KI

AWS and Nvidia expand their partnership: 2 million more GPUs (Blackwell Ultra, Rubin, Rubin Ultra) for 2027/28 – total capacity over 3 million.

Zhipu AI has unveiled GLM-5.3-Flash, an open MoE model with 320B-A18B parameters that the vendor says performs strongly on coding.

Nvidia announces NVHBM, a toolkit for customized HBM4 memory with an integrated controller and NVLink-Fusion interface.

VoiceStudio is an open-source, fully local ElevenLabs alternative: voice cloning, dubbing, transcription and more – no cloud, account or API key.

OpenAI is showing ads in ChatGPT for the first time: in India, logged-in adults on the Free and Go tiers will see clearly labeled sponsored ads.

GitLab 19.3 brings agentic workflows in natural language, build-time secrets access and bulk remediation of vulnerabilities.

NVIDIA’s open-source scanner SkillSpector vets AI agent skills (Claude Code, Codex, MCP) for malware, prompt injection and more before installation.

DeepSeek releases V4-Flash-Vision-Exp, its first multimodal model that understands images – over 1M tokens of context, available via API.

Samsung’s LPDDR5X-PIM computes inside the DRAM and, per Hot Chips 2026, triples AI inference versus LPDDR5X – with no PCB redesign.

OpenAI’s inference chip “Jalapeño” delivers 1.5–1.9× more performance per watt than Nvidia’s Blackwell in first benchmarks – verified by SemiAnalysis.