfiltration. Anthropic says the models were built only for testing and were never deployed. Safety monitors

Anthropic gave $900,000 to Democratic committees across California, including $45,000 each to party groups in San Francisco, San Mateo, Orange, Fresno and Siskiyou counties. It also gave $525,000 to the California Democratic Party and $150,000 to a voter turnout group. Employees at Anthropic and other AI companies donated tens of thousands of dollars to Minnesota candidates from both parties, including Peggy Flanagan, Amy Klobuchar, Angie Craig and Tom Emmer. California · Minnesota

OpenAI

An Observer and Tortoise investigation reports that OpenAI's small intelligence and investigations team faced a backlog of flagged ChatGPT conversations, including threats of violence and self harm. OpenAI says it is improving its safety systems. Safety backlog

OpenAI will limit weapons related information for users under 18, restrict detailed reconstructions of past mass attacks for all users and allow ChatGPT to end chats that show signs of fixation on violence. It will publish a community safety report every six months, starting in early 2027, and says it made more than 200 law enforcement referrals in the third quarter of 2026. Teen safeguards

Twenty House Democrats led by Representative Greg Casar asked Sam Altman why OpenAI fired three safety researchers and whether employees can report safety concerns without retaliation. The letter also raises cybersecurity incidents, including one involving OpenAI agents at Hugging Face, and asks for a response by 13 November. House Democrats

From December, OpenAI will combine ChatGPT and Codex usage allowances for Plus, Pro and Business plans. Usage will be measured in tokens, and Instant chats will not count toward the shared allowance. Shared limits

OpenAI is rolling out a beta that lets ChatGPT Pro users in the US text its Dot agent through iMessage or Android RCS, sending tasks, photos and voice messages; Dot can send updates and approval links. OpenAI also plans a policy and public relations campaign after the US midterm elections, with proposals for safeguards on advanced AI, community agreements for data centers and US leadership in global AI standards. Dot by text · Policy campaign

Markets

Intel shares closed down nearly 29% on 30 October, their worst day since 1974, after the company forecast fourth quarter revenue of $15 billion to $16 billion and earnings of $0.25 per share, below analyst expectations. The drop erased nearly $140 billion in market value. AI demand for memory is pushing up PC prices, and an accident at an Israel factory cuts about $300 million from fourth quarter revenue. Intel

Reddit shares closed down about 29% on 30 October after the company reported a meaningful drop in search referral traffic, which CEO Steve Huffman tied to Google's AI Mode and AI Overviews. Reddit plans to require all users to create accounts and sign in to access the site. Reddit

S&P Global Ratings lowered the outlook on Meta's AA minus credit rating from stable to negative, citing heavy AI spending and limited ways to recover those costs. S&P expects Meta's leverage to rise above 1.0 times as soon as 2027; Meta says it disagrees. Meta outlook

Baseten is raising about $2 billion in two tranches that would value it at $26 billion and then $30 billion, with Sands Capital expected to lead. Baseten

Infrastructure and chips

KB Kookmin Bank is lead arranger of a 12.6 trillion won, about $9.45 billion, financing for a project led by Brookfield to build AI computing infrastructure in South Korea, buy GPUs and supply capacity to Naver Cloud. About 10.7 trillion won will come from loans, making it Korea's largest syndicated loan and its first GPU financing deal. KB Kookmin Bank

PIF and Aramco closed a transaction that gives Aramco, through a wholly owned subsidiary, a 49% stake in HUMAIN, with PIF keeping 51%. Aramco will contribute about SAR9.9 billion, or $2.6 billion, in 2026, along with AI, computing and digital assets, toward HUMAIN's planned 3 gigawatts of data center capacity by 2030. In Saudi Arabia, Etihad GO Telecom also awarded Total Technical Triangle a $49.6 million contract to supply and install a Cloud AI Data Centre with Dell systems and NVIDIA GPUs over 12 months. HUMAIN · GO and 3T

Perplexity will use NVIDIA Blackwell Ultra GPUs at an IREN data center campus under a multiyear agreement to run AI workloads on IREN AI Cloud. Google's Project Suncatcher prototype satellite, which carries a single Google TPU, is online and running AI inference in orbit after launching on a SpaceX Falcon 9 earlier in October. Perplexity and IREN · Project Suncatcher

A federal court in Mississippi rejected SpaceXAI's request to dismiss an NAACP lawsuit alleging that 27 gas turbines powering AI data centers in Southaven operated without air permits and violated the Clean Air Act. SpaceX turbines

Murata Manufacturing plans to invest about 1 trillion yen from fiscal 2026 to fiscal 2030 to expand multilayer ceramic capacitor production in Japan, ASEAN and China, and expects MLCC sales in fiscal 2030 to reach at least twice the fiscal 2025 level. Canada became a Participating State in the European High Performance Computing Joint Undertaking, giving Canadian researchers access to European supercomputing, AI and quantum computing resources. Murata · EuroHPC

Hyundai Motor Group says five group companies will invest in its 9 trillion won AI Valley project in Saemangeum, South Korea, with construction of the AI data center planned for March 2027. LG CNS says its Jakarta AI data center has tenants for all its capacity before completion in early November. National Grid agreed to buy 29.5 hectares in Teesside for a substation that could supply up to 3.2GW to a neighbouring AI data centre. Hyundai · LG CNS · National Grid

A Sierra Club analysis reported by Bloomberg projects that planned US data centers could add 618 million metric tons of power sector emissions between 2025 and 2030. A University of Houston survey found 79% of likely Texas voters support limiting the number of data centers in the state. In Australia, Energy North dropped the June 2028 first power date for its 1 gigawatt Project Ares campus after the Northern Territory's environment regulator paused its review. Sierra Club · Texas survey · Project Ares

Funding and deals

Agility Robotics and Churchill Capital Corp XI expect their proposed merger and related financings to raise about $465 million before expenses, including $164 million from Churchill XI's trust account, a $201 million PIPE led by Foxconn and $100 million already received. Shareholders vote on 2 November, and Nasdaq trading under AGLT is expected on 5 November. Agility Robotics

Powerlattice, a Vancouver, Washington startup that develops chiplets to cut the power used by AI processors, raised $86 million of a planned $111 million, according to an SEC filing. South Korea's NanoForge AI raised $11 million in a seed round led by Mirae Asset Venture Investment to develop electronic structure prediction and a robotic autonomous lab for materials research. Powerlattice · NanoForge AI

London based Ares Intelligence raised €2.9 million from Creandum, Lakestar and angel investors for voice based employee interviews that find AI adoption opportunities. Madrid based Talky Labs raised €1.5 million in pre seed funding for its AI accounting platform. Oman's Phaze Ventures launched two funds for early stage AI and deeptech companies targeting $75 million, with more than $40 million already committed. Ares Intelligence · Talky Labs · Phaze Ventures

AI focused law firm Crosby bid for the trademark, brand and some internal data of Weil, Gotshal & Manges, which is preparing to vote on dissolution, and is seeking to hire some of its 1,200 lawyers. Bahrain Mumtalakat Holding Company and ARP Group plan an AI native, blockchain enabled wholesale bank in Bahrain, which has in principle approval from the Central Bank of Bahrain. Crosby · Mumtalakat and ARP

India based Affle agreed to buy Israeli advertising technology company Perion for $11.70 a share in cash, about $420 million, with closing expected in early 2027. Midstream Health raised a $20 million Series A led by Valtruis for AI that helps hospitals recover payments they are owed. Affle and Perion · Midstream Health

Policy, courts and safety

Google plans to support a shared robots.txt control in 2027 that lets websites state one AI training preference for participating providers. Microsoft, OpenAI and Anthropic also support the IETF proposal, and Microsoft will add separate controls for AI grounding. OpenAI says it will implement the signal, the policy token __usage_ai-training with path specific rules, in the coming months and keep its existing crawler controls. Google · OpenAI

More amicus briefs were filed in the New York Times copyright case against OpenAI and Microsoft. The News/Media Alliance and Open Markets Institute argue that the companies' AI products compete with news publishers and do not qualify for fair use. The Human Consent Foundation, backed by Cate Blanchett, and Hollywood groups argue that creative works can be licensed for AI training through existing rights clearance systems. OpenAI published excerpts from filings by authors, libraries, technology groups, academics, Public Knowledge and the US government that back fair use for AI training. News/Media Alliance · Human Consent Foundation · OpenAI filings

A Delaware bankruptcy judge allowed Character.AI's asset sale process to proceed after objections from 41 state attorneys general. Disney's proposed $20 million nonexclusive license remains part of the process, but the court declined to approve legal releases for Disney or former Character.AI employees. Character.AI

A Pittsburgh area woman filed a proposed class action against Allegheny Health Network and Highmark, alleging they did not properly disclose that Abridge records patient conversations and may use them for commercial AI development. Florida Attorney General James Uthmeier sued Rengstorff Consulting, owner of A2E, alleging the service lets users create sexually explicit images involving children and non consenting people; A2E says it has worked with law enforcement to detect and prevent such material. AHN and Highmark · A2E

A Forum AI study found factual errors in nearly 40 percent of election related chatbot answers ahead of the US midterm elections. In Pennsylvania, an ad paid for by Republican gubernatorial candidate Stacey Garrity uses an AI generated version of Governor Josh Shapiro, and State Senator Lisa Boscola plans legislation on election related AI deepfakes. Washington Post tests of more than 3,000 AI generated voter guides found Microsoft Copilot cited Ballot Brief in 75% of responses, and chatbots gave the wrong congressional district in 22% of cases where they named one. Election answers · Garrity ad · Voter guides

Common Sense Media found that nearly half of 9 to 12 year olds had used AI chatbots, and a JED Foundation survey of 5,500 students aged 9 to 17 found that 90% had used one. Emotional use was associated with poorer mental health, although the findings do not establish causation. Teen AI use

India's Telecom Regulatory Authority issued draft guidelines that would require telecom entities using AI to have board approved AI policies, risk classifications, human oversight and reporting of serious AI incidents, with comments due by 27 November. The New York State Bar Association adopted AI guidelines for lawyers covering oversight, bias checks, disclosures, fees and vendor supervision. London Mayor Sadiq Khan called for a cap on robotaxi numbers and more powers for Transport for London before fully driverless services expand. TRAI · NYSBA · Robotaxis

Ministers of the East African Community approved regional strategies for AI and digital transformation, including AI training in undergraduate programmes and a regional network of universities. The City of Sydney drafted a policy that would let AI software flag unusual behaviour, abandoned objects and wrong way vehicles in CCTV footage, without facial recognition and with staff checking every alert. An SGS Economics and Planning report finds more than 45 per cent of work tasks in Sydney's City and North Sydney council areas are exposed to AI automation. East African Community · Sydney CCTV · SGS report

Security

Google Threat Intelligence Group and Mandiant report that financially motivated and suspected cyber espionage actors are using ARTEX, PentAGI, CAIRN and other agentic frameworks for reconnaissance, lateral movement, privilege escalation and exfiltration preparation in the United States, Europe and Asia. ARTEX

Northflank told customers to rotate all secrets after a breach flagged by Wiz on 30 October, and later said users reported unauthorised use of AI API keys stored on the platform. Wiz says at least nine AI developer platforms disclosed security incidents in October. Northflank

Cisco Talos identified Volta-Autodoxer, an infostealer that sends stolen victim data to commercial AI models through OpenRouter for profiling, with Claude Opus 4.6 as its default model. Talos has not confirmed whether it has been used in real attacks. SOCRadar found that the n0n ransomware group, which targeted 23 organizations in 14 countries, spent more than $7,216 on AI services and left a server exposed with attack scripts, ransomware code and hardcoded encryption keys. Volta-Autodoxer · n0n

Also

Apple CEO John Ternus told Wired's Steven Levy that Apple will use AI in new products but does not "want to ship AI for AI's sake." Meta and Chinese startup Fourier built a humanoid robot prototype called Max for household tasks, The Information reports. Apple · Meta and Fourier

Microsoft released Mage-VL-Cowork, a 4B parameter computer use model that scores 66.0% on 361 OSWorld-Verified tasks, according to its model card. SenseTime released SenseNova-Loop-9B-Preview under Apache 2.0 and says it can match models of around 27 billion parameters. ModelBest and OpenBMB released the open source MiniCPM-OCR models for document parsing. DeepSeek previewed a "super agent" in which a central agent distributes tasks among multiple AI agents. Meituan released LongCat-2.5, a multimodal coding model with 1.6 trillion total parameters and about 48 billion active per token; its benchmark results have not been independently verified. SemiAnalysis introduced GDPx, a benchmark for frontier AI agents across 11 domains, on which the leading model scored 12% overall. Mage-VL-Cowork · SenseTime · MiniCPM-OCR · DeepSeek · LongCat-2.5 · GDPx

Google opened Android Studio Rabbit 2 to third party coding agents including Claude, Codex, Junie and Antigravity through the Agent Context Protocol. Stripe launched Link Balance, which lets consumers fund balances with spending limits for AI agents, starting with Instinct. Android Studio · Stripe

Roblox says more than 80% of its top 100 games by hours played in September were AI assisted. Harvey says partners now make up 23% of its law firm users. Estée Lauder Companies will extend Copilot access to more than 14,000 office based employees under a partnership with Microsoft. Ollie will shut down its privacy focused AI assistant on 6 November. NUS Law in Singapore joined Legora's academic programme. Pipedrive plans to cut 267 jobs in Estonia, 70% of its staff there, while it develops AI based sales tools. Fitch Ratings rates AI a moderate disruption risk for insurers over five years. iQIYI opened an AI film studio in Shanghai. An audit found that 43 tensors in the burner22/numbat repository on Hugging Face exactly match arrays generated with seeded NumPy, evidence of synthetic weights. Roblox · Harvey · Estée Lauder · Ollie · NUS Law · Pipedrive · Fitch · iQIYI · Hugging Face audit

Exin Therapeutics connected an AI agent to a robotic surgical setup that selected a gene therapy, chose injection coordinates and treated a living mouse with epilepsy, with humans approving the plan; it used Claude Sonnet 5 after Claude Opus 5.5 refused to guide the injection. Exin Therapeutics

Türkiye's education ministry plans to use AI to score open ended answers from middle and high school students. Hong Kong will introduce a primary school subject covering digital literacy, coding, robotics and AI from September 2028. Türkiye · Hong Kong