AI News — 2026-08-25
- Xiaomi built its own AI-capable mobile chip to challenge Qualcomm. The Xring O3 system-on-a-chip brings more power-efficient image processing and better on-device AI to Xiaomi’s flagship devices. — Bloomberg
- Nvidia’s Groq racks will go online later this year. The Groq 3 LPX rack enters full production following Nvidia’s $20B asset purchase, targeting low-latency inference for AI agents. — CNBC
- OpenAI is pushing to build AI agents for nearly every task. The company wants intelligence to go beyond answering questions and start executing on users’ ideas directly. — TechCrunch
- Hugging Face is reportedly fielding acquisition offers around $13B. That would be nearly triple its last valuation of $4.5B set in 2023. — TechCrunch
- General Intuition is raising at a $6B valuation to push into robotics. Valor and Point72 are backing the foundation-model startup just weeks after its $2.3B round. — TechCrunch
- A Goldman Sachs partner warned AI risks “cognitive atrophy” among bankers. Marquee lead Chris Churchman said outsourcing reasoning to models could erode juniors’ analytical skills. — CNBC
- The SEC is probing Situational Awareness, the AI-focused hedge fund that nearly imploded. Regulators are subpoenaing banks that backed and traded for the fund founded by ex-OpenAI researcher Leopold Aschenbrenner. — TechCrunch
- DeepMind alumni’s startup Inherent says its AI agent beat Anthropic and OpenAI at replicating research. Faraday reproduced published papers’ findings without being told the answers in advance. — TechCrunch
- Instinct’s AI assistant is drawing privacy and security scrutiny. Its terms grant broad, perpetual rights to user data, including screen captures and keystrokes. — TechCrunch
- NVent is buying power distribution firm Maverick Power for $1.75B. The deal, nVent’s biggest since its 2018 spinoff, chases surging demand from AI data center buildouts. — Bloomberg
- OpenAI’s Jalapeño inference chip beat Nvidia’s GB300 in benchmark tests. Built with Broadcom, it keeps the KV cache local to compute to cut data movement, leading on tokens per kilowatt and latency on SemiAnalysis’s InferenceX. — TechCrunch
- Apple’s new Mac Studio runs local LLMs nearly 4x faster. The M5 Max drives the Studio, while the Mac Mini gets the first M6 — Apple’s debut chip on TSMC’s 2nm process. — CNBC
- Keenable is building a web index designed for agents rather than people. Its crawl of over 100 billion documents is already serving retrieval in production at several AI labs and inference providers. — TechCrunch