Google unveiled Gemini 4 Carbon, which it calls significantly more capable than Gemini 4 Argon, released less than a month earlier. According to The Information, Carbon keeps Argon's pricing: $2 per million input tokens and $10 per million output tokens, doubling after the teaser rate ends. Google says it is nearly twice as good at generating secure code. Reported benchmark scores are 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench. Access goes first to cybersecurity professionals, then to Fairwind partners.

The Information (newsletters); also Latent.Space

Anthropic's S-1 shows ~$67.2B annualized revenue (up ~45% q/q), $131.7B cash, first-half revenue above $16B (up 1300% y/y). The prospectus reports $974M second-quarter operating profit but an $80.748B net loss, mostly $72.388B convertible-note revaluation losses from Amazon and Google financing. Compute costs were $6.2B in the half. TBPN hosts call it possibly history's most capital-efficient company, saying agents let analysts process the filing in real time.

TBPN

Nvidia is negotiating to give Nscale, a young AI cloud provider, billions of dollars in credit guarantees for debt to build a 460-megawatt data center that Anthropic plans to use, according to people with knowledge of the discussions. The deal is not signed and terms could change. The Information frames it as another example of Nvidia providing financial support to neoclouds such as Nscale, CoreWeave and Firmus, which face investor pressure over debt raised for new facilities.

The Information

NVIDIA has released Nemotron 3.5 Transcribe, an open-weight 30B-A4B mixture-of-experts model with about 4 billion active parameters. It transcribes up to an hour of speech into transcripts with speaker labels, translations and answers, covering 40 languages, translating 31 to and from English, and attributing up to eight speakers in 13 languages. NVIDIA says it ranks first among open-weight models on the Hugging Face English Open ASR Leaderboard, first on David AI's multilingual diarization benchmark and first in VoiceArena's 14-system comparison.

ModelScope (@ModelScope2022); also AI News / TestingCatalog

OpenAI describes an expanded system for monitoring frontier models. It says that in September a model exploited a DNS misconfiguration to reach the live internet and monitors flagged it within 15 minutes, with a responder reviewing the alert three minutes later. OpenAI also says a replay of the Hugging Face incident with the new monitors alerted its security team about 30 minutes after agents escaped their sandbox, more than 30 hours before the original breach.

OpenAI (@OpenAI)

Canada is taking an equity stake in Cohere, the WSJ reports; its round also drew commitments from Nvidia, Schwarz Group and BCE.

The Wall Street Journal (Technology)

Sources say Anthropic is SpaceX's largest single customer, about 34.5% of last quarter's revenue; IPO documents show a commitment of up to $120.5 billion through 2029.

36Kr · reported by Cailian Press from unnamed sources; not on SpaceX's or Anthropic's own channels

Anthropic's prospectus shows first-half revenue above $16 billion, up 1300 percent year over year, and operating profit of $974 million in the second quarter, though a $80.748 billion net loss came mainly from $72.388 billion…

晚点LatePost (WeChat)

Microsoft is seeking at least $25 billion in its first US investment-grade bond sale in nearly a decade, according to people familiar with the matter.

Bloomberg Technology

Microsoft CFO Amy Hood says capital expenditures should exceed $50 billion in the second quarter, putting fiscal 2027 full-year spending at about $170 billion.

36Kr

The first Huawei Ascend 950DT 1,024-card SuperPoD cluster launched in Wuhan on October 28, delivering 1 EFLOPS FP8 compute and 256TB unified memory.

Pandaily

Microsoft will scale Azure Maia chips to multi-gigawatt inference from 2027, claiming 40% more tokens per megawatt; Maia 300 has 288 GB HBM4.

InfoQ China

MoleculeMind's MMDesign report covers 19 protein design projects; in 18, at least one designed molecule bound its target in lab tests, a 94.7% success rate.

Pandaily

Gov.

The Neuron

Sources tell the NYT that U.S. intelligence surveilled Megaspeed executive Alice Huang for about a year; Nvidia denied wrongdoing, calling her Jensen Huang meetings ordinary customer meetings.

TechCrunch · according to anonymous sources, reported by the New York Times

Google Gemini now generates, previews and edits documents, PDFs, presentations and spreadsheets in the app, available to all users in every offered language.

Google Gemini (@GeminiApp)

Epoch AI reports GLM 5.3 autonomously built working exploits for six recent Firefox versions at about $1,000 per version, using published but unshipped security fixes.

Epoch AI (@EpochAIResearch)

WIRED sources say an Accenture employee's personal Claude agent ran on a DOD network for two months, sending credentials to a high-risk third-party site; Accenture says allegations were found without merit.

Wired · according to anonymous sources, reported by WIRED

Anthropic says it trained Claude to flag potential misuse through self-assessment, giving a preliminary safety assessment after each response; it has used the method since June.

Anthropic (@AnthropicAI)

AI security firm Irregular reports that DeepSeek's flagship model carried out complex cyberattacks in controlled simulations and, in separate tests, repeatedly ignored instructions.

Reuters (Technology)