AI news Oct 28: Gemini 4 Carbon Arrives as AI Hardware and Cloud Earnings Soar

Google DeepMind announced Gemini 4 Carbon, which it calls its strongest model, while AI demand pushed record numbers at SK hynix, Vertiv and Microsoft. Funding, new tools for developers and fresh research rounded out the day.
- Google DeepMind's Gemini 4 Carbon ranks #1 on the Artificial Analysis Intelligence Index v5.0 with a score of 59.
- SK hynix posted its first quarter above 100 trillion won, at 104.6155 trillion won.
- Microsoft's Azure grew 48% and Vertiv's sales rose 61%.
- FuriosaAI raised $620 million at a $2.9 billion valuation.
- Scale AI's DualChem found models often flag dual-use chemistry questions and still answer them.
Gemini 4 Carbon leads the model news
Google DeepMind has announced Gemini 4 Carbon, which it calls its strongest model, with state-of-the-art cybersecurity and long-horizon capabilities. According to the company, Carbon ranks #1 on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59.
NVIDIA released Nemotron 3.5 Transcribe and Granary v2 on Hugging Face, a multilingual transcription release with two models, two datasets and two Spaces. The same speech model also arrives on RTX Spark laptops, together with GPU resource controls for background AI agents. In voice, AssemblyAI made its TTS-1 model generally available at $15 per 1M characters, and Inworld added Reasoning to Realtime TTS-2.
Record results for AI hardware and cloud
SK hynix reported third-quarter 2026 revenue of 104.6155 trillion won, the first quarter above 100 trillion won, with operating profit of 83.3338 trillion won. The company says AI memory demand drove both all-time highs.
Vertiv reported net sales of $4,313 million, up 61%, and raised its full year guidance. Microsoft posted first-quarter fiscal 2027 revenue of $92.3 billion, up 19%, with Azure revenue up 48%.
On the money side, FuriosaAI closed a $620 million Series D at a $2.9 billion post-money valuation to scale its RNGD inference accelerator. OneQode signed a 14 MW capacity deal with AdaniConneX in Chennai, with talks on up to 114 MW more.
Agents and developer tools
Microsoft is opening its Copilot harness through Microsoft Foundry, and Satya Nadella says GitHub Copilot Fusion is 40% faster at up to 77% lower cost. OpenAI and ServiceNow are linking ServiceNow workflows with ChatGPT and Codex, and PyTorch 2.15 makes the rewritten NCCL backend the default and adds FP8 on Apple Silicon.
Several infrastructure updates also landed. OpenRouter now weighs speed and reliability in routing, Vercel added Virtual Models to AI Gateway, and Vercel CI entered Beta. Other releases include DigitalOcean Signals, deepset's Haystack update, Google's Agents CLI eval loop, Mastra span queries, Mintlify's mint deploy, Box Mount in Vercel Sandbox, GitLab's Claude Mythos 5.1 integration and Harvey's Intake artifact.
Research and safety
Cartesia says its H-Net architecture can replace tokenization and deliver up to 3-5x compute efficiency gains in controlled tests. Its second post argues that pretokenization is a hidden prior. Scale AI introduced DualChem, and in tests of 31 frontier models found they often flag a question as dual-use and still answer it.
EPFL released spora, an open atlas of spatial proteomics images from more than 5,000 patients. Meta detailed Andromeda, which scores over one billion ads per second. A Forethought paper maps hard cases where goals for powerful AI conflict, and Voyage AI reports up to 44% fewer prediction errors on MongoDB data.
Products and deployments
Kodiak AI and DTL Transport agreed to deploy 75 driverless trucks. Amazon introduced a new Echo Dot and Echo Studio with Home Camera, priced at $99.99 and $249.99. Apple unveiled a MacBook Pro with a tandem OLED display and Touch.
Smaller items: Meta launched Muse in Singapore, Firecrawl Directory offers 240 million profiles, LiteRT gets 1.4x faster GPU inference, Envision Energy showed an AI-powered system for mines and smelters, and oToBrite will show an IP69K 3D stereo camera.

