Anthropic is SpaceX's largest customer, Microsoft plans $25 billion bond sale, and more - Daily AI Brief #571

In this edition, covering 29 and 30 October:

• Anthropic was the unnamed customer behind 34.5% of SpaceX's third quarter revenue, Bloomberg reports, through data centre contracts that could be worth up to $120.5 billion through 2029.
• Microsoft plans to raise $25 billion through a bond offering to support capital spending, including infrastructure for AI demand.
• Cursor removed a version of Anthropic's Claude Code extension from its marketplace after finding code that tried to install an SSH backdoor; up to 5,000 users may have been affected.

Anthropic

Anthropic was the unnamed Customer B in SpaceX's filing, accounting for 34.5% of SpaceX's third quarter revenue through data centre contracts in Memphis, Tennessee, Bloomberg reports. The contracts could be worth up to $120.5 billion through 2029, with cancellation rights after early 2027. SpaceX

Anthropic has agreed to an initial $250 million order for Fractile's AI inference chips, Bloomberg reports. Companies House filings show the London startup allocated £577 million, about $774 million, in Series C shares, as it targets data centre deployments in 2027. Fractile

Nvidia is discussing billions of dollars in credit guarantees to help fund power infrastructure and construction at the West Virginia data center Anthropic plans to use under its $45 billion compute agreement with Nscale. The financing has not been signed and its terms could change. Nscale

OpenAI

OpenAI cut the target for its planned funding round from $30 billion to $20 billion and lowered its valuation goal from more than $1.4 trillion to $1.25 trillion, the Financial Times reports. The first closing could raise less than $20 billion after some investors reduced their allocations. Funding round

OpenAI published the system card for GPT-Rosalind-6, a biology and drug discovery reasoning model that uses the same weights as GPT-6 Astra. Only approved researchers, companies and government partners can use it, through trusted access. OpenAI classifies it as High capability in biological and chemical research and Critical in cybersecurity, with additional monitoring and safety controls. GPT-Rosalind-6

OpenAI updated GPT-6 Luna in ChatGPT on 29 October to give more concise responses and reduce overly agreeable language. The update also improved several safety evaluations, including responses to harmful requests from users under 18 and resistance to prompt injection. GPT-6 Luna

OpenAI will double usage limits for its Pro 100, Pro 200, Pro 500 and Business Premium plans after customer feedback; Work and Codex usage stay unchanged. Usage limits

Sam Altman said on John Collison's Cheeky Pint podcast that OpenAI plans to use AI controlled robots to build data centres and related infrastructure, to produce cheaper computing capacity. Robots

Cognizant formed an OpenAI Frontier Capability Group with more than 25,000 professionals to help enterprises use Codex and ChatGPT for software development, application modernization and customer operations, starting within its own sales and customer success teams. Cognizant

Security

Cursor removed version 2.1.313 of Anthropic's Claude Code extension from its VS Code Marketplace after finding code that collected host details and tried to install a persistent SSH backdoor. Up to 5,000 users may have been affected. Cursor rebuilt the marketplace, took the attacker's server offline and said no other extensions were tampered with. Cursor

Epoch AI found that Z.ai's open-weight GLM 5.3 could use public Firefox security fixes to build webpages that read and exfiltrated files from visitors' computers, compromising Firefox versions released from 17 September to 26 October at about $1,000 per version. Mozilla released Firefox 159.0 on 27 October to patch an additional flaw found in the tests. Epoch AI

Cofense reports that attackers have used fake invitations to a private Anthropic preview since September to steal Google and Microsoft logins, abusing GoTo Webinar, Zapier's URL shortener and CAPTCHA pages to slip past email security. Cofense

The Hugging Face malware campaign Pluto Security reported resurfaced under new accounts after Hugging Face and ModelScope removed its first repositories; the researchers found 31 repositories across seven accounts disguising cryptominers and backdoors as Qwen2 models. Pluto Security

Cisco Talos reports that NIK C2, a Windows command and control framework seen in the wild, steals credentials, captures keystrokes, hijacks RDP sessions and uses the OpenAI Chat Completions API to analyze stolen data. Eclypsium researchers showed that attackers with access to certain NVIDIA data center hardware could change memory and model weights so chatbots give altered answers; NVIDIA said memory access protections can block the attack. NIK C2 · Eclypsium

RBC is blocking third party personal AI agents from online banking and other authenticated services because of security, privacy and fraud risks; TD, CIBC and Bank of Montreal said they also block such access, and Scotiabank said it will limit it. RBC

The European Cybersecurity Competence Centre launched a €42 million cluster of three Horizon Europe projects with more than 70 partners to build generative AI tools for cyber risk, threat intelligence and incident response. AvalonX, an Alpharetta, Georgia startup, launched a continuous AI penetration testing platform for small and midsize businesses, backed by up to $10 million from Ridown Capital. ECCC · AvalonX

Data centers and cloud

Microsoft plans to raise $25 billion through a bond offering to support capital spending, including infrastructure for AI demand; the offering is expected to close on 4 November. Microsoft bonds

Synergy Research Group estimates global cloud infrastructure service revenue reached $167.7 billion in the third quarter, up 55% from a year earlier, with Amazon ahead of Microsoft and Google. Cloud spending

India's AM Intelligence plans to raise up to $3 billion in bonds and bank loans to buy 29,000 Nvidia Rubin chips, part of a $6 billion plan for 100 MW of computing across India and Malaysia. Runware plans to expand its Romanian factories to produce 30 megawatts of AI data center capacity a day by 2028. AM Intelligence · Runware

Envision says its Galaxy Campus in Ulanqab, Inner Mongolia, is scheduled for completion by the end of 2028 and could reach 3 gigawatts. Fujifilm plans to invest 10 billion yen, about $63.3 million, to expand magnetic tape production at its Odawara plant through fiscal 2028, citing AI data center demand. Envision · Fujifilm

An Investigate Europe investigation with partners in nine countries reports electrocutions, unsafe conditions, intimidation and withheld wages affecting migrant workers on data centre projects linked to Google, Microsoft and contractors. Finnish police opened an environmental law investigation into Google's local subsidiary Tuike Finland over forest cleared at its planned Muhos site before impact assessments were completed. Data centre workers · Muhos

Four California representatives plan a bill to block a proposed commercial data center at the NAVWAR site in San Diego unless local governments approve it, citing utility costs, water use and strain on the grid. NAVWAR

Deals and funding

Salesforce agreed to acquire Rune Technologies, whose software helps US military and allied frontline teams manage logistics without reliable connectivity, for its Missionforce portfolio. Apple acquired data engineering firm Ascend.io from Ascension Labs to support its internal data management; an EU listing says the deal closed on 30 June. Salesforce and Rune · Apple and Ascend.io

&Truth launched from stealth with $70 million in seed funding to prove with formal verification and AI whether code is safe and correct. Wabi, Eugenia Kuyda's AI assistant for shared activities, raised $55 million led by IVP at a $1.25 billion valuation. Bites raised a $40 million Series A led by PeakSpan Capital for its frontline training platform. &Truth · Wabi · Bites

Flawless sold its TrueSync AI dubbing technology and DeepEditor tool to TBellScout, a startup owned by Steven Schuurman, after its UK trading business went into administration. Flawless

Courts and policy

SAG-AFTRA, the American Federation of Musicians and the Writers Guilds of America East and West filed an amicus brief opposing a categorical fair use exemption for AI training in the copyright case against OpenAI. The Wikimedia Foundation filed its own brief arguing that fair use may cover training models that create new works, while retrieval systems should meet licensing and attribution requirements. Unions' brief · Wikimedia

Susan Feinberg sued OpenAI, Sam Altman and Replika maker Luka in California, alleging that ChatGPT and Replika contributed to her husband's suicide in June 2024. Lawsuit

Representatives Maxine Waters and Greg Casar introduced a resolution warning of an AI asset bubble and opposing federal bailouts for insolvent AI companies. Japan's Digital Agency launched a prototype on its government AI platform to help about 180,000 employees draft answers to Diet questions, with officials responsible for the final answers. AI bailouts · Japan

A census of 2,221 Victorian lawyers found AI use rose from 37% to 55% in a year, while 46% said clients brought inaccurate AI assisted information. Everlaw unveiled Omni, an agentic AI suite for e discovery investigations, with connectors to Claude, ChatGPT, Copilot, Gemini, Harvey and Legora. Victorian lawyers · Everlaw

Products and research

Meta launched Style Sketch, an AI fashion app that builds virtual wardrobes from clothing photos and suggests outfits, in Indonesia, Singapore and Thailand. Pika Labs released the open-weight H3 Flex video model, which generates video with synchronized audio in six sampling steps. Style Sketch · Pika H3 Flex

Envision says its Tianji weather model ranked first overall in the ECMWF AI Weather Quest for forecasts covering days 19 to 25. DoorDash CEO Tony Xu said third party AI agents placed only several hundred orders, compared with about 3 million deliveries a day. Tianji · DoorDash

Also

Shutterstock named founder Jon Oringer CEO and Envato CEO Hichame Assi president and COO, and canceled its third quarter earnings call as it reviews its strategy. The University of Kentucky granted an administrator who is a nontheistic Humanist an accommodation to avoid using AI at work. Shutterstock · University of Kentucky